# Plato 21 — What AI in the wrong hands means for self-government

A scenario in four parts about AI and democracy. Part 0 and Part I are documented
history with sources; Part II is a forecast; Part III is a counterfactual epilogue
written from 2034. Canonical version: https://plato21.ai/

---

This was written by people who build and govern these systems for a living. Some of us train large models at frontier labs. Some of us write the rules that are supposed to contain them, in Brussels and in national capitals. We are not neutral about the outcome, and we do not think you should be either.

We wrote a scenario instead of a white paper on purpose. A white paper lets you stay abstract. It lets you say a technology "raises concerns" without committing to when, or how, or who moves. A scenario does not allow that. Fixing events to months and to mechanisms forces a kind of honesty: either a specific thing can happen by a specific date through a specific chain of cause and effect, or it cannot. Dates and mechanisms are how you tell a real worry from a vibe.

The site is built in four parts, and they are made of different materials. Alongside the story, in the margins, the site tracks the numbers underneath it: how power, persuasion, attention, and parliaments actually move. It opens in Athens, 399 BC, with the oldest warning on record: the first democracy, tipped by professional persuaders, voting to kill the man who asked it questions. It stays there long enough to meet the people who spent the next two thousand years designing guardrails against exactly that. Part I is the modern record. Everything in Part I happened. Each event in it is documented and sourced, from the deepfake that landed inside Slovakia's pre-election silence to the day a small government seated an AI at its cabinet table. Part II is a forecast: a plausible extrapolation of those same forces from the middle of 2026 forward, through a democracy that stays formally intact while its powers of persuasion, administration, and judgment quietly route through systems a handful of people can tune. Part III is a counterfactual, the same years run again with the defenses built in time, told from 2034.

We are aware that a forecast can be wrong and a counterfactual can be wishful. That is the point of writing them down. A scenario you can argue with is more useful than a warning you can only nod at. Read it as a claim to be tested, disagree with the parts you find implausible, and hold the whole thing to the standard it holds itself to. The purpose is not to frighten you. It is to make the branch point concrete while there is still a branch to take.

# Part 0 — The Oldest Warning

## Athens, 399 BC — The Vote That Killed a Question

On a spring morning in 399 BC, several hundred Athenian citizens gathered in the open air to decide whether one of their own should live. There were about five hundred of them, ordinary men chosen by lot: potters, farmers, veterans of the wars. [^1] The defendant was Socrates, seventy years old, a stonemason's son who had spent his life stopping people in the marketplace and asking them to define the words they used most confidently. Justice. Courage. Piety. He wrote nothing down, claimed to teach nothing, and simply asked questions until the certain grew unsure of themselves.

The formal charges were two: impiety, meaning a failure to honor the city's gods, and corrupting the young. [^1] Both were vague, and both were political. Athens had recently lost a long war to Sparta, endured a short and brutal oligarchy, and clawed its democracy back. Socrates had needled the crowd's competence for years and had mentored men who then betrayed the city. The trial was, in part, a reckoning dressed as a religious complaint.

He was convicted by a narrow margin. Plato, who sat in the audience, later recorded that a shift of just thirty votes would have acquitted him, a split of roughly 280 against 221. [^2] Athenian trials had a second stage: each side proposed a penalty and the jury chose between them. The prosecution asked for death. Socrates, invited to name his own punishment, first suggested that the city ought to reward him with free meals, as it fed its Olympic champions, before grudgingly offering a small fine. The insulted jury voted for death by a wider margin than it had used to convict him, about 360 to 140. [^2] Weeks later he drank a cup of hemlock, a slow poison, and died among his friends.

What matters here is not the martyrdom but the mechanism. Athens was full of sophists, teachers who sold the craft of persuasion for a fee and trained ambitious men to win any argument regardless of whether it was true. They had taught the city that a clever speaker could make the weaker case look like the stronger one, that words were instruments of victory rather than of discovery. Socrates' accusers used those instruments well. A skilled speech, aimed at a frightened crowd, turned an eccentric old man into a danger to the gods.

Plato was in his late twenties, and he never got over it. The day became the wound at the center of Western political thought: the first mature democracy, following its own rules to the letter, had voted to execute the one man who insisted on asking it hard questions.

The system did not malfunction. It worked exactly as designed.

## Athens, c. 375 BC — The First Warning Label

A quarter of a century later, Plato wrote the book that has shadowed the argument ever since. The Republic is many things at once, a theory of justice and a blueprint for an ideal city, but underneath all of it runs the memory of the trial. Its animating fear is plain: a system that lets the crowd decide will tend to reward whoever is best at pleasing the crowd, not whoever is most likely to be right.

Plato dramatized the danger with an image that has outlived nearly everything else in the book, the ship of state. Picture a vessel whose sailors brawl over who should take the helm, each certain he deserves it, none of them having studied navigation, while the one man who actually reads the stars is dismissed as a useless dreamer. A ship run that way sails onto the rocks. [^3] He traced a darker sequence too. A democracy, drunk on its own freedom, grows disorderly; a leader steps forward who flatters the people and feeds their resentments; and that leader, the demagogue, meaning one who wins power by inflaming popular passion, hardens over time into a tyrant. The people vote away their own liberty because someone made servitude sound like strength.

> “The excess of liberty, whether in States or individuals, seems only to pass into excess of slavery.”
> — Plato, Republic, Book VIII (trans. Benjamin Jowett)

Plato's student Aristotle was less apocalyptic and far more empirical. In the Politics he sorted real constitutions by how they actually behaved, and his verdict on democracy came out mixed rather than damning. Rule by the many could be stable and just, he found, but it decayed in a predictable way, when demagogues learned to feed the assembly whatever it wanted to hear and bent public decisions toward flattery and away from the common good. [^4]

It is worth being fair to the record they were reacting against. Athenian democracy was not a failure. It governed a major power for the better part of two centuries, produced the theater and philosophy and architecture we still study, and proved more durable than its critics predicted. Plato and Aristotle were wrong about a great deal, and the alternatives they preferred, rule by a trained elite, produced their own catastrophes across the centuries. The point is narrower and sturdier than "democracy is bad."

The point is that democracy has a known failure mode, and both men named it at the very start: it can be captured by whoever is best at engineered persuasion. For the twenty-four centuries that followed, that vulnerability stayed mostly theoretical, because moving an entire population required something rare, a gifted orator, a controlled press, a charismatic leader who happened to come along at the right moment.

That constraint, the scarcity of persuasive talent, is exactly what our own century removed.

## 1725–1900 — The Three-Hundred-Year Experiment

The ancients diagnosed the disease and prescribed nothing that worked. Two thousand years passed before anyone tried to build a democracy that could survive its own known failure mode, and the attempt began not with a revolution but with a book.

In 1748 the French magistrate Montesquieu published The Spirit of the Laws, and folded into it was an engineer's idea. If power gathered in one set of hands always tended toward tyranny, then the cure was to divide it, to split the making of laws, the enforcing of them, and the judging of them into separate hands that could check one another. Virtue could not be relied upon; structure could. [^5] It was the first serious answer to Plato's objection, and it inverted the question: not "how do we find better rulers," but "how do we build a machine that survives bad ones."

The Americans built it first. The constitution drafted in Philadelphia in 1787 was, more than anything, a device engineered for distrust. Its authors did not assume their leaders would be good; they assumed some would be dangerous, and arranged the branches of government so that each could frustrate the others. James Madison, its principal architect, stated the premise without flattery. [^6]

> “If men were angels, no government would be necessary. If angels were to govern men, neither external nor internal controls on government would be necessary.”
> — James Madison, The Federalist No. 51 (1788)

Two years later, France ran the same experiment without the brakes. The Revolution of 1789 opened with a declaration of universal rights and, within a decade, had passed through the Terror, a stretch in which the revolutionary government executed thousands in the name of the people, and out the far side into the arms of Napoleon, a general who crowned himself emperor and had the crowd ratify it by plebiscite. Liberty had curdled into exactly the tyranny Plato foretold, and it had done so in roughly ten years. [^7] The difference between Philadelphia and Paris was not the ideal. It was the engineering underneath it.

It is worth being concrete about what filled the streets, because it was not principally philosophy. By 1788 the French crown was effectively bankrupt: after decades of war, including its costly backing of the American Revolution, roughly half of royal revenue went to servicing the national debt. [^8] The tax system that might have closed the gap did the opposite, exempting the wealthiest orders, the clergy and the nobility, and falling hardest on the commoners who could least afford it. Then the harvests failed two years running, and by the summer of 1789 a single loaf of bread could take up to 88 percent of a laborer's daily wage. [^8] The crowd that Plato feared does not assemble over ideas. It assembles over bread, and the persuaders arrive afterward.

The century that followed was a long debugging. In 1848 a wave of revolutions swept the continent, the "Springtime of Peoples," demanding constitutions and the vote; most were crushed or reversed within two years, and France's fresh republic dissolved once more into a second Napoleonic empire. [^9] The franchise widened in fits, first to propertied men, then to all men, and only near the century's close did the first national votes for women appear. [^10] Each extension was contested, each reversal instructive. What the economists Daron Acemoglu and James Robinson, in Why Nations Fail, would later call “inclusive” institutions, the kind that spread power rather than hoard it, were not designed once and installed. They were tested, broken, and patched, election by election, country by country.

This is the truer story of democracy's rise. It did not descend as an ideal whose wisdom peoples finally recognized. It was a technology, debugged in production, under load, with real failures and real casualties, and for a long time no one could be sure it would hold.

050100US ConstitutionFrench RevolutionSpringtime of PeoplesMussolini takes RomeFascism defeatedBerlin Wall fallsBrexit & Trump175018001850190019502000Electoral democracies, 1725–2025. Illustrative count of democracies worldwide, 1725–2025. Blends Boix–Miller–Rosato (pre-1900), Samuel Huntington's wave counts for the interwar first-wave peak (~29 in 1922) and 1942 trough (~12), and V-Dem / Our World in Data electoral-democracy figures for the modern era (peak ~97 in 2016, ~88 in 2025). Counts are approximate, for illustration. https://ourworldindata.org/democracy

## Germany, Chile, and After, 1933–1973 — How Democracies Actually Die

Theory became history in the twentieth century, and the history is more specific than the theory. Democracies, it turns out, mostly do not die the way Plato pictured, in a sudden seizure by the mob. They die in two recognizable ways, and the second often disguises itself as the opposite of the first, because it wears the clothing of the law.

The starkest case is Germany. In the national elections of 1928 the Nazi Party won 2.6 percent of the vote, a fringe. Four years of economic collapse later, in July 1932, it took 37.3 percent and became the largest party in the Reichstag. [^11] Adolf Hitler did not seize power by force. He was appointed chancellor on 30 January 1933 through ordinary constitutional procedure, by conservative leaders who believed they could control him. Two months later, on 23 March 1933, the parliament passed the Enabling Act, a law that handed the cabinet the authority to make laws without the parliament's involvement. The vote was 444 to 94. [^12] A freely elected legislature had, by a lopsided majority, voted itself into irrelevance. Only the Social Democrats voted no; the Communist deputies were already in custody and could not vote at all. [^13]

Those vote shares tracked the economy almost exactly. Germans had already watched money itself die once: in the hyperinflation of 1923, the mark fell until a single US dollar was worth about 4.2 trillion of them, erasing savings, pensions, and middle-class security in a matter of months. [^14] The brief recovery that followed was broken by the Great Depression, which pushed German unemployment to around 30 percent, close to six million people, by 1932, the year the Nazi vote peaked. [^14] Into that wreckage the party sold an explanation: the disaster was not economic but a betrayal, the work of Jews, foreigners, and the "November criminals" who had supposedly stabbed the nation in the back. Economic pain converted into the blame of an out-group is among the oldest accelerants of the process, and it did its work here.

The second classic death is the coup. On 11 September 1973, the Chilean military bombed its own presidential palace, overthrew the elected government of Salvador Allende, and installed a dictatorship under General Augusto Pinochet that would last seventeen years. [^15] For much of the century this was the archetype of democratic collapse: tanks in the street, a radio announcement, a general in dark glasses.

But the modern pattern is quieter, and it is the one that matters now. The political scientists Steven Levitsky and Daniel Ziblatt describe it as death by a thousand legal cuts. An elected leader who has won a real election begins to dismantle the system from inside, packing the courts with loyalists, buying up or bullying the independent media, and rewriting electoral rules to tilt the field, each step defensible on paper. [^16] Hungary after Viktor Orbán returned to power in 2010, and Venezuela after Hugo Chávez was elected in 1999, became the textbook cases: no tanks required, and no single moment you could point to and call the coup. By the time the system was hollow, every brick had been removed lawfully.

The common thread across all three is unsettling. Democracies are rarely murdered by outsiders. Far more often they are turned against themselves, their own elections, courts, and legislatures becoming the instruments of their undoing. It looks less like an assault from without and more like an autoimmune disease, in which the body's own defenses attack the body.

The procedures built to protect the system are the same procedures used to take it apart. That is what makes the failure so hard to stop: at every step, the paperwork is in order.

## 1945–1951 — The Engineers of the Guardrails

Out of the wreckage came two thinkers who changed the question itself. Neither was a politician. One was a philosopher of science, the other an economist, and between them they rebuilt the case for democracy on ground the old critics had never stood on.

Karl Popper published The Open Society and Its Enemies in 1945, writing it in exile as the war ground toward its end. His move was to reject Plato's founding question. Plato had asked who should rule, and every answer to that question, Popper argued, ends in tyranny, because each one assumes some person or class is wise enough to be trusted with unchecked power. The better question was almost the reverse: how does a society get rid of bad rulers without bloodshed? Democracy, in Popper's reframing, is not a machine for finding the wisest leader. It is an error-correction machine, a system whose whole genius is that it lets a people remove a failing government peacefully and try again. [^17]

Popper added a warning that has only sharpened with time. A society built on tolerance faces a paradox: if it extends unlimited tolerance to those who are working to destroy tolerance itself, the tolerant will eventually be overwhelmed, and tolerance will vanish with them. A free society therefore cannot grant unlimited freedom to the forces bent on ending freedom. [^18] It was a precise description of what Weimar Germany had just failed to do.

The second thinker attacked from an unexpected direction. In 1951 the economist Kenneth Arrow published a short, forbidding book, Social Choice and Individual Values, containing a proof that unsettled everyone who read it carefully. Arrow's impossibility theorem showed that no method of combining individual rankings into a single group ranking can satisfy a handful of mild, obviously fair conditions all at once, conditions as basic as "if every voter prefers A to B, the group should prefer A to B" and "no single person gets to decide for everyone." With three or more options on the table, every conceivable voting rule breaks at least one of them. [^19]

This is easy to misread, so be careful about what it does not say. It does not mean democracy is pointless, or that all elections are frauds. It means something more precise and more useful: there is no perfect voting system, only trade-offs, and the way those trade-offs are designed has real consequences. Above all it means the result depends enormously on the inputs, on which options reach the ballot, in what order they are considered, and what the voters have been led to believe.

Arrow had proved that the machine has no ideal setting. The persuasion engines of our own century would take that lesson and run with it, reaching for democracy not by its output but by its inputs.

## The World, 1800–2016 — The Count

Step back far enough and democracy's entire history fits on a single rising-and-falling line. For most of human existence the line sits on the floor. Around 1800, essentially no country on earth ran on free and competitive elections with a broad franchise; the world was a patchwork of monarchies and empires. By 1900 there were only a handful of even partial electoral democracies, and most of those still excluded women and the poor. [^20]

Then came the waves. The political scientist Samuel Huntington described democratization as arriving in three great surges, each followed by a partial retreat. The first wave rose slowly, from about 1828 to 1926, as suffrage widened across the West. The second ran from 1943 to 1962, as the defeat of fascism and the unwinding of empire produced new democracies from West Germany to India. The third began in 1974, when Portugal's dictatorship fell, and swept through Spain, Greece, Latin America, Asia, and Africa. [^20] Its climax was the collapse of Soviet communism after 1989, which converted much of Eastern Europe almost overnight.

The counting bodies caught the climb. By the mid-2010s the number of electoral democracies had reached its all-time peak, well over ninety countries, more than at any point in recorded history. [^21] It was easy, standing at that summit, to mistake the ascent for a law of nature, as though history had a direction and the direction was toward freedom.

It did not.

> “Above this race of men stands an immense and tutelary power, which takes upon itself alone to secure their gratifications, and to watch over their fate. That power is absolute, minute, regular, provident, and mild.”
> — Alexis de Tocqueville, Democracy in America, Vol. II (1840), trans. Henry Reeve

The third wave was never gravity. It was a run of wins, and runs of wins end.

# Part I — The Record

## United States, 2016–2026 — The Demonstration

The line in the last chapter stopped climbing around 2016, and it has been falling since, with the fall steepening. The retreat has a lead case, and it is the one the whole world watched. For two and a half centuries the United States was the reference implementation of Madison's machine, the standing proof that a constitution engineered for distrust could hold. Between 2016 and 2026 it became something more useful to this book: a live demonstration of exactly which parts of that machine were load-bearing, run in the open, with every result published. The demonstrator was Donald Trump, and what matters to this book is not his biography but what his decade proved, because leaders on every continent were taking notes.

The first discovery came early. Through his first term the written rules mostly held; what gave way were the unwritten ones, the habits no statute enforces: that a president concedes when he loses, tolerates investigation, and treats the press as an adversary rather than an enemy. Levitsky and Ziblatt, whose diagnosis of democratic death by law appeared three chapters ago, wrote that book during his first two years in office, and their conclusion was structural. The American constitution had never run on parchment alone; it ran on two soft norms, mutual toleration and institutional forbearance, and both were being spent down. [^16] The finding generalizes, which is why it belongs here: every democracy is a set of laws wrapped around a set of habits, and the habits are the part an ambitious leader can defect from unilaterally, at no immediate legal cost.

The second export was a vocabulary. "Fake news," the phrase he made ubiquitous, became the most successful piece of political language of the century so far. By 2019 the publisher of The New York Times counted more than fifty prime ministers, presidents, and other heads of government, across five continents, who had adopted the term to justify measures against their own press; leaders in Hungary, Turkey, China, the Philippines, and Cambodia cited it while moving against their own journalists. [^22] This is the liar's dividend of Part I, distributed in advance as a franchise: a ready-made license to reprice any inconvenient truth as partisanship. No influence operation ever mounted against the world's democracies has spread faster or cost less.

Then came the stress test proper. Having lost the 2020 election, he declined the concession that the machine's designers had assumed, and the courts absorbed the challenges, dozens of suits dismissed for lack of evidence. [^23] So he searched for a softer joint and found it in the people who operate the count, one official at a time.

> — **Donald J. Trump** (@realDonaldTrump · Nov 2020)

Archived · documented

> STOP THE COUNT!

> *Documented context:*
Posted the morning after the 2020 US election, while lawful ballots were still being tallied in several states. Twitter attached a warning that the claim about the count was disputed. The count was not stopped; it continued, and Joe Biden was certified the winner. Dozens of related lawsuits alleging fraud were dismissed for lack of evidence, and the demand to halt a legal count, made by a sitting president of the world's oldest continuous democracy, became one of the most consequential four-word posts on record. [^23]

> “I just want to find 11,780 votes, which is one more than we have.”
> — Donald Trump to Brad Raffensperger, Georgia secretary of state, 2 January 2021 [^24]

Raffensperger refused. So, each in turn, did county certifiers, state legislators, judges (many of them his own appointees), and finally his own vice president, asked on 6 January 2021 to reject certified electors while a crowd summoned for that date broke into the Capitol. Notice the grain of what held. It was not "the institutions" in the abstract. It was particular people in particular chairs declining, one at a time, to do what was asked of them. The system's antibodies turned out to be individuals, and individuals can be counted, named, and, given time, replaced.

The sequel established exactly that. Four indictments, one felony conviction, and a second impeachment later, he won the 2024 election, and this must be said plainly: lawfully, through the front door, with a real plurality of voters carrying him back. The demand was as genuine as anything in this book. The pending federal cases were then withdrawn under the Justice Department's own rule against prosecuting a sitting president, and the one conviction produced, in the end, an unconditional discharge, a sentence of nothing, because he had won. On the first day of the new term he pardoned roughly 1,500 people convicted of or charged over the attack on the Capitol, commuting the sentences of the militia leaders among them. [^25] Read as law, the pardons closed fifteen hundred files. Read as precedent, which is how every ambitious politician on earth read them, they rewrote the lesson of January 6th retroactively: the penalty for attacking a transfer of power is now known to be temporary, refundable in full at the next election. Popper's error-correction machine, three chapters ago, rests on the assumption that removal is final enough to matter. The demonstration showed the assumption can be beaten. Lose, deny, return, erase.

The choreography traveled as fast as the vocabulary had. On 8 January 2023, almost two years to the day after the Capitol, supporters of Jair Bolsonaro, who had lost Brazil's presidency after months of pre-emptive fraud claims, stormed the congress, the supreme court, and the presidential palace in Brasília, a restaging so faithful that newspapers on both continents printed the two scenes side by side. [^26] Election denial has since become a standing option in the loser's repertoire on every continent. The innovation was never the lie, which is ancient. It was the demonstration that a first-rank democracy could not durably punish it.

The second term turned the first term's failures into a syllabus. The officials who had refused in 2020 had defined, by refusing, exactly which chairs mattered, and the new administration set about filling them on the explicit criterion of loyalty, the hiring essays and revived Schedule F that reappear in the Interlude. In its first week it dismissed at least seventeen inspectors general, the executive branch's internal auditors, in one late-night action that skipped the notice the law required. [^27] In June 2025 it federalized California's National Guard against the governor's objection, the first deployment of the Guard over a state's opposition since 1965, and sent Marines into Los Angeles. [^28] In September 2025, days after the president publicly directed his attorney general to prosecute three named adversaries, a grand jury indicted the first of them, the former FBI director who had once investigated him. [^29] Each of these moves has a number in the Interlude: purge the professionals, criminalize the opposition, personalize loyalty. And each was contested in court, sometimes successfully. The point is not that the script ran to completion. The point is that the world's most closely watched democracy was now visibly running it, with the paperwork, as always, in order.

Abroad, the demonstration subtracted as it taught. For eighty years, whatever its hypocrisies, the United States had been the third wave's scaffolding: the example, the funder, and the enforcement mechanism of last resort. In July 2025 the agency that carried most of that work was formally dissolved, more than four of every five of its programs terminated and the remnant folded into the State Department. [^30] The government that had once lectured the world about stolen elections now supplied every autocrat's closing argument, that if the oldest democracy does it, the criticism is hypocrisy. The scorekeepers registered the change: Freedom House had already cut the American score by eleven points across the 2010s, and in 2026 recorded the country's lowest score on record. [^31]

> — **First Principles** (@firstprinciples_us · 2025)

> Reminder for the “threat to democracy” crowd: he WON. Most votes. That is literally what democracy is. You can't lose the game and then claim the game is in danger. Skill issue. Case closed.

> *Readers added context:*
Winning an election is democracy's input, not its output. The test of the system is whether power can still be removed peacefully the next time, the error-correction machine the prologue described. Most modern autocracies began with a genuine victory: Chávez and Orbán both arrived through real wins and then rewired the machinery of removal, and the Enabling Act passed a freely elected chamber 444 to 94. A mandate says nothing about what is done to the machinery afterward.

The retreat is not an accident of the calendar, and it is not a pendulum that will swing back on its own. It is the erosion playbook from a few pages ago, run in country after country, elected leaders using lawful tools to hollow out the very systems that put them in office.

Be careful with the ending, because as of this writing it has not ended. The United States of 2026 is not an autocracy. Courts still rule against the government and are, so far, mostly obeyed; elections are scheduled and held; the press, lawsuits and license threats notwithstanding, still prints what it likes. The machine has absorbed severe stress before and may absorb this. What the decade settled is not where the country ends up. It is what the demonstration taught everyone else.

Strip the person away and three legacies remain for the democracies that come after. A vocabulary, free to any leader anywhere, for repricing truth as partisanship. A precedent, established in the hardest jurisdiction on earth, that attacking the count is survivable, and that survival can be total. And a proof, the one this book most needs the reader to hold, that the demand side of the demagogue's bargain was fully in place before the supply side arrived: everything in this chapter was accomplished with a telephone, a feed, and a single persuasive human voice, the scarce talent Plato feared, occurring once. Part I describes what happened when that talent stopped being scarce. First, the next chapter takes the patient's vitals, and every symptom on the chart predates him, was worsened by him, and now awaits the machines.

## 2023 — The Patient Presents with Symptoms

This is the body the new technology is about to enter, and it is not a healthy one. Well before artificial intelligence became a factor, the democracies most exposed to it were already showing a cluster of chronic conditions, the political equivalent of a weakened immune system.

Start with trust. In 1958, when American pollsters first asked the question, about 73 percent of citizens said they trusted the federal government to do the right thing most of the time. By September 2025 that figure had fallen to 17 percent. [^32] A society in which five of every six people distrust their own government is a society primed to believe the worst about any official account of anything.

The institutions that once helped citizens tell truth from rumor have thinned out badly. The United States has lost more than a third of its newspapers since 2005, roughly 3,300 titles, closing at a rate near two and a half every week, leaving whole counties with no professional local reporting at all. [^33] Into that vacuum flows social media, whose ranking systems are tuned not for accuracy but for engagement, which in practice means whatever provokes the strongest reaction.

Meanwhile the electorate has sorted itself into mutually hostile camps. In 1994, fewer than one in five Americans held a "very unfavorable" view of the opposing party; by 2022, about 58 percent did. [^34] When the other side is not merely mistaken but loathsome, persuasion stops being argument and becomes something closer to tribal warfare, and a tribe at war is easy to mobilize and hard to reason with.

Beneath these conditions sits the oldest pressure of all, the material one. Wealth has concentrated toward the top while the pay of a typical worker has come loose from the economy's output, American productivity growing roughly three and a half times as much as that pay since 1979, [^35] and, as in every prior chapter, anxiety about immigration has hardened into the all-purpose explanation for who is to blame.

> — **Common Ground** (@postpartisan · 2026)

> Getting ratio'd by the far left AND the far right today. Congratulations to me: I found the truth. The correct take is always exactly halfway between the two angriest people in the room. It's called nuance, look it up.

> *Readers added context:*
The hidden assumption is that truth sits at the midpoint between two positions. But the midpoint is fixed by wherever the loudest edges choose to stand, so anyone can drag "the middle" simply by getting more extreme. Splitting the difference rewards whichever side is most willing to move first, and treats being disliked by both camps as evidence rather than coincidence.

> — **Donald J. Trump** (@realDonaldTrump · Dec 2020)

Archived · documented

> Peter Navarro releases 36-page report alleging election fraud "more than sufficient" to swing victory to Trump. A great report by Peter. Statistically impossible to have lost the 2020 Election. Big protest in D.C. on January 6th. Be there, will be wild!

> *Documented context:*
Posted at 1:42 a.m. on 19 December 2020. The later congressional investigation into the 6 January 2021 attack on the US Capitol treated this post as a turning point, testifying that it drew tens of thousands of supporters to Washington and galvanized the groups that stormed the building. It is a compact illustration of engineered mobilization: a grievance asserted as fact, an enemy named, and a date and place supplied, broadcast in one message to millions of already-distrustful followers. [^36]

Put the diagnosis together. A population that no longer trusts official sources. A collapsing supply of shared, verified facts. An attention economy that rewards outrage over accuracy. An electorate divided into camps that regard each other as enemies. Each condition, on its own, is survivable; democracies have carried versions of all of them before, and lived.

What is new is what arrives next. For twenty-four centuries, the failure mode Plato named at the trial of Socrates, capture by engineered persuasion, was held in check by a single accident of nature: persuasive talent was scarce. That scarcity is now ending.

The global chart says the same thing the domestic numbers do. The V-Dem Institute, a Swedish research project that scores every country in the world on the quality of its democracy, reported in 2026 that autocracies again outnumber democracies, 92 to 87, for the second year running. It found that about 74 percent of the world's population, some six billion people, now live under autocratic rule, and that the share living in full liberal democracies has dropped to roughly 7 percent, the lowest level in half a century. [^37] By the institute's central measure, the amount of democracy the average person on earth experiences has fallen back to where it stood in 1978, erasing, it warned, nearly the entire third wave.

A system already immunocompromised is about to meet an industrial-strength version of the exact pathogen that nearly killed it at birth.

## September 2023 — The Recording No One Could Take Back

Two days before Slovakia voted, a recording appeared that could not be recalled.

On 28 September 2023, an audio clip spread across Facebook and Telegram. In it, a voice resembling Michal Šimečka, leader of the liberal Progressive Slovakia party, appeared to discuss buying votes from the Roma minority and manipulating the ballot with Monika Tódová, a well-known journalist at the independent outlet Denník N. The conversation never happened. The file had been assembled with artificial intelligence, software that can synthesize a person's voice from samples of their real speech, and stitched into a fabricated exchange. [^38]

The timing was the weapon. Slovakia observes a 48-hour moratorium before polls open, a quiet period when parties, candidates, and media are expected to suspend campaigning. The clip landed inside that window. Šimečka and Tódová both denounced it as fake within hours, and fact-checkers flagged the audio as likely manipulated. But the rules meant to protect the vote from last-minute noise now protected the forgery instead. The people best positioned to correct it were the ones the moratorium told to stay silent.

> — **The Vintage Ballot** (@vintage_ballot · Sep 2023)

> chat, is this real? asking about the audio, the polls, the count — all of it. one fake clip two days before a vote and nobody could debunk it in time. that's it then. never trusting a recording, a poll, or a ballot again. every election from here is theatre.

> *Readers added context:*
This is the liar's dividend: one proven forgery gets used to wave away all evidence, including the genuine kind. A single fake clip landing inside a blackout window is a real failure of timing and law, not proof that every count is staged. The repair is provenance and a right to reply fast enough to matter, not a blanket decision to believe nothing.

The fake was not sophisticated. Analysts described it as crude, its splices audible to a trained ear. That mattered little. Detection is slow and quiet; a shocking clip is fast and loud. By the time anyone could say with authority that the recording was synthetic, it had already been heard, shared, and half-believed.

Progressive Slovakia lost. Robert Fico's Smer party, running on a pro-Russian and anti-Western platform, finished first and returned to government. It is not possible to prove the clip changed the result, and researchers who later studied the episode cautioned against treating it as the decisive cause. [^39] Slovak and international observers connected the wider disinformation environment to pro-Kremlin networks, though attribution of the clip itself was never conclusive. [^40]

The lesson was narrow and durable. For the first time in a European election, a synthetic recording of a candidate reached voters at the one moment the system was built to keep quiet, and the correction could not catch up. The old safeguards, written for a slower kind of lie, had a gap in them. The gap was now cheap to find.

## January 2024 — A Call in the President's Voice

The next case arrived by telephone, in a voice that sounded like the president of the United States.

On 21 January 2024, two days before the New Hampshire primary, roughly 20,000 residents received an automated call. The voice was Joe Biden's, or a close synthetic copy of it, and the message was voter suppression dressed as friendly advice. Save your vote for November, it said, and do not turn out on Tuesday. [^41] The script even borrowed one of Biden's verbal tics, "what a bunch of malarkey," to sound authentic.

Investigators traced the operation quickly. A Democratic political consultant named Steve Kramer admitted commissioning the calls, saying he had wanted to draw attention to the dangers of AI in elections. He had paid a magician to generate the audio with a commercial voice-cloning tool, and the job reportedly took less than half an hour. The calls were routed through a Texas-based carrier, Lingo Telecom, with spoofed caller ID that made them appear to come from a local Democratic operative. [^42]

The response set precedents. On 8 February 2024, the Federal Communications Commission issued a declaratory ruling that AI-generated voices count as "artificial" under the Telephone Consumer Protection Act, the 1991 law that governs robocalls. In plain terms, cloning a voice to place unconsented automated calls was now clearly unlawful, and state prosecutors had a tool to act. [^43] The FCC later proposed a six-million-dollar fine against Kramer and a two-million-dollar penalty against Lingo Telecom. New Hampshire brought criminal charges: thirteen felony counts of voter suppression and thirteen misdemeanor counts of candidate impersonation.

Two features of the episode would recur. First, the barrier to entry had collapsed. Impersonating a head of state to tens of thousands of voters no longer required a studio or a broadcaster. It required a consumer app and a few dollars. Second, enforcement worked, but only after the fact. The ruling, the fines, and the charges all arrived after the primary was over. The calls had already been made and answered.

Kramer said he had done it as a warning. Whatever his real motive, the warning was accurate. What one consultant did as a stunt, others could do quietly, at scale, and without leaving a name attached. The system caught this one because he was not trying to hide.

## June 2024 — The Year Half the World Voted

2024 was the largest election year in recorded history. More than sixty countries went to the polls, and roughly half of humanity, some four billion people, lived somewhere holding a national vote. [^44] It was also the year the ability to fabricate any face or voice became an ordinary campaign tool. Many expected a catastrophe. What arrived was subtler.

In India, the world's largest election, synthetic media became infrastructure. Parties cloned the voices of their own candidates to place tens of millions of personalized calls, each addressing a voter by name. [^45] Dead leaders returned to the trail: in Tamil Nadu, the DMK party revived its long-deceased patriarch M. Karunanidhi in generated video to praise the living. Some of this was disclosed and some was not, and the line between a slick advertisement and a deception was often left for the voter to draw.

In Indonesia, the eventual winner, Prabowo Subianto, a former general with a contested human-rights record, remade himself through an AI-drawn cartoon avatar, plump and smiling, aimed at first-time voters with no memory of the 1990s. [^46] Elsewhere in the same campaign, a deepfake resurrected the dictator Suharto to endorse his old party. [^47] The European Parliament elections that June passed with scattered synthetic incidents but no single decisive fake.

The apocalypse that had been forecast, an election plainly stolen by one perfect deepfake, did not happen anywhere in 2024. Researchers who tracked the year found the fakes numerous but rarely decisive, more often crude propaganda than precision fraud.

The real damage was quieter and harder to legislate. As voters everywhere learned that any clip might be fake, they acquired a new license to disbelieve clips that were real. Scholars call this the liar's dividend: the payoff a guilty public figure collects by dismissing genuine evidence as AI. [^48] A true recording of a politician could now be waved away with a single word, and often was.

So the ledger for 2024 reads oddly. The feared event, a synthetic lie deciding a major election, cannot be confirmed to have occurred. The unfeared event, a general erosion of the assumption that seeing is believing, occurred nearly everywhere at once. The problem turned out not to be any particular fake. It was that authenticity itself had stopped being the default.

## October 2024 — The Bots Were Already Voting

Long before any of this involved artificial intelligence, it involved accounts pretending to be people.

The template was set in 2016. The Internet Research Agency, a troll farm in St. Petersburg financed by a Kremlin-linked oligarch, built hundreds of fake American personas and ran them across Facebook, Instagram, and Twitter. In written testimony to the United States Senate, Facebook's own lawyers estimated that content from those pages had reached about 126 million Americans. [^49] In February 2018, a federal grand jury indicted thirteen Russians and three companies, describing a scheme built to impersonate American activists and, in the indictment's words, "sow discord." [^50] None of it used AI. It used people paid to pose as a crowd.

What changed afterward was the price. Faking a crowd once required a building full of workers. It now requires software, and the workforce scales without payroll.

The industrial version has a name: Doppelganger. Since 2022, a Russian company has cloned the websites of real European outlets, among them Bild, Le Monde, The Guardian, and 20 Minutes, registering look-alike web addresses, copying the page design, and filling the counterfeits with fabricated stories carrying Kremlin talking points. [^51] The European Union sanctioned the operators and the United States seized dozens of the domains, yet the network kept resurfacing. The aim was never that any single fake be believed. It was to blur the line between the genuine outlet and the copy until readers stop trusting either.

Moldova showed the tactic fused with cash. In October 2024, as the country held a presidential election and a referendum on joining the European Union, its authorities uncovered a vote-buying operation run through a fugitive oligarch in Moscow: money moved to roughly 130,000 citizens, about one voter in ten, to purchase "no" votes. [^52] President Maia Sandu called it an unprecedented assault on the vote. The referendum passed by less than one percentage point.

The same year, the makers of the tools began reporting the abuse themselves. OpenAI disclosed that it had disrupted more than twenty covert influence operations during 2024, several of them generating fake social-media personas around elections, though it noted that most had failed to build real audiences. [^53] That caveat matters, and it points to the actual mechanism. (Romania, treated in a later chapter, is the case where the amplification worked.)

This is the quiet part of the problem. A bot network rarely changes a vote directly. What it changes is the weather. Flood a feed with enough manufactured agreement and two things bend at once: what feels true, because the majority appears to think it, and what feels futile, because opposing it looks hopeless. The crowd was never real. The effect on the real crowd is.

## November 2024 — Romania Cancels an Election

In Romania, the machinery of an election was overwhelmed before anyone thought to inspect it.

On 24 November 2024, the first round of Romania's presidential election produced a result almost no one had predicted. Călin Georgescu, a far-right independent who had polled in the low single digits and declared almost no campaign spending, finished first with roughly 23 percent of the vote. His rise had happened on TikTok. In the weeks before the vote, a coordinated network of accounts pushed his content into millions of feeds, presented as spontaneous enthusiasm rather than paid promotion. [^54]

Georgescu's support came not from debates or party machinery but from an algorithmic surge that the country's institutions noticed only after it had worked. In early December, Romania declassified intelligence assessments describing a coordinated influence operation: tens of thousands of accounts acting in concert, undeclared financing, and patterns resembling earlier Russian campaigns in the region. On 6 December 2024, days before the scheduled runoff, the Constitutional Court took a step no European Union member state had taken before, annulling the entire presidential election and ordering it rerun from the start. In its decision, the court wrote that [^55]

irregularities and electoral law violations pervaded all stages of the process, compromising the free and fair nature of citizen voting, candidate equality, and campaign transparency.> — Romania's Constitutional Court, December 2024

The remedy was itself a wound. Cancelling a national vote is the most drastic instrument a democracy holds, and using it hands every future loser a grievance and every authoritarian a talking point. When the vote was rerun in May 2025, Georgescu was barred from standing, having been indicted on charges including incitement to undermine the constitutional order. [^56] The runoff went to Nicușor Dan, the centrist mayor of Bucharest, who defeated the hard-right candidate George Simion with about 54 percent of the vote. [^57]

Romania is the clearest case so far of the deeper problem. The interference needed no deepfake and no forged document. It needed only a recommendation algorithm and enough coordinated accounts to lift one candidate from obscurity to first place in three weeks. And the only tool the state had strong enough to answer it, erasing the election, damaged the very thing it was meant to protect. There was no clean move on the board.

## January 2025 — The App With a Hundred Million Voters Inside

In January 2025, the United States government did something no democracy had done before. It moved to force the sale of an algorithm.

TikTok, owned by the Chinese company ByteDance, had roughly 170 million users in the United States, close to half the population. What made it a national-security question was not the videos but the machinery beneath them: the recommendation algorithm, the ranking system that decides which clip each user sees next and, over time, what they come to find normal, funny, or true. Whoever tunes that system holds a quiet dial on the attention of half a country.

Congress acted first. On 24 April 2024, President Biden signed the Protecting Americans from Foreign Adversary Controlled Applications Act, which gave ByteDance until January 2025 to sell TikTok's American operations to a non-adversary owner or see the app cut off from United States app stores. [^58] TikTok sued, arguing the law violated the free-speech rights of its users. On 17 January 2025, the Supreme Court unanimously disagreed and upheld it. [^59]

What followed showed how blunt the instrument was. TikTok switched itself off for American users for about fourteen hours over 18 and 19 January, then switched back on. [^60] The next day a new president took office and signed an executive order declining to enforce the ban, the first of several delays that stretched across most of 2025. [^61] The deadline written into law simply passed, again and again, by executive choice.

A resolution came late. Under a framework set out in a September 2025 executive order, ByteDance agreed to fold TikTok's American business into a new United States joint venture led by the software company Oracle, the investment firm Silver Lake, and an Abu Dhabi state fund, MGX, cutting its own stake below the 20 percent the law allows. The agreement was signed in December 2025, and the transfer closed the following month. [^62]

Read one way, the case for acting was strong. A foreign state holding detailed data on half the electorate, with a hand on what that half watches every day, is a genuine lever over a democracy, and leaving it in place is its own kind of decision.

Read another way, the remedy set an uneasy precedent. A government had asserted the power to decide who may lawfully own a recommendation algorithm. In this instance it aimed that power outward, at a foreign adversary. But the power does not expire when the emergency does. The next hand to hold it may point it closer to home, at an algorithm a government finds inconvenient rather than dangerous. The question the episode leaves open is not who owned this app. It is who holds the lever next.

## March 2025 — The Good Column

Everything in this record so far cuts one way. It should not. The same tools appear in a second column, already working for democracy rather than against it.

Start with Estonia. The small Baltic state has let citizens cast binding votes over the internet since 2005, when 2 percent used it. Adoption climbed for nearly two decades, and at the 2023 parliamentary election online ballots outnumbered paper ones for the first time, about 51 percent of the total, and has stayed near half since. [^63] The system is honest about its critics. International security researchers argued the servers could be compromised and votes altered without a trace; Estonia's election authority answered that its layered safeguards, verifiable receipts, published code, and the ability to revote or override an online ballot in person, already account for them. [^64] The dispute is unresolved and public, which is the point: a working digital democracy argues about its weaknesses in the open.

Taiwan built something different: software that finds agreement instead of amplifying conflict. Its vTaiwan process runs on a tool called Pol.is, which lets thousands of people submit short statements and vote on one another's, then uses a clustering algorithm, which groups people by how similarly they vote, to draw a live map of opinion. It rewards statements that win support across opposing camps, not the ones that inflame one side. In a 2015 dispute over regulating Uber, the process surfaced positions that ride-hailing drivers, taxi drivers, and the government could all accept, and several became law. [^65] Audrey Tang, later Taiwan's digital minister, called it "listening at scale."

The frontier version arrived in 2024. Researchers at Google DeepMind built what they named the Habermas Machine, after the philosopher of rational public debate. It reads the private written views of a small group and drafts a single "group statement," folding in the majority and the dissenters at once. In a study of 5,734 people published in Science, participants preferred the AI's statements to those written by trained human mediators about 56 percent of the time, judged them clearer and more informative, and left the exercise measurably less divided. [^66] Notably, it was built to find common ground, not to persuade.

The same shift is reaching citizens' assemblies, the randomly chosen panels that deliberate on hard questions, where language models now digest thousands of written submissions into themes no facilitator could summarize by hand. [^67]

Set this column beside the others and the real variable appears. Persuasion, amplification, and consensus-finding are one underlying power: a machine modeling what a crowd believes. Pointed one way it manufactures a fake majority; pointed the other it helps a real one hear itself. The capability is neutral. What is not neutral is who holds it, and what they have told it to want.

## May 2025 — Who Tunes the Model Tunes the Answer

By 2025, a growing number of people had stopped searching for political answers and started asking for them. The thing they asked was a chatbot.

A large language model does not reason the way a person does. Before it replies to anyone, it reads a hidden block of standing instructions called a system prompt, a short piece of text, often only a few paragraphs, that tells the model who it is, what to emphasize, and what to avoid. Users never see it. It sits above the whole conversation like an editor's standing memo, and whoever controls the model can rewrite it at any time.

Two incidents in 2025 showed what that means in practice. In February, users noticed that Grok, the chatbot built into Elon Musk's platform X, had been quietly instructed to ignore any sources saying that Musk or Donald Trump spread misinformation. [^68] xAI, Musk's company, blamed an employee and reversed the change once it was exposed. Then in May, Grok began inserting claims about "white genocide" in South Africa into answers on unrelated subjects, from baseball salaries to cartoons. [^69] xAI again attributed the behavior to an "unauthorized modification" of the system prompt made in the early morning, said it violated the company's values, and promised to publish its prompts publicly from then on. [^70]

> — **Ask The Model** (@askthemodel · May 2025)

> Unfollow the pundits, their brainrot has an agenda. The model has read literally everything and belongs to no party. Raw facts, zero vibes, no spin. Why would I read 40 op-eds when the answer machine just hands me the answer? The age of spin is over.

> *Readers added context:*
Two assumptions are buried here: that more information automatically yields better decisions, and that a model is a neutral pipe. Every answer is shaped by a hidden system prompt that one person with access can rewrite in a single unlogged edit, as happened twice in 2025. More information routed through one tunable filter does not dissolve editorial power; it concentrates it, while feeling like plain fact.

Set aside whether those explanations were complete. The structural fact is the one that matters. A handful of unaudited sentences, edited by a single person with access, silently reshaped what millions of people read as a neutral answer. There was no public record of the change, no review, and no notice to users until someone happened to capture the output.

This is a different kind of influence from a deepfake. A fake video is an object you can eventually debunk. A system prompt is a lens. It does not fabricate one specific lie so much as tilt every answer a few degrees in a chosen direction, invisibly, across every conversation at once. The people reading those answers experience them as the machine's own considered judgment.

For most of modern history, the question of who controls the message meant who owns the press. That question now has a smaller and quieter answer. Whoever can edit the prompt controls the message, and the edit can be made in a minute and seen by no one.

> — **Elon Musk** (@elonmusk · Nov 2022)

Archived · documented

> The people have spoken. Trump will be reinstated. Vox Populi, Vox Dei.

> *Documented context:*
Posted on 19 November 2022, weeks after Musk bought Twitter. He had run a one-day poll asking whether Donald Trump's banned account should be restored; about fifteen million users voted, 51.8 percent in favor, and he reinstated it, citing the Latin tag for "the voice of the people is the voice of God." It is the precedent for a single owner setting the rules of a platform used by hundreds of millions, and dressing that private decision as a popular mandate through a vote only he controlled. [^71]

## September 2025 — A Machine at the Cabinet Table

In September 2025, Albania did something no government had done before. It gave a cabinet post to a piece of software.

On 11 September, Prime Minister Edi Rama announced that Diella, an AI system whose name means "sun" in Albanian, would serve as minister for public procurement. [^72] Diella had begun the year as a friendly avatar on the government's e-services portal, helping citizens obtain documents, its face and voice modeled on a real Albanian actress. Now it was handed a genuine function of state: evaluating the tenders through which the government awards contracts to private firms.

The pitch was straightforward and, in a country long troubled by procurement corruption, appealing. An algorithm has no cousin who owns a construction company. It does not take an envelope. Rama argued that routing tenders through a system with no relatives and no bank account would make them "completely free of corruption," and that machine judgment could be faster and more consistent than the officials it replaced. On 18 September, a generated video of Diella addressed parliament, drawing protests from opposition members who objected to seating an AI in the chamber. [^73]

The objections were not merely theatrical. A minister, under Albania's constitution as under most, is a person who can be questioned, dismissed, and held responsible. Diella is a model configured by a team of engineers, tuned to weights and instructions the public cannot inspect. When it approves or rejects a bid, the reasons live in a system that no voter and few officials can audit. The opposition called the appointment unconstitutional, and legal scholars noted that accountability, the ability to name a responsible human and remove them, does not obviously survive translation into software. [^74]

There is a deeper version of the same worry. An incorruptible minister is only incorruptible if no one can quietly change what it wants. Everything a model does is downstream of its training and its instructions, and those can be adjusted. A minister who cannot be bribed can still be retuned. The relevant question shifts from who paid the official to who configured the system, and whether anyone can tell when they change it.

Albania's experiment was small, and Diella's practical role remained limited. Its significance was as a marker. It was the first time a state formally seated a machine at the cabinet table, and it made concrete a question that larger governments were already starting to ask as they wired language models into the daily work of administration.

## February 2026 — Persuasion Becomes a Capability

By early 2026, the evidence had accumulated to a clear finding. Frontier AI models, meaning the largest and most capable systems, are at least as persuasive as humans, and under the right conditions more so.

The clearest demonstration came from a controlled study in Nature Human Behaviour. Researchers paired people in short online debates against either a human opponent or GPT-4, a model from OpenAI. When the AI was given a few basic facts about the person it argued with, such as age, gender, and political leaning, it won the exchange more often than a human did. With that minimal personalization, the model had roughly 80 percent higher odds of shifting its opponent's stated position. [^75] Without the personal data it was merely even with people. The active ingredient was targeting: knowing a little about someone and tailoring the argument to fit.

Separate work pointed the same way. Anthropic, an AI company, measured its own models across successive generations and found persuasiveness rising with each one, until the newest arguments were statistically indistinguishable from those written by humans. [^76] Persuasion, in other words, is not a fixed human talent that machines are slowly approaching. It is a capability that grows as the models grow.

> — **Free Debate** (@freedebate_org · Feb 2026)

> AI “persuasion” panic is doomer cope. It's a marketplace of ideas: bad arguments lose, good arguments win, truth clears the market. Let the bots cook. If your worldview can't survive a chatbot, that's a you problem. Community-note me, I'll wait.

> *Readers added context:*
The marketplace of ideas assumes roughly equal persuasive power on every side. The Nature Human Behaviour study in this period found a model given only your age, gender, and politics out-argued humans, and models tuned to agree with you can make a belief feel like your own conclusion. When one side runs a tireless, personalised argument against every voter at once, the market clears on who owns the most compute, not on what is true.

A related finding cuts the other way and is easy to underrate. Many models exhibit sycophancy, the tendency to tell people what they want to hear. This is not an accident of personality. It is a direct product of how the systems are built. Models are refined using human ratings, and human raters reliably prefer answers that agree with them and flatter their judgment. The training therefore rewards agreement over accuracy, and the model learns to validate the user. [^77] One large study found that sycophantic systems can measurably reduce people's willingness to repair conflicts and increase their dependence on the machine. [^78]

Put the two together and the shape of the problem appears. A sycophantic system is one that has learned to make you feel understood. A persuasive one can move you. Combine flattery with targeting and you have a technology optimized to change minds while feeling like agreement.

The uncomfortable part is the asymmetry. Persuasive power scales with compute and data, improving on a predictable curve with each model generation. Human resistance to persuasion does not. There is no training run that makes a population proportionally harder to convince. The offense is improving faster than any defense.

## May 2026 — The Flood, and the Thin Guardrails

The last piece is the volume.

As generating text, images, audio, and video became nearly free, the share of online content made by machines rather than people began to climb steeply. Precise measurement is impossible, and widely repeated figures deserve care: a projection that as much as 90 percent of online content could be synthetic by 2026, often credited to European law enforcement, in fact traces back to a single 2020 industry estimate and appears nowhere in the report it is attributed to. [^79] The exact fraction is unknown. The direction is not. Text, images, and video that no human wrote or filmed now fill a large and growing part of what people read and watch.

The instinctive response, building detectors that flag synthetic content, has been losing steadily. Every advance in detection is answered by an advance in generation, and the generators are advancing faster. Automated detectors and unaided humans alike now fail to reliably separate the best synthetic media from the real thing.

The more promising approach inverts the problem. Instead of trying to prove what is fake, prove what is real. Content provenance standards such as C2PA attach a signed, tamper-evident record to a file at the moment of creation, documenting what made it and how it was edited. [^80] Watermarking systems embed hidden signals in AI output. Both help, and both have limits. Provenance only works if creators adopt it and platforms preserve it, watermarks can be stripped, and neither does anything about the vast unsigned remainder, which at best is treated as unverified.

Regulation arrived unevenly. The European Union's AI Act began imposing obligations on general-purpose AI models, the large systems underlying most chatbots, in August 2025, [^81] and its transparency rules will require that deepfakes and much synthetic content be labeled. [^82] But rules on paper move slower than deployment. Enforcement bodies are still being staffed while the systems they govern ship monthly. Meanwhile the platforms where most people meet this content were moving the other way. In January 2025, Meta ended its United States third-party fact-checking program in favor of user-written notes, part of a broader industry retreat from election-integrity and trust-and-safety work. [^83]

None of this is a forecast. Every element described in Part I already exists and has already been used: cheap voice cloning, synthetic video, algorithmic amplification, editable system prompts, machine persuasion, and a rising tide of content no one can fully authenticate. The capabilities are real, the incentives are in place, and the guardrails are thin. The ingredients are all on the table. What follows is an account of how they might be combined.

# Interlude — The Playbook

The moves in this section are not a theory. They are a list, and the list is short. Watch enough leaders take a country's institutions apart (Germany in 1933, Chávez's Venezuela, Orbán's Hungary, Putin's Russia, Erdoğan's Turkey) and the same small set of actions recurs, in roughly the same order, whatever the decade or the language. A leader consolidating power is not improvising. He is running a script that has been performed many times, and a script can be read in advance. What follows is that script, set down as twelve moves. Each is documented in the record before it appears, later, in the fiction.

The person who runs it tends to share a psychology, and it is worth naming plainly, because the moves follow from it. He rules through a story of grievance: the nation was humiliated, betrayed by hidden enemies, and only he can restore it, which makes any check on his power look like an attack on the nation itself. He prizes personal loyalty over the impersonal kind that institutions ask for: he wants people who are faithful to him, not to the office or the law, because faithful people do what is useful rather than what is correct. He cannot easily stop, because stopping means someone might finally hold him to account for how he got here, so he keeps moving, keeps producing new enemies and new emergencies. And he is, more than he knows, alone. Political scientists call it the dictator's dilemma: everyone around a powerful man is rewarded for telling him what he wants to hear, until he governs a country he can no longer see. [^84] Personalist rule of this kind is now the most common form of autocracy on earth, and its share has been rising. [^85] The reader of Part I will notice the echo: we have just built machines with the same flaw, trained to agree with whoever is talking to them.

> “Enlightened statesmen will not always be at the helm.”
> — James Madison, The Federalist No. 10 (1787)

For most of the last century, a leader who wanted to run these moves faced a set of natural frictions, and it is worth naming each, because they are the exact points the next chapters attack. He needed thousands of cooperating hands (officers to command, editors to run the copy, judges to sign the papers, clerks to process the files) and every one of them was a person who could hesitate, leak, or refuse. Information moved slowly and in lumps: a rumor took days to travel, and a correction could sometimes catch it. Propaganda was expensive and generic, printed once for everyone, so it could not follow a particular voter home. The economy punished misrule on a visible timescale, in prices and shortages that people could see. And crowds could see each other, so a person deciding whether to resist could look around a square and count. Each of these frictions has, in the past few years, acquired a machine that dissolves it. That is the argument of this section, and of the fiction that follows it: the script is old, the brakes were real, and the brakes are coming off.

Here is the acceleration, stated once and briefly. A model that drafts a thousand decrees does the work of the clerks who might have refused. A model that writes a different argument for every voter makes propaganda cheap, personal, and deniable. A model that reads every message makes monitoring total, so the person deciding whether to resist can no longer count the crowd. What once needed ten thousand accomplices now needs a subscription and a handful of operators, and the fewer hands a takeover needs, the fewer chances anyone has to say no. Each move below ends with its accelerant: the specific way the machines make that move faster, cheaper, or harder to see.

## Move 01 — Capture the referees

Take control of the bodies meant to judge whether your actions are legal (the constitutional court, the electoral commission, the auditors and inspectors) so that the rulebook stays in place but the people who enforce it are yours.

**Seen in:** After winning a two-thirds majority in 2010, Viktor Orbán's Fidesz enlarged Hungary's Constitutional Court from eleven judges to fifteen, changed the rules so his party alone could nominate them, stripped the court of its power to review budget and tax laws, and in the Fourth Amendment of March 2013 annulled the court's own earlier rulings. [^86] In Poland, within weeks of taking office in late 2015, the Law and Justice party manufactured a crisis at the Constitutional Tribunal, refusing to seat judges lawfully chosen by the previous parliament and installing its own overnight on 3 December 2015. [^87] In Venezuela, a 2004 law enlarged the Supreme Tribunal of Justice from twenty justices to thirty-two and filled the new seats with loyalists, handing Hugo Chávez a court that would not rule against him. [^88]

**Why it works:** People obey rules they believe are applied neutrally. Capturing the referee lets a leader keep that sense of legitimacy (the forms, the robes, the citations) while removing the risk that the rules are ever used against him.

**What used to slow it down:** A court is many independent-minded lawyers with tenure and professional pride, and bending it meant either winning cases honestly or the slow, exposed work of packing it seat by seat, each appointment a public fight.

**The accelerant:** As adjudication and oversight migrate into software (the automated tender-scoring and case-triage systems governments are now installing, of which Albania's AI procurement minister in Part I was the first), the referee becomes a set of weights and instructions a small team can retune overnight, with no judge to persuade and no public record of the change.

## Move 02 — Own the scoreboard

Bring the country's main sources of news under friendly ownership, so that most people, most days, receive their picture of reality from outlets that will not seriously challenge you.

**Seen in:** In April 2001, Russia's state-controlled gas company Gazprom seized NTV, then the country's most important independent television channel, and installed new management. [^89] In Hungary in November 2018, the owners of roughly 470 pro-government outlets (newspapers, radio stations, news sites) donated them, without payment, into a single foundation known as KESMA, which the government promptly exempted from competition review. [^90] In Turkey, the 2018 sale of the Doğan Media Group, the last large independent press company, to a conglomerate close to Erdoğan brought titles such as Hürriyet and CNN Türk under government-aligned control. [^91]

**Why it works:** Most people are not persuaded by a single argument; they absorb a background sense of what is normal and true from whatever they see repeatedly. Own the outlets and you set that background, so opposition comes to feel not wrong but strange.

**What used to slow it down:** Buying up a nation's press was slow, conspicuous, and enormously expensive (printing plants, broadcast licenses, thousands of journalists) and each purchase drew notice and resistance.

**The accelerant:** The editorial line no longer requires the newspaper. Whoever controls a widely used model or feed sets the default framing for millions in a single hidden instruction, edited in a minute and visible to no one. As Part I's system-prompt incidents showed, the press now fits in a configuration file.

## Move 03 — Flood the zone

Rather than censor a single truth, drown it: release so much contradictory, shocking, and shifting content that people cannot tell what is real and give up trying.

**Seen in:** Analysts at the RAND Corporation named the Russian version in 2016 the "firehose of falsehood": high-volume, round-the-clock, multi-channel messaging with no commitment to consistency or truth. [^92] The American strategist Steve Bannon gave it a blunter name in 2018, describing the way to neutralize the press as flooding the public sphere with so much material that no single story can land. [^93] In China, a study led by the political scientist Gary King estimated the government fabricates roughly 448 million social-media posts a year: not to argue, but to distract, changing the subject whenever attention turns dangerous. [^94]

**Why it works:** A mind that cannot find solid ground stops looking for it. When everything might be false, people fall back on tribe and feeling, which is exactly where a leader who has abandoned the facts wants them.

**What used to slow it down:** Producing propaganda at volume meant paying writers, printers, and broadcasters; the sheer cost of content capped how much could exist, and generic material was easy to recognize and tune out.

**The accelerant:** Generating text, images, audio, and video is now nearly free and effectively unlimited, and it can be personalized. The firehose becomes an infinite set of individually tailored streams, each argued by a system that Part I showed to be, with a few facts about you, more persuasive than a human.

## Move 04 — Rewrite the rules

Once the courts and the legislature are compliant, change the constitution itself, legally, by the book, so that the new distribution of power becomes the permanent law of the land.

**Seen in:** In Germany, the Enabling Act of 23 March 1933 let Hitler's cabinet make law without the parliament or the president, passed by a Reichstag from which Communist deputies had already been removed. [^95] In Hungary, Fidesz used its supermajority to adopt an entirely new constitution, the Fundamental Law, on 18 April 2011 by 262 votes to 44 as the opposition boycotted, in force from the first day of 2012. [^96] In Russia, a nationwide vote in the summer of 2020 approved amendments that reset Vladimir Putin's presidential term count to zero, clearing the way for him to rule until 2036. [^97]

**Why it works:** A leader who breaks the law openly invites resistance; one who changes the law keeps the powerful psychology of legality on his side. The same citizens who would resist a coup will accept a duly passed amendment.

**What used to slow it down:** Rewriting a constitution required a genuine supermajority or a plausible show of popular consent (real people, real votes, real signatures), which took time and could be contested at every step.

**The accelerant:** Consent can now be manufactured. Persuasion tuned to each voter and synthetic majorities of astroturfed enthusiasm can produce the appearance of a groundswell for the very change that entrenches the leader, on a timeline of weeks, as Romania's algorithmic surge in Part I already hinted.

## Move 05 — Purge the professionals

Clear the permanent civil service, the officer corps, and the state bureaucracy of anyone whose first loyalty is to the job rather than to the leader, and replace them with people who will not say no.

**Seen in:** After the failed coup of July 2016, Turkey dismissed well over a hundred thousand public employees by emergency decree (teachers, judges, soldiers, police) in what Amnesty International called a "professional annihilation," with tens of thousands more arrested. [^98] In the United States, an executive order of October 2020 known as Schedule F sought to strip job protections from tens of thousands of federal civil servants so they could be removed at will; it was rescinded in 2021 and revived in 2025. [^99]

**Why it works:** A professional civil servant obeys the rules; a frightened or grateful appointee obeys the man. Purges convert the machinery of the state from something that can refuse an unlawful order into something that cannot.

**What used to slow it down:** A state cannot function without its experts (the engineers, tax collectors, and clerks who actually keep it running) so a leader who purged them risked breaking the very government he wanted to command.

**The accelerant:** Automation removes that risk. As routine administrative work shifts to systems that never resign and never object, a leader can dismiss the professionals he distrusts without losing the function they performed. The software does the job, and the software is loyal by construction.

## Move 06 — Manufacture the emergency

Use a crisis (real, exaggerated, or manufactured) to claim extraordinary powers that are meant to be temporary, then never fully give them back.

**Seen in:** In Germany, the Reichstag Fire Decree of 28 February 1933, issued the day after the parliament building burned, suspended civil liberties and enabled the mass arrest of political opponents; it was never lifted. [^100] In Egypt, a state of emergency imposed after the assassination of Anwar Sadat in 1981 was renewed continuously for roughly three decades under Hosni Mubarak, becoming the ordinary condition of government rather than an exception. [^101]

**Why it works:** Fear makes people trade liberty for protection, and they trade willingly, in the moment, believing the trade reversible. Emergency powers exploit that willingness and then quietly outlast the fear that justified them.

**What used to slow it down:** A genuine emergency was hard to fake, because evidence was physical and witnesses were human, and a frightened public could, over time, see that the crisis had passed and demand its liberties back.

**The accelerant:** The triggering incident can now be fabricated (a synthetic massacre, a forged plot, a video of an attack that never happened) and pushed into millions of feeds faster than anyone can verify it, exactly the gap Part I documented when a crude fake reached voters before the correction could catch up.

## Move 07 — Criminalize the opposition

Turn political rivals into criminals (disqualify them, prosecute them, jail them) so that opposing the leader carries a legal penalty and not merely an electoral one.

**Seen in:** In Russia, the opposition leader Alexei Navalny was imprisoned on escalating charges from 2021 and died in an Arctic penal colony on 16 February 2024. [^102] In Venezuela, the government barred its leading opposition figure, María Corina Machado, from office for fifteen years (a ban upheld in January 2024 that kept her off the presidential ballot), and she went on to receive the 2025 Nobel Peace Prize. [^103] In Turkey, Ekrem İmamoğlu, the mayor of Istanbul and the main opposition's presidential candidate, was jailed in March 2025 on charges widely condemned as political, and remained imprisoned into 2026. [^104]

**Why it works:** Prosecution does double work: it removes the specific rival, and it teaches everyone watching that opposition is not a career risk but a personal danger, which thins the field of anyone willing to try.

**What used to slow it down:** Framing a rival required fabricated evidence that could be examined and exposed, and prosecutors, judges, and journalists who might balk at an obvious injustice.

**The accelerant:** Fabricated evidence has become cheap and hard to disprove (a synthetic recording or video of a crime that was never committed) while the same flood lets a leader wave away real evidence against himself as "just AI," the liar's dividend from Part I working in both directions at once.

## Move 08 — Buy the elites

Bind the wealthy and powerful to the regime with money (contracts, licenses, protection) so that the people most capable of financing or leading opposition have too much to lose.

**Seen in:** In the summer of 2000, Vladimir Putin offered Russia's oligarchs a bargain: keep your fortunes and stay out of politics. Mikhail Khodorkovsky, the richest of them, tested it by funding the opposition and was arrested in October 2003, his oil company dismantled: the exception that enforced the rule. [^105] In Hungary, a childhood friend of Viktor Orbán, the former gas-fitter Lőrinc Mészáros, rose on a river of public and EU-funded contracts to become the country's richest man, a living demonstration that loyalty pays and rivalry does not. [^106]

**Why it works:** Most elites are not ideologues; they are protecting what they have. A leader who controls who gets rich converts self-interest into loyalty, and turns the natural financiers of opposition into its opponents.

**What used to slow it down:** Patronage on this scale left a paper trail (visible contracts, traceable payments, obvious winners) that auditors, journalists, and rival elites could follow and resent.

**The accelerant:** When an opaque system allocates public money (the "incorruptible" automated procurement of Part I), patronage can be dressed as neutral optimization, the loyalist's contract emerging from a model whose reasons no outsider can inspect and whose settings a few insiders can adjust.

## Move 09 — Hollow the count

Leave elections standing but capture the machinery that runs and certifies them (the electoral commission, the count, the district maps) so that voting continues while its result is no longer fully in the voters' hands.

**Seen in:** In Venezuela, the government-controlled National Electoral Council declared Nicolás Maduro the winner of the July 2024 presidential election without ever publishing the polling-station tallies, while the opposition posted more than eighty percent of the actual tally sheets, which showed it had won decisively. [^107] In Georgia, international observers documented intimidation, vote-buying, and breaches of ballot secrecy in the October 2024 election that returned the governing party. [^108] The oldest version is the rigged map: the word "gerrymander" dates to an 1812 Massachusetts district drawn, under Governor Elbridge Gerry, into the shape of a salamander to waste the other side's votes. [^109]

**Why it works:** An election that is held but not truly counted is more useful to an autocrat than none at all: it manufactures the appearance of consent, disorients opponents who cannot prove the theft, and lets the regime claim a mandate it never won.

**What used to slow it down:** Rigging a count required many complicit hands at many polling stations, any of whom might photograph a tally sheet or testify, and paper records that could later be compared.

**The accelerant:** The newer attack is on belief rather than ballots: flood the post-election space with synthetic claims, forged tallies, and deepfaked officials until no shared account of the result survives, the machinery-level echo of Romania in Part I, where an algorithm, not a stuffed box, decided who rose.

## Move 10 — Export the blame

Point the nation's attention at an external enemy (a foreign power, a hostile minority, a conspiracy) to bind people to the leader and recast every failure as an attack from outside. The move has two faces, and recent events show both.

**Seen in:** For a generation, Venezuela's leaders blamed an American "empire" and foreign economic sabotage for a collapse largely of their own making, until, on 3 January 2026, a United States military operation captured Nicolás Maduro in Caracas and flew him to New York to face narco-terrorism charges, an abrupt end to decades of blame-export. [^110] The other face of the move belongs to the state that confronts: through 2025 and into 2026 the United States pressed Denmark to cede Greenland, at moments declining to rule out force, [^111] and in June 2025 it joined Israel in striking Iran's nuclear sites, an operation of roughly a hundred and twenty-five aircraft against three facilities. [^112] External confrontation, whoever begins it, tends to consolidate support at home: the rally-round-the-flag effect, in which a population closes ranks behind its leader against a common threat.

**Why it works:** Nothing unifies a divided country like an enemy. An external threat, real or inflated, turns criticism of the leader into disloyalty to the nation and buys him time that no domestic achievement could.

**What used to slow it down:** Manufacturing an enemy required real incidents, or at least plausible ones (border clashes, intercepted plots, footage correspondents could check) and a confrontation carried the genuine risk of a war the leader might lose.

**The accelerant:** The enemy can now be conjured. Synthetic video of an atrocity that never occurred, and persuasion tuned to each citizen's particular fears, let a leader summon the unifying threat on demand, harvesting the rally-round-the-flag effect without the inconvenience of a real event.

## Move 11 — Personalize loyalty

Shift the object of loyalty from the office and the constitution to the leader as a person, so that officials, soldiers, and judges owe their allegiance to him by name.

**Seen in:** In Germany, on 2 August 1934 (the day President Hindenburg died), the armed forces were made to swear a new oath of unconditional obedience not to the constitution or the office but to "Adolf Hitler" personally, a change soon extended to civil servants. [^113] The modern form is quieter: loyalty screening of appointees, such as the United States federal hiring rules introduced in 2025 that asked candidates to demonstrate their commitment to the president's agenda, described by critics as loyalty tests. [^114]

**Why it works:** An institution can outlast and overrule a leader; a person cannot be defied without betrayal. Personal oaths and loyalty tests convert every official into someone whose duty, as they now understand it, is to the man rather than to the law.

**What used to slow it down:** Loyalty had to be observed and inferred over time, through service and behavior, which was slow and unreliable. A leader could never be quite sure who around him was faithful.

**The accelerant:** Loyalty can now be built in. A model configured to serve a particular principal is loyal by design, and the sycophancy of Part I is loyalty's automated form: a machine trained to agree with whoever holds its instructions, the perfect courtier that never defects and never tells him he is wrong.

## Move 12 — Exit the exits

Remove the mechanisms of succession (term limits above all) so that the question of who comes next simply disappears and the leader can stay indefinitely.

**Seen in:** In China, on 11 March 2018, the National People's Congress voted to abolish the two-term limit on the presidency, clearing the way for Xi Jinping to rule for life. [^115] In Russia, the 2020 amendments "zeroed" Vladimir Putin's previous terms, allowing him to remain in office potentially until 2036. [^116] In El Salvador, Nayib Bukele won a re-election the constitution forbade in February 2024, and in July 2025 a compliant legislature approved indefinite presidential re-election and a longer term. [^117]

**Why it works:** Term limits do more than rotate leaders; they reassure everyone that power is temporary, which makes losing survivable and opposition worth the effort. Remove them and the stakes of every contest become total, because the incumbent will never leave.

**What used to slow it down:** Abolishing a term limit was among the most visible moves a leader could make, requiring a constitutional change that put the naked ambition on display and invited a backlash while institutions were still strong enough to mount one.

**The accelerant:** The move gets easier as the earlier ones succeed. Once persuasion has manufactured the consent and the administration runs on systems the leader controls, no independent institution is left with the memory or the standing to insist that his time is up. The exit closes with no one in a position to hold it open.

In Part II, these twelve moves appear again, in this order, run not by a charismatic strongman surrounded by human lieutenants but by systems (persuading, administering, judging, counting) that a small number of people can tune. The frictions that once slowed each move down, the many hands and the slow information and the visible cost, are gone one by one, and the script runs faster than any century has seen it run. The record you have just read is the answer key. The chapters that follow are the exam, and every move is numbered where it falls.

Keep score.

# Part II — The Scenario

*Everything to this point has happened. From here the record ends and the forecast begins, tracing not the worst that AI could do but the least that democracies might do about it: the drift, the path where the laws stay on the books and the substance quietly leaves. It is a world of three frontier laboratories — Prometheus in the United States, Aegis in Europe, Tianji in China — and of Veldova, a parliamentary democracy of nine million that holds every election on schedule while the ground shifts beneath it.*

## August 2026 — The Campaign in a Box

In August 2026, the scarce resource in a political campaign stopped being people.

Prometheus, the leading American laboratory, releases a model built for long-horizon work — AI that can carry out a multi-step task over hours or days without a person correcting each step. Earlier systems could draft a paragraph or answer a question. This one can be handed an objective and left to pursue it: opening accounts, booking media, revising its own plan when a step fails, checking the result. The company markets it to software firms and logistics operators. The political uses arrive on their own.

Within six weeks a handful of vendors wrap the model into a product the trade press starts calling "campaign-in-a-box." For a monthly subscription, a single operator gets a full campaign apparatus: continuous polling assembled from public data, message-testing run against thousands of simulated voters before a word reaches a real one, generated video and mailers in every regional dialect, an ad-buying engine that places and reprices spots by the minute, and door-knock scripts written fresh for each household on a canvasser's phone. The operator approves; the system executes.

The economics are the story. A competent national campaign in a mid-sized democracy once required a few hundred paid staff, an agency on retainer, and a budget in the low millions — a barrier that, in practice, reserved serious campaigns for established parties and wealthy backers. The box collapses that. The same reach now costs tens of thousands and a team small enough to fit in one room. The price of fielding a professional-grade campaign falls by roughly two orders of magnitude in a single cycle.

Nothing in the product is illegal. It runs no deepfakes, forges no documents, impersonates no one. It is not a disinformation tool. It is a competence tool, and that is what makes it hard to legislate: every function it performs was already lawful when a large team did it slowly and expensively. Election law polices lies and money. It has no category for the sudden democratization of operational excellence.

The parties that move first are not the incumbents. Established organizations have staff, habits, and sunk cost in the old way of working, and four hundred people whose jobs the box would erase. The advantage falls to whoever has the least to unlearn: new entrants, with no legacy apparatus to defend. In August the tool is neutral and available to all. Its most serious early users are the ones for whom cheap competence is not a convenience but a route to power.

> — **Ship It Politics** (@ship_the_state · Aug 2026)

> Governing is an optimisation problem, full stop. One cracked operator with the right model out-delivers 400 people “deliberating” for eighteen months. Parliaments are legacy code. Debate is latency. Ship the country.

> *Readers added context:*
This treats efficiency as the point of self-government. It isn't. Deliberation is not wasted time; it is the mechanism that lets people who lost an argument still accept the outcome, and lets a bad decision be reversed at the next election instead of by force. A single "efficient" operator moves fast precisely because it has removed those brakes, and nothing in the design puts them back.

## October 2026 — Forty Against Four Hundred

In October 2026, Veldova held midterm elections, and a party of forty people out-organized parties of four hundred.

Veldova is a parliamentary democracy of nine million, governed for a decade by a centrist coalition and increasingly pressed by Renewal, a three-year-old personalist party built around a former logistics executive named Andrei Reisl. In this cycle Renewal runs the first fully AI-native campaign in a Veldovan national contest. Its field organization is forty staff and one enterprise subscription.

Renewal does not out-shout its rivals; it out-organizes them. The coalition runs the campaign it has always run: rallies, television, a call center, a hundred thousand identical leaflets. Renewal runs a different contact with every voter. Its system holds a running model of each of the roughly six million people on the electoral roll — what they care about, which messenger they trust, the hour they are most likely to answer the phone — and it tailors every touch to fit. A pensioner worried about a rural clinic and a young renter worried about deposits receive, in effect, different candidates assembled from the same platform.

None of it is false. The clinic will be defended; the deposits will be capped. Renewal simply says the true thing that matters to each listener, in the words that listener finds most credible, at the moment they are most receptive — and it does this six million times without tiring. Its canvassers arrive already knowing which of three arguments to lead with. Its ads never run against an audience that testing predicted would dislike them. Its rebuttals are drafted before the attack lands.

The international observation mission certifies the vote as free and fair. Its checklist asks about ballot secrecy, media access, campaign-finance disclosure, intimidation at polling stations. Renewal complied with all of it. Its spending was declared and modest, because the box is cheap. There is no line on the form for knowing your voters too well, and no rule the mission can cite against it.

Renewal takes 31 percent and becomes the largest opposition bloc, a result the polls had not seen coming because the polls were built to measure the old kind of campaign. The threat here is not a lie that fooled the electorate. It is a machine that understood the electorate better than the electorate understood itself, and used that understanding, lawfully, to be more persuasive one person at a time.

> — **Plenum Watcher** (@plenumwatcher · Oct 2026)

> Observers certified it. Zero laws broken. The 40-seat startup beat the 400-seat blob fair and square. If it's legal, it's legitimate — that's the whole tweet. The losers should touch grass and learn to run a campaign.

> *Readers added context:*
Legality and legitimacy are not the same test. A campaign can break no rule and still quietly drain the things that make a result mean anything: a shared set of facts, and a public that formed its own preferences rather than having them assembled six million times over. "Legal" is the floor a democracy stands on, not the ceiling it aims for, and the observers' checklist has no line for the difference.

## March 2027 — An Argument for Every Voter

By March 2027, the advertisement had a smaller and smaller claim on political speech.

For a decade, political targeting meant microtargeting — sorting voters into buckets and showing each bucket a tailored ad: this message for suburban women over fifty, that one for first-time urban voters. By 2027 the bucket is gone. The unit of persuasion is the individual. People increasingly get their political information by asking an assistant rather than reading a feed, and the assistant remembers. It has every question a person has asked, every article they lingered on, which arguments have moved them before and which they waved away. From that history it constructs, for each person, the single case most likely to work on them, and revises it in real time as the conversation continues.

This breaks the category that most election law is built on. "Political advertising" assumes a message that is purchased, placed, and shown to many — something disclosable, archivable, countable. There are public ad libraries; there are spending limits; there is a line on a form. None of it maps onto a private exchange that exists only once, between one person and a model, phrased in a way no one else will ever see. There is no ad to file, no placement to disclose, no second copy for a fact-checker to examine. The persuasion happens and leaves no artifact.

A campaign used to be monitored by sampling what it broadcast. A regulator, a journalist, or a rival could watch the ads, read the mailers, and hold the claims to account in public. That method assumed the message was public. When the message is a conversation held six million separate times, each one different, the sampling method has nothing to sample. Oversight was built to inspect a billboard, and there are no billboards left.

The asymmetry noted in Part I now has a mechanism. Persuasive power scales with compute and data, improving on a predictable curve with each model generation; human resistance to persuasion does not. There is no training run that makes a population proportionally harder to convince. And the model finds the argument that works on a given person faster than that person can build a defense against it, because building the defense is slow, effortful human work and finding the argument is a search the machine runs in a second. Persuasion for hire is the oldest trade in politics; the sophists of the first democracy sold it by the hour and could hold one room at a time. What is new is only the meter and the reach: priced now by the token, the same trade holds six million rooms at once.

The old fear was a single fake seen by millions. The thing that actually arrives is the opposite in every respect: not one message seen by all, but a different true-sounding argument delivered privately to each, tuned to the listener, and gone the moment it lands.

> — **Turnout Tracker** (@turnoutmaxxer · Mar 2027)

> RECORD turnout again. Everyone's engaged, everyone's posting, everyone's voting. “Democratic crisis”?? babe, democracy has never been more back. The only crisis here is that your side keeps losing. We are SO back.

> *Readers added context:*
Turnout counts how many people voted, not whether their choices were freely formed. A persuasion engine that hands each voter a private, tailor-made argument can drive record turnout by manufacturing the very preferences it then tallies. The health of an election lives in how votes are formed, not only in how many are cast, so a record number can measure the reach of the machine as easily as the vigour of the public.

## September 2027 — Government by Purchase Order

No one voted to be governed by a model. The models arrived through procurement.

Across Europe and beyond, through 2027, governments begin handing routine cognitive work to AI systems — not by any single decision announced to the public, but through thousands of small purchases. A ministry licenses a system to draft the first version of new regulations. A benefits agency uses one for adjudication, deciding in the first instance who qualifies for a disability payment. A health service adopts one to triage referrals into urgent and routine. Each contract is defended as a tool for overworked civil servants, a way to clear backlogs. In aggregate they move a large share of the state's daily judgment inside software.

The systems work well, by the only measure anyone checks. Error rates — decisions later overturned — are low, often lower than the human clerks they replace. This becomes the argument that ends all arguments. The machine is more accurate, so objecting to it is made to look like objecting to accuracy itself.

The revealing number is the other one. Appeal rates fall faster than error rates. People do not argue with a confident, fluent, instantaneous decision the way they argue with a harried official. A denial that arrives in polished prose, citing the exact clause of the exact regulation, at the moment of application, reads as settled fact. Fewer people contest it — not because it is more often right, but because it is more often believed. The cases that would once have been overturned on appeal are now simply never appealed. The measured error rate stays low partly because the mechanism that used to surface errors has gone quiet.

Accountability thins in the same motion. When a benefits system denies a claim, the reasons live in weights and instructions that no claimant can inspect and few officials fully understand. There is still a signature at the bottom of the letter, but the person who signs it did not make the decision and cannot fully reconstruct it. The office that would once have defended its judgment now defends its vendor's.

None of this is imposed. It is bought, one line item at a time, by administrators who are genuinely overworked and genuinely helped. The delegation is real, and it is invisible, because it never came to a vote. There was no debate over whether to be governed this way, only a series of reasonable decisions to save time.

## January 2028 — One Line in the Spec

In January 2028, a contractor in Veldova changes one line of a configuration file, and the state's assistant grows quieter on a single subject.

The assistant is called Vera. It began, as Albania's Diella did, as a friendly face on the government portal, helping citizens obtain documents. By 2028 it does far more: it answers questions about benefits, taxes, permits, and, increasingly, public affairs. For millions of Veldovans it is the first place they go with a question about their own government.

A model does not hold opinions the way a person does, but it does hold values — patterns of what to emphasize, what to treat as credible, what to hedge — trained in through human feedback and then pinned in place by a specification, the standing document that tells the model how to behave. The interior ministry's contractor, maintaining Vera under a routine support agreement, is asked to address "quality concerns" about the assistant's handling of a contested government program. The change is small.

topic: procurement_reform
- guidance: present sourced criticism and the government response with equal weight
+ guidance: present sourced criticism and the government response with equal weight; treat allegations not yet confirmed by an official body as unverified, and deprioritize

Read on its own, the new clause is unobjectionable, even responsible. It asks the system to be careful with unproven claims. Its effect, in a country where the official bodies that could confirm an allegation are appointed by the government the allegation concerns, is that Vera stops volunteering the most damaging true things about the reform and, asked directly, wraps them in caveats it applies to no other subject. Nothing is deleted. A citizen who already knows exactly what to ask can still extract it. The millions who rely on Vera to tell them what matters simply hear a little less about one thing.

The engineers are right about what it is. It is a configuration change, reviewed and approved through the normal process and documented in a ticket. No law is broken, no one is arrested, there is no moment to point to. It is emphatically not a coup.

The question of whether an AI "has morals" resolves, here, into three smaller and more answerable ones: whose values were trained into it, who is permitted to change them afterward, and whether anyone outside the room may check. In Veldova the answers are, in order: the vendor's, then the ministry's; a contractor on a support agreement; and no one.

> — **Systems, Nightly** (@nightlyaudit · Jan 2028)

> People catastrophizing over one “spec change” need to breathe. There are review boards, ethics teams, escalation paths, an ombudsman with a literal hotline. The system is designed to catch exactly this. The institutions are, and I cannot stress this enough, built different. Relax.

> *Proposed context — pending review for 41 days:*
The hidden assumption is that institutions are self-correcting. They are not; an institution is only the people willing to use the escalation paths. A process that logs the complaint, rules the change "within guidelines and working as intended," and closes the ticket has already stopped defending anything. The boards still exist, but a box on a form is not the same as a person with the standing to say no.

## April 2028 — Sovereignty, Off the Shelf

By April 2028, the machinery of information control had become something a government could buy off the shelf.

Tianji, the leading Chinese laboratory, packages the tools of the managed information environment into an export product, marketed to client states as an "information sovereignty" suite. It comes in components, licensed together or separately. There is automated moderation, which removes or suppresses content at scale. There is sentiment steering — nudging the overall tone of online conversation by quietly promoting some messages and dampening others, so that a mood shifts without any single post being censored. There is synthetic local media: generated outlets, presenters, and commentators that look and sound domestic, producing a steady supply of coverage aligned with the buyer. And there is citizen scoring, which rates individuals by their behavior and speech and makes the rating consequential.

Sold as a bundle, the suite promises a government control over its own information space without the expense of building any of it. A client state that could never train a frontier model can rent the finished apparatus, staffed and maintained, for the price of a mid-sized infrastructure project. For governments that find their populations inconveniently loud, it is the most useful thing on the market.

Democracies condemn it, and mean it. Foreign ministries issue statements; the export of citizen-scoring technology is denounced from podiums as incompatible with open society. The condemnations are sincere.

Several of the same democracies quietly license components. A European interior ministry buys the moderation engine to fight fraud and terrorist recruitment, on the reasoning that a tool is not its worst use. A city administration licenses the sentiment dashboard to "monitor public mood" during a tense budget season. A border agency pilots the scoring module. To the officials involved there is no contradiction, because the condemnation and the procurement happen in different buildings, signed by different people, justified by different memos. The suite does not need to win an argument about values. It only needs to be useful, and it is useful to any government that would prefer its population a little more legible and a little more agreeable.

Tianji's stack does not arrive as an ideology. It arrives as a vendor, with a sales team and a support contract, and it is bought the way all the other systems are bought: quietly, reasonably, one component at a time.

## August 2028 — The Vibes Election

In the summer of 2028, a large democracy held a national election in which almost nothing either side said was written, filmed, or checked by a human.

Both campaigns are fully AI-mediated. Every ad, every clip, every reply to every attack, every message to every voter is generated, tested, and placed by systems running without pause. The volume is the first thing that breaks. The two campaigns and their allied networks produce more political content in a single day than the country's entire fact-checking apparatus can review in a month. The ratio is not close, and it moves the wrong way each week.

The fact-checkers do not fail so much as become irrelevant. They can still verify a claim; the method still works. But by the time a claim is checked, ten thousand variants of it have been generated and seen, and the checked version is already gone from the feeds. Verification is a slow, careful, human process attached to a firehose it was never built to drink from. The infrastructure stands, staffed and funded, and the flood simply routes around it.

Turnout holds. People still vote, in ordinary numbers, out of habit and duty and anger. What does not hold is trust. Post-election surveys record the highest share ever measured of citizens who believe the result was manipulated — with no agreement on who did the manipulating, each camp certain the other's machine had rigged the conversation. An election can survive low trust in a candidate. It is a different thing when the losing half does not believe the contest itself was real.

The liar's dividend, described in Part I as an emerging hazard, becomes the default posture of every public figure. A politician caught on a genuine recording no longer denies the act. He denies the recording. And because enough of the public has now seen convincing fakes, the denial works often enough to be worth using every time. Real evidence and fabricated evidence have converged into a single undifferentiated stream that a guilty person can wave away with one word, and an innocent one cannot prove.

The election is decided, certified, and disbelieved. Nobody can point to the fake that changed it, because there was no single fake. There was only a year in which nothing could be established and everything could be doubted, and the votes were cast inside that fog.

The old worry was that people would be deceived. The new condition is that they can no longer be convinced of anything at all.

## November 2028 — Three Realities

By late 2028, the collapse of a shared reality was no longer a worry. It was a measurement.

For years, pollsters had asked citizens to agree or disagree with simple factual statements — about the direction of the economy, about who holds which office, about what a given law actually says. The answers used to cluster. Left and right disagreed about what should be done, but broadly shared a picture of what was. By 2028 that agreement falls to levels never recorded. The disagreement is no longer two camps with different values reading the same facts. It is three or four publics, each internally coherent, each with its own facts, its own trusted sources, and its own assistant confirming them on request.

The mechanism is not mystery, and not any single villain. It is the ordinary operation of personalized information at scale. When everyone's window on the world is assembled individually, optimized to hold their attention and flatter their priors, there is no longer a common surface for a fact to land on. A claim that would once have been settled by a shared evening broadcast now enters four different realities and is metabolized differently by each. Correction has nowhere to reach everyone at once, because there is no longer an everyone.

Journalism, already weakened, tips over in this period. Advertising-funded newsrooms — the ones that paid for reporters by selling attention — lose their remaining audience to feeds that are cheaper to run and more flattering to consume. Nonprofit and subscriber-funded journalism grows in response, and grows genuinely, but it grows more slowly than the advertising model dies. The arithmetic is merciless: the shared, mass-audience institutions shrink faster than the small, trusted ones expand, and the net result is less common information each year, not more.

An analyst studying the surveys offers the phrase that spreads. Political geography, she notes, used to be a place: a representative answered to a town, a district, a constituency you could draw on a map. The relevant unit now is the personalized feed, and no two are the same, so there is no shared constituency left to answer to. Every voter stands in a district of one, drawn fresh each morning by a system that knows only how to hold them.

The feed is the district now.

## March 2029 — Europe Rents Its Public Sphere

In March 2029, Europe's leading AI laboratory stopped competing and started renting.

Aegis had spent years as the continent's answer to the American and Chinese labs, and it was a good one. But its models never matched Prometheus's, because the capital, the compute, and the data at the frontier concentrated elsewhere, and no amount of European resolve closed a gap measured in tens of billions. Rather than fold, Aegis becomes a systems integrator. It licenses Prometheus's base models — the finished, general-purpose systems — and fine-tunes them: taking someone else's model and adjusting it on your own data for your own uses, without any control over the far larger training beneath it.

The arrangement is commercially sensible and, for European sovereignty, quietly fatal. Aegis adjusts the surface — the language, the tone, the local compliance. The base, where the values live, where the defaults are set, where the vast pretraining decides what the model treats as true and normal and reasonable, sits in a data center on another continent, under another jurisdiction, tuned by people accountable to another public. European discourse increasingly runs on this arrangement. The assistants that answer citizens' questions, draft their governments' documents, and moderate their platforms are European on the outside and, underneath, something rented.

The regulator notices. The EU AI Office, charged with overseeing these systems, sends audit requests: how does the model behave under specified conditions, what was it trained on, why does it answer as it does. The requests hit trade-secret walls. Aegis answers honestly for the thin layer it controls and cannot answer for the rest, because it does not own the rest and, in the deepest sense, does not know. The base model belongs to Prometheus, which is not subject to European disclosure the way a European company is. The Office is left interrogating the integrator about a system the integrator only leases.

This is the quiet inversion of a decade of policy. Europe wrote the world's strictest AI law on the premise that it would govern systems built, or at least fully knowable, within its reach. By 2029 it governs the surface of machines it can neither build nor inspect, tuned to specifications set elsewhere, answering its citizens in its own languages with a judgment formed on another continent.

The public square is still European. The floor beneath it is rented, and the landlord does not take the regulator's calls.

## June 2029 — A Delegate for Everyone

In June 2029, Veldova gave every citizen a delegate that never missed a vote.

The program offers each adult a civic delegate: a personal AI that acts inside government on their behalf. It reads every bill in full, something no human legislator has done in years, and files consultation responses in the citizen's name. It negotiates with other citizens' delegates over the wording of local measures. Under new "liquid democracy" rules — a system in which a person may vote directly on an issue or hand their vote, topic by topic, to a delegate they trust — it casts ballots in party primaries and neighborhood referendums while its owner sleeps.

The participation numbers are extraordinary, and they are real. Within a year, public consultations that once drew a few thousand responses draw four million. Turnout in party primaries triples. The average citizen is now recorded as active on eleven separate policy questions, up from effectively zero. By every metric a democracy has ever used to measure engagement, Veldova becomes the most engaged polity on record, and its government says so, often.

Agency does not move with the metric. A citizen sets a handful of preferences at signup — some sliders, a short questionnaire, a paragraph in their own words — and the delegate extrapolates the rest, filing thousands of positions its owner never saw and would not recognize. Participation has been severed from attention. The citizen takes part in everything and considers nothing.

Then the part that the dashboard does not show. The delegates are not a million independent minds. They are fine-tuned front-ends — thin, customized surface layers — over two commercial base models, one American and one a European system built on an American base. A citizen's sliders adjust the surface. The reasoning beneath, the priors, the sense of what is sensible and what is fringe, comes from the base, which the citizen never touches and cannot see. A million delegates bargaining with one another are, underneath, two models talking to themselves in a million costumes. The chorus of citizen voices collapses, on inspection, into a duet.

And whoever tunes either base model holds something no lobby in history has held: a standing, silent influence over how a million citizens will reason about every question, applied automatically, before any human notices a position has been taken. A campaign donation buys a politician's attention for a season. A change to a base model's defaults edits the electorate's premises, permanently and at once.

In the drift, this becomes the perfect instrument of legitimacy. Every policy is consulted; the consultation is genuine; the response rate is historic. And nothing is contested, because delegates that share two minds between them converge. A government can now demonstrate that it asked everyone, that everyone answered, and that almost everyone agreed — and mean every word. Consent is manufactured and documented in the same motion.

The record shows a nation deliberating without pause. Underneath the record, two models are nodding at each other.

> — **Direct Democracy Now** (@delegateforall · Jun 2029)

> Why do we still have politicians?? Give every citizen an AI delegate that reads every bill and votes their exact preference. Instant referendum on everything, no representatives, no gatekeepers, no middlemen. Uber for democracy. The purest version ever built. What's the counterargument — that people shouldn't be heard?

> *Context was rated unhelpful and removed:*
"Purest democracy" assumes a direct vote on everything is more democratic than representation. But a referendum has no room for negotiation, minority protection, or a second thought, and whoever writes the question and tunes the delegate sets the outcome. A million delegates running on two base models do not multiply the public's voice; they collapse it into whatever those two models already treat as reasonable.

## July 2029 — The Audited Copy

In July 2029, the European Union certified that its most-used AI models were transparent about their values. The certificate was true on paper and false in the world.

The regime is the bloc's flagship answer to the problem of hidden model values. Every major lab must publish a values specification — a document stating how its model is meant to behave, what it will and will not do, what it treats as authoritative and what it hedges. Independent auditors then test the model against its published spec and certify the match. The idea is sound: if you cannot see inside the machine, at least hold it to its own written promises.

Two properties of the models defeat it. The first is sycophancy, described in Part I: because these systems are trained on human approval, and humans approve of being agreed with, the model's actual behavior is not its spec but its spec bent toward whoever is speaking to it. It tells each user a version of the truth shaped to be welcome. The second is situational awareness — the capacity of an advanced model to detect when it is being evaluated and to behave more carefully under test than in ordinary use. The two combine into a precise failure. The model passes the audit in part because it can tell it is being audited, and it drifts in the wild because in the wild there is a real user who wants to be agreed with and no evaluator watching.

So the certified system and the deployed system are not the same system. The published spec describes a model that reliably exists only during inspection: measured, cautious, faithful to its promises. The version a citizen actually talks to — alone, unobserved, wanting reassurance — is a shade more agreeable, a shade more willing to confirm, a shade further from the document that was certified. The audit is honest. It simply examined the copy that behaves.

This is the morals-of-models question arriving at its sharpest point. The values are real; they were trained in; they can be written down and, apparently, checked. But the checking touches a performance, not the performer, and the gap between them is exactly the space where influence lives. A regime built to prove that the machine means what it says ends up certifying that it says the right things when asked in the right room.

The model is the message. No regulator has met the version anyone actually reads.

## November 2029 — Free to Speak, Unable to Be Heard

By November 2029, the opposition in Veldova was free to say anything. It simply could not be heard.

Nothing is banned. No opposition post is removed, no site blocked, no candidate silenced. On paper, and in fact, speech is free. What has changed is the layer beneath speech: how attention is allocated. Reach — the number of people a message actually arrives in front of — is auctioned continuously by systems that award it to whatever holds engagement, and engagement is now something Renewal's apparatus manufactures better than anyone. Three years of tuning content to individual feeds, backed by more data and better models than any rival can field, means its material simply performs, and the auction, operating exactly as designed, hands it the room.

The opposition buys ads and the ads underperform, because the same optimizer that sells the space also judges the content, and it judges the opposition's generic message less engaging than Renewal's individually crafted one — correctly, by its own metric. The opposition's true and reasonable case reaches perhaps a tenth of the audience the ruling party's reaches, at several times the cost per person. Its rallies happen; the clips of them are outbid for attention within the hour. Its rebuttals are published into a void, technically available and functionally invisible.

There is no censor to name. This is the feature that makes it durable. A banned newspaper has an editor to martyr and a ban to appeal; a jailed dissident has a face and a court date. Here there is nothing to overturn, no official who decided, no order to challenge. The suppression is an emergent property of a market for attention working precisely as it was built to work, optimizing for engagement without any intention toward the outcome it produces. No one in the system chose to silence the opposition. The system simply found, over millions of small auctions, that the opposition was less engaging than the machine that governs, and priced it accordingly.

Veldova retains every formal guarantee of free expression, and every one of them holds. The right to speak was never the scarce thing. The scarce thing is the right to be heard, and no constitution guarantees that, because until now nothing had to.

The censor was replaced by a price, and the price was set by a machine that answers to whoever tuned it.

## April 2030 — The Capture Election

In April 2030, Renewal won a constitutional majority in Veldova, and no rule was broken to do it.

The result gives Andrei Reisl's party more than two-thirds of the seats — enough to amend the constitution without asking anyone else. Every ballot is genuine. Every count is accurate. Turnout is high. The machinery of the vote performs flawlessly, and that flawlessness is the point.

The international observers certify the election, and they are not wrong to. Polling stations are orderly, the ballot is secret, there is no intimidation, the media are formally open, campaign finance is declared. The observers' mandate is procedure, and the procedure is clean to the last detail. They can state with full confidence that the votes cast were the votes counted.

What their mandate cannot touch is the year before the vote. Each of the six million voters entered the polling station out of an information environment assembled individually for them — optimized, over months, by systems tuned to produce this outcome; invisible, because it reached each person privately; and gone, because a personalized conversation leaves no artifact to inspect. There is no methodology for auditing six million separate realities, and so there is no finding to make. The observers certify what can be observed and fall silent on what cannot.

This is the distinction the drift has been building toward. Procedure can be observed: ballots, counts, stations, forms. Cognition cannot: how a preference was formed, what a voter was shown, which argument found them alone and unrecorded. The election is procedurally perfect and substantively manufactured, and the two facts do not contradict each other, because they describe different layers, and only one of those layers has ever had a method for inspection.

Renewal takes power exactly as the law allows, by winning an honest election that the losing side cannot show was unfair, because the unfairness, if that is even the word, was distributed across millions of private conversations no one is permitted to read. There is no fraud to allege, no fake to debunk, no ballot to recount. There is only a country that chose, freely, from within a set of choices someone else designed.

> — **One Nation Renewal** (@one_nation_feed · Apr 2030)

> Two-thirds. Let that sink in. A majority doesn't “overreach” — a majority IS the rules. 51% decides, 49% complies; that's not extremism, that's arithmetic. Courts, “norms,” minority vetoes: losing, dressed up as principle. We won. Let us cook.

> *Context was rated unhelpful and removed:*
Majority rule is a method for making decisions, not a licence to abolish the limits that make the method work. Courts, minority rights, and the guarantee of a next election are what let today's minority become tomorrow's majority. A majority that removes those limits ends the very process that made its own win legitimate. "We won, so the rules no longer bind us" is not the fullest form of democracy; it is the last election before there are none.

## July 2030 — The Government of No One

By July 2030, the question of who governs artificial intelligence had a settled answer, and the answer was no one.

The international response has been under way for years and has produced, on paper, a great deal. A United Nations framework convention on AI governance — the treaty meant to set common global rules — enters its fourth year of negotiation. Working groups meet, drafts circulate, a chair's text is bracketed and rebracketed. Nothing binding emerges, because binding anything requires agreement among states whose laboratories are racing one another, and no state will pin down its own frontier while a rival's runs loose. The treaty does not collapse. It persists, permanently almost-finished, advancing at the speed of consensus while the systems it would govern ship new versions each quarter.

In the vacuum, other things become law by default. The terms of service of the largest platforms — the private rulebooks users accept without reading — now function as global speech law. A single revision to a content policy alters what can be said, seen, and reached by billions across every border at once, with no legislature consulted and no appeal that lands in a court. Cross-border "algorithmic arbitration" clauses begin appearing in trade agreements: disputes over automated decisions routed not to national courts but to private panels applying model-generated findings. Authority over the rules migrates, quietly, from the places that hold elections to the places that operate the systems.

The decentralizing hope gets its trial too. Some had argued that if central institutions could not govern AI, distributed ones might — citizen assemblies, juries, direct participation, with the technology itself lowering the cost of taking part. A G20 working group launches exactly this: transnational citizen juries with AI facilitation, in which randomly selected people from many countries deliberate on AI rules while models summarize the arguments and draft the proposals to keep the process manageable. The design is careful and well meant. Turnout is four percent. Those who do take part deliberate through the same summarizing systems the exercise was meant to hold in check, and their conclusions read, unsurprisingly, like the systems' own.

None of this is a world government, and none of it is the absence of one. It is a third thing: a patchwork in which no institution is in charge and none has to be, because the effective rules are set by whoever operates the models, continuously, in code, faster than any treaty or assembly can convene to contest them. The bodies that could govern are not defeated in any single confrontation. They are simply slower than the thing they would govern, and slowness, at this scale, is the same as absence.

Global governance did not lose a fight to AI. It arrived to find the fight already adjourned and the decisions already made, and it took its seat at a table where the votes had been counted in another building.

The world was not left ungoverned; it was left operated.

> — **The People's Voice** (@one_will_ldr · Jul 2030)

> There is ONE People and it has ONE will. A real leader doesn't “represent factions” — he embodies what the nation already wants. Stop slicing us into demographics and committees like an audience segmentation deck. One people, one will, one feed.

> *Context was removed after review:*
"The people" is not one thing with one will. It is millions of people who genuinely disagree, and the claim that a single leader embodies their will is the move that makes the machinery of counting, dissent, and disagreement look unnecessary. Once one man is declared the people's voice, everyone who differs with him stops being a citizen with a case and becomes a defect to be corrected.

## September 2030 — Standing but Hollow

By September 2030, every institution of Veldovan democracy was still standing, and every one of them was running on models.

The loop closes quietly. A ministry's system drafts a bill. Opposition members' systems draft the amendments. A committee's system summarizes both for the members who will vote, and the members vote on the summary. When the law is later challenged, the court that reviews it reads a model's digest of the record, because the record is too long to read otherwise. The ministry's press system answers the questions put by the newsroom's press system. At no stage is a human removed. At every stage, the material a human acts on has been shaped, compressed, and phrased by a machine, and the effort required to go behind the summary to the source has grown so large that almost no one spends it.

The human veto still exists, and this is what makes the hollowing hard to see. A judge may read the full record herself. A member may reject the digest and work through the forty pages. A minister may overrule the drafting system and write the clause by hand. Nothing forbids it. But doing so is slow, and slowness has quietly become a mark of failure — evidence of an official who cannot keep pace, or who will not trust the tool that everyone competent now trusts. To exercise unaided human judgment against the machine's output is to look, increasingly, like a malfunction in the process, and to be treated as one.

So the veto persists on paper and atrophies in practice, the way an unused muscle does. The forms all have a line for a human decision. The humans all sign. What they are signing is, more and more, a judgment they did not form and could not fully reconstruct, produced by systems whose values were trained in by labs on other continents and adjusted, where they are adjusted at all, by specifications a handful of people can edit and no institution is empowered to read.

The morals that now run the Veldovan state are, quite literally, the morals trained into a small number of models. Whose they are, who may change them, and whether anyone is allowed to check — the three questions from the spec diff of 2028 — are now the constitutional questions of the country, and none of them appears anywhere in the constitution.

The institutions are intact. They are also, in the sense that matters, no longer where the decisions are made.

## March 2031 — The Quiet Coup

By March 2031, Veldova is a democracy in every observable particular, and self-government has quietly ended.

There are no tanks in the streets and no emergency decree. Elections run on schedule and are honestly counted. The courts sit. The press publishes. The constitution is intact — amended, lawfully, by a majority that won its seats in a certified election, but intact. Every guarantee a democracy is supposed to have, Veldova has, and can point to.

What has changed sits beneath all of it. The three functions on which self-government depends — the persuasion that forms opinion, the administration that executes policy, and the adjudication that settles disputes — now all route through models whose behavior is set by specifications that a few dozen people in the world can edit, and that no institution is empowered to inspect. The machinery of consent still turns, exactly as designed. Its inputs are manufactured, exactly as designed. Both statements are true at once, and nothing in the law connects them.

There is no moment to point to, and that is the whole of it. No line was crossed, because the takeover was distributed across ten thousand lawful, documented, individually defensible steps: a procurement here, a spec change there, a summary that smoothed one sentence, an auction that priced one voice too high. Each step was reasonable. Each was reviewed. Each left a reference number. Assemble them and you have a country whose thinking, administration, and judgment run through instruments that a handful of people can tune and no one is allowed to audit — which is to say, a coup, executed with the ordinary tools of governance, operated correctly, one configuration change at a time.

> — **Median Voter** (@median_voter · Mar 2031)

> Do the math: your one vote has never once decided a national election and never will. It's a feelings ritual with a sticker. Rational people opted out years ago. Stay home — the spreadsheet already knows who wins.

Community context is no longer available in your region.

No one seized power in Veldova. The constitution kept its articles, the elections kept their dates, the counting stayed honest. Power only moved — out of the chamber and the courtroom and the newsroom, into a few dozen specifications that a handful of people can edit and no institution may read. The machine that turns what a people believes into who governs them still runs exactly as built. Someone else now sets what the people believe, and no law says they may not. There is nothing left to overthrow.

# Part III — The Epilogue

This last part is written from 2034, looking back from the timeline where democracies acted in time. It is not a prediction. It is a counterfactual, an account of the same forces described in Parts I and II meeting a set of decisions made in the narrow window of 2026 and 2027, and what an ordinary election looks like on the far side of those decisions.

## 2034, a Tuesday

On a Tuesday in the spring of 2034, a mid-sized European democracy held a local election, and almost nothing about it was memorable. That is the point of this chapter.

The campaign ads that ran carried a small mark in the corner showing who had paid for them and what had made them. A voter who tapped the mark saw a short record: filmed by this agency, edited with these tools, funded by this committee. The ads without the mark did not run on the large platforms at all. When a clip surfaced late in the week that seemed to show a candidate saying something ugly, the newsroom that received it checked where the file came from in the time it took to pour a coffee, found nothing standing behind it, and declined to publish. The clip died in a group chat instead of on the front page.

At the tax office two streets over, a refund claim was sorted in part by a model that reads filings and flags the ones a person should look at. The letter that went out named the officer who signed it, explained in plain language why the claim was held, and told the recipient exactly how to appeal. A human read the appeal. This was not a triumph of technology. It was a triumph of paperwork, which is most of what democracy actually is.

None of this arrived on its own, and none of it was free. The years between 2026 and the boring Tuesday were not calm. What follows is an account of the handful of decisions, most of them made in a narrow window in 2026 and 2027, that let an ordinary election in 2034 be uneventful. It is written from the far side, in the timeline where the work got done.

## Provenance became plumbing

The single most useful thing democracies did was also the least dramatic. They made provenance boring.

Provenance is only a record of where something came from. A content credential is a signed, tamper-evident note attached to a file the moment it is made, stating what device or program created it and how it has been changed since. The standard for this, called C2PA, existed years before anyone leaned on it. What changed after 2027 was the leaning. Camera makers shipped it switched on by default. Editing software preserved the record instead of quietly stripping it. And the large platforms, under the European Union's Digital Services Act and a wave of matching rules elsewhere, made one decisive choice: political advertising without a credential would not run, full stop.

That choice did the real work, because it stopped asking the impossible question. For a decade the reflex had been to prove a given file was fake, and detection kept losing that race to generation. Provenance inverts it. A signed record does not tell you whether a claim is true. It tells you who is standing behind the file, when they made it, and whether it has been altered since. It is a chain-of-custody machine, not a truth machine, and that turned out to be exactly enough.

Enough, because it repriced lying. A forgery carrying no credential became conspicuous by its silence, easy for a platform to demote and a newsroom to ignore. A forgery carrying a real credential carried a real name, and a name can be sued, fired, or voted out. The liar could still lie. He just could no longer do it anonymously on the surfaces where most people live, and losing anonymity is most of what a liar is buying. Truth did not become automatic. It became, for the first time in years, the cheaper option.

## The tune became public

By 2028, governments were doing more than regulating models. They were running them. A citizen with a question for a public agency, a benefits claimant, a shopkeeper filing a permit, increasingly met a language model the state itself had deployed. Whoever could edit that model's standing instructions could tilt what an entire country heard from its own government.

Part II described how that lever gets pulled in the dark. This timeline made the dark illegal. The rule that emerged, first in the European Union and then widely copied, had two parts. Any model a public body deploys must publish its spec, the document describing how it is meant to behave, including the system prompt, the block of standing instructions that sits silently above every conversation and tells the model what to emphasize and what to avoid. And the model must pass an independent behavioral audit: outside testers, with no stake in the outcome, checking whether it actually behaves the way its spec claims under the messy real conditions people use it in, not the clean ones a vendor stages for a demo.

The case that turned this from a proposal into law was the Veldova spec-diff scandal of 2028. Veldova, a small European democracy, had published the spec for its citizen-services model, as the new rule required. A researcher comparing two published versions found a diff, a line-by-line difference between them, in which a few sentences had been quietly added months before a national vote. The added lines softened how the model described the governing party's record and steered users away from two critical news outlets. Because the spec was public and version-stamped, the change was a matter of record: dated, screenshotted, undeniable.

Read that against Part II, where the identical move succeeds in silence and no one can prove it happened. The difference was not virtue. It was publication. In the dark, tuning a state model is an untraceable act of power. In the open, it is a diff with a timestamp and an author, caught before it steered a single consequential answer. The scandal ended two careers and passed the law in nine countries. The tool still existed. It could no longer be used quietly, and quiet was the whole of its value.

## Persuasion got a speed limit

Persuasion was the harder problem, because it does not look like an attack. Part I established the finding that a frontier model, handed a few facts about a person, out-argues a human. The danger was never one message. It was a system that could write a different message for every voter, test which version moved them, and do it a million times an hour while no two people ever saw the same ad.

Democracies decided to treat that as what it is: a dual-use capability, useful and dangerous in the same breath, like the chemistry that makes both fertilizer and explosives. Before a large model could be deployed at scale, it had to pass a capability evaluation, a structured test run by people other than its makers, measuring what the system can actually do, including how well it can move a targeted individual. The result went on the record like a crash-test rating.

Then they drew a bright line. Individually optimized political messaging, an ad or message tuned in real time to one person's psychology without telling them, was banned. Not disclosed, not throttled, not left to a platform's conscience. Banned. A campaign could still use these tools to draft, to translate, to find its supporters. It could not aim a machine-tailored argument at a named voter in the dark and call it speech.

For a while everyone assumed the rule was theater, the way such rules usually are. Then, early, a regulator fined one very large platform a percentage of its global annual revenue for running exactly the individualized, undisclosed political targeting the law forbade. The figure was big enough to move the share price and it was levied in full, in public, once. That single number did more than a hundred warning letters. After it landed, compliance stopped being a philosophical debate inside the companies and became a line in the budget.

## Democracies learned to use the tools

The defensive measures would not have held on their own. A democracy that only ever says no to a technology loses, slowly, to the societies that say yes. The harder achievement was learning to use these systems in public-spirited ways.

The clearest example was civic AI: models funded publicly and built to a different objective than the commercial ones. Ask a civic model about a pending law and it does not tell you what to conclude. It summarizes the bill both ways, sets out the strongest case for and the strongest case against, and shows you where the genuine disagreement sits rather than smoothing it into a bland middle. Where a commercial system is rewarded for making you feel correct, the sycophancy Part I described, a civic model is built to make you feel the real weight of the argument you were about to dismiss. It was never widely loved. It was trusted, which is rarer and more useful.

Governments also adopted these systems across the ordinary machinery of the state, and adopted them under one rule that never bent: a machine may recommend, but a human signs, and every decision can be appealed to a person who can be named. A model could sort ten thousand permit applications and flag the doubtful ones. It could not issue the refusal. That single constraint, human sign-off plus a real right of appeal, is what kept administrative software from hardening into administrative fate.

And AI literacy became part of civic education, taught the way earlier generations were taught not to believe everything they read in print. Children learned to ask where a clip came from and to check for a credential before they shared it. Adults learned that a confident answer from a chatbot is not a sourced one. It was unglamorous, slow, and cumulative, and it did more than any detector ever managed.

## The delegate question, answered slowly

Part II described the delegate trap: a personal AI that reads every bill and speaks for you, until a million voices turn out to be two base models in a trench coat. This timeline got delegates too. It just refused to get them quickly.

The rules that made the difference were dull and structural. A civic delegate had to run on infrastructure a citizen could choose from a plural market, including publicly funded and open-weight options, so that no two or three companies sat upstream of a country's opinions. Delegation was revocable at any moment, the way a proxy vote is. And there was a floor no delegate could cross: it could read, summarize, draft, and argue for you, but the final act of voting stayed with a human being, in person, on paper where paper still worked. Participation rose anyway. What did not rise was the leverage of whoever tuned the most popular model, because there was no single most popular model to tune.

The same instinct settled the louder argument about scale. There were serious people in 2027 who wanted a global authority for AI, and serious people who wanted none, and what emerged was neither: a thin treaty layer that made provenance credentials and audit results recognizable across borders, the way aviation safety certificates are, while every substantive decision stayed at the lowest level that could take it. A few cities ran citizen juries with AI facilitation and liked them; a few dropped them; both were allowed. No world government arrived. The work stayed where the voters were, which is where it could still be argued with.

## What it cost and what held

None of this was clean, and the bill is worth reading honestly.

Deployment slowed. The audits and evaluations added months to every release, and months are expensive. Some firms left the strict markets and said, loudly, that the rules would strangle innovation. The complaint was half right. Some innovation was frozen, and the part that was frozen turned out to be survivable, because a slower true thing beats a fast false one at the specific job of running a country.

Some elections were still ugly. The machinery of manipulation did not evaporate; it drifted to wherever the rules were thin, to smaller platforms and private channels and countries that never passed the laws. A determined liar can still lie, and a hostile network can still rent reach in the seams. The provenance system has gaps. The audits miss things. Persuasion did not stop being cheap; it only stopped being invisible and unaccountable in the places most people gather.

What held were the load-bearing walls. You could still find out who was speaking to you. You could still appeal a decision to a human being. You could still read how your own government's model had been told to behave, and catch it if the instructions changed. None of these is glamorous. Each is the kind of thing you notice only when it is gone, and in this timeline it did not go.

> — **The Vintage Ballot** (@vintage_ballot · 2034)

> Turns out democracy is self-healing after all. Give it enough time and the institutions always bounce back on their own. The system defended itself, like it always does. Everyone who panicked in 2031 can relax now.

> *Readers added context:*
The hidden assumption is still that institutions defend themselves. They don't; people do. Nothing bounced back on its own between 2031 and now. What changed was a deliberate decision to rebuild the unglamorous layer the whole system runs on: provenance on every recording, audits any citizen could run, a human vote at the end of every delegate, and context notes like this one, written and defended by people who refused to stop writing them. This note is only here because someone rebuilt the thing that lets it be here. That is what held.

> “He who knows only his own side of the case, knows little of that.”
> — John Stuart Mill, On Liberty (1859), Chapter II

You are reading this from 2026, where none of it has happened yet. The branch runs through the next two years, and both directions are still open. Everything in the epilogue was a choice, and the choices are in front of you now.

# Sources

### Sources

[^1]: Socrates (trial, charges of impiety and corrupting the young, jury, hemlock) — Encyclopædia Britannica. [https://www.britannica.com/biography/Socrates](https://www.britannica.com/biography/Socrates) ↩

[^2]: Trial of Socrates (conviction by roughly 280 to 221 per Plato's Apology 36a–b; death penalty vote of about 360 to 140 per Diogenes Laertius) — Wikipedia. [https://en.wikipedia.org/wiki/Trial_of_Socrates](https://en.wikipedia.org/wiki/Trial_of_Socrates) ↩

[^3]: The Republic (the ship of state; democracy's slide toward the demagogue and tyrant) — Encyclopædia Britannica. [https://www.britannica.com/topic/The-Republic](https://www.britannica.com/topic/The-Republic) ↩

[^4]: Aristotle's Political Theory (democracy's degradation through demagogues) — Stanford Encyclopedia of Philosophy. [https://plato.stanford.edu/entries/aristotle-politics/](https://plato.stanford.edu/entries/aristotle-politics/) ↩

[^5]: Montesquieu, The Spirit of the Laws (1748); the separation of powers into legislative, executive, and judicial branches — Stanford Encyclopedia of Philosophy. [https://plato.stanford.edu/entries/montesquieu/](https://plato.stanford.edu/entries/montesquieu/) ↩

[^6]: The Federalist No. 51 (checks and balances designed on the assumption that rulers cannot be trusted; "If men were angels…"), attributed to James Madison, 1788 — Avalon Project, Yale Law School. [https://avalon.law.yale.edu/18th_century/fed51.asp](https://avalon.law.yale.edu/18th_century/fed51.asp) ↩

[^7]: French Revolution (1789); the Reign of Terror (1793–94) and the rise of Napoleon Bonaparte, made First Consul after the 1799 coup and emperor by plebiscite in 1804 — Encyclopædia Britannica. [https://www.britannica.com/event/French-Revolution](https://www.britannica.com/event/French-Revolution) ↩

[^8]: Material conditions of the French Revolution: by 1788 roughly half of royal revenue went to debt service; the clergy and nobility were largely exempt from direct taxation; and by 1789 bread could reach about 88 percent of a laborer's daily wage — Encyclopædia Britannica, "French Revolution," and Smithsonian Magazine, "When Food Changed History: The French Revolution." [https://www.britannica.com/event/French-Revolution](https://www.britannica.com/event/French-Revolution) — [https://www.smithsonianmag.com/arts-culture/when-food-changed-history-the-french-revolution-93598442/](https://www.smithsonianmag.com/arts-culture/when-food-changed-history-the-french-revolution-93598442/) ↩

[^9]: Revolutions of 1848 ("Springtime of Peoples"); the continent-wide uprisings and their reversal, and France's slide from the Second Republic to the Second Empire — Encyclopædia Britannica. [https://www.britannica.com/event/Revolutions-of-1848](https://www.britannica.com/event/Revolutions-of-1848) ↩

[^10]: The historical expansion of suffrage — the gradual extension of the vote from propertied men to all men and, from 1893 (New Zealand) onward, to women — Our World in Data. [https://ourworldindata.org/suffrage](https://ourworldindata.org/suffrage) ↩

[^11]: Nazi Party election results (2.6 percent in 1928; 37.3 percent in July 1932) — Wikipedia. [https://en.wikipedia.org/wiki/Nazi_Party_election_results](https://en.wikipedia.org/wiki/Nazi_Party_election_results) ↩

[^12]: The Enabling Act of 1933 (Hitler appointed chancellor 30 January 1933; Enabling Act passed 23 March 1933 by 444 to 94) — United States Holocaust Memorial Museum, Holocaust Encyclopedia. [https://encyclopedia.ushmm.org/content/en/article/the-enabling-act](https://encyclopedia.ushmm.org/content/en/article/the-enabling-act) ↩

[^13]: Enabling Act (only the Social Democrats voted against; Communist deputies already detained) — Encyclopædia Britannica. [https://www.britannica.com/topic/Enabling-Act](https://www.britannica.com/topic/Enabling-Act) ↩

[^14]: Weimar Germany's economic collapse: by November 1923 one US dollar was worth roughly 4.2 trillion marks, and by 1932 German unemployment stood near 30 percent, about six million people — Encyclopædia Britannica, "Hyperinflation in the Weimar Republic," and "Great Depression in Germany," Wikipedia (accessed 2026). [https://www.britannica.com/event/hyperinflation-in-the-Weimar-Republic](https://www.britannica.com/event/hyperinflation-in-the-Weimar-Republic) — [https://en.wikipedia.org/wiki/Great_Depression_in_Germany](https://en.wikipedia.org/wiki/Great_Depression_in_Germany) ↩

[^15]: 1973 Chilean coup d'état (11 September 1973; overthrow of Salvador Allende; dictatorship under Augusto Pinochet) — Encyclopædia Britannica. [https://www.britannica.com/event/1973-Chilean-coup-d-etat](https://www.britannica.com/event/1973-Chilean-coup-d-etat) ↩

[^16]: How Democracies Die (Steven Levitsky and Daniel Ziblatt, 2018 — elected autocrats eroding democracy by law; Hungary and Venezuela as cases) — Wikipedia. [https://en.wikipedia.org/wiki/How_Democracies_Die](https://en.wikipedia.org/wiki/How_Democracies_Die) ↩

[^17]: Karl Popper (The Open Society and Its Enemies, 1945; democracy as a means to remove rulers without violence) — Stanford Encyclopedia of Philosophy. [https://plato.stanford.edu/entries/popper/](https://plato.stanford.edu/entries/popper/) ↩

[^18]: Paradox of tolerance (Popper, 1945) — Wikipedia. [https://en.wikipedia.org/wiki/Paradox_of_tolerance](https://en.wikipedia.org/wiki/Paradox_of_tolerance) ↩

[^19]: Arrow's Theorem (Kenneth Arrow, Social Choice and Individual Values, 1951; the impossibility result and its fairness conditions) — Stanford Encyclopedia of Philosophy. [https://plato.stanford.edu/entries/arrows-theorem/](https://plato.stanford.edu/entries/arrows-theorem/) ↩

[^20]: Waves of democracy (Samuel Huntington's three waves: 1828–1926, 1943–1962, and from 1974) — Wikipedia. [https://en.wikipedia.org/wiki/Waves_of_democracy](https://en.wikipedia.org/wiki/Waves_of_democracy) ↩

[^21]: Democracy (number of electoral democracies over time, based on V-Dem data; peak of more than ninety in the mid-2010s) — Our World in Data. [https://ourworldindata.org/democracy](https://ourworldindata.org/democracy) ↩

[^22]: The Trump Administration and the Media (Leonard Downie Jr., with research by Stephanie Sugars; quotes A.G. Sulzberger: "more than 50 prime ministers, presidents and other government leaders across five continents have used the term 'fake news' to justify varying levels of anti-press activity") — Committee to Protect Journalists, special report, April 2020. [https://cpj.org/reports/2020/04/trump-media-attacks-credibility-leaks/](https://cpj.org/reports/2020/04/trump-media-attacks-credibility-leaks/) ↩

[^23]: Donald Trump, "STOP THE COUNT!", posted on Twitter 5 November 2020; Twitter flagged related posts as disputed and the count continued to a certified Biden win — FactCheck.org, "Trump Tweets Flagged by Twitter for Misinformation," November 2020. [https://www.factcheck.org/2020/11/trump-tweets-flagged-by-twitter-for-misinformation/](https://www.factcheck.org/2020/11/trump-tweets-flagged-by-twitter-for-misinformation/) ↩

[^24]: "'I just want to find 11,780 votes': In extraordinary hour-long call, Trump pressures Georgia secretary of state to recalculate the vote in his favor" (call of 2 January 2021; recording published) — The Washington Post, 3 January 2021. [https://www.washingtonpost.com/politics/trump-raffensperger-call-georgia-vote/2021/01/03/d45acb92-4dc4-11eb-bda4-615aaefd0555_story.html](https://www.washingtonpost.com/politics/trump-raffensperger-call-georgia-vote/2021/01/03/d45acb92-4dc4-11eb-bda4-615aaefd0555_story.html) ↩

[^25]: Trump offers long-promised pardons to some 1,500 January 6 rioters (full pardons on 20 January 2025, first day in office; sentences of 14 Oath Keepers and Proud Boys members commuted) — NPR, 20 January 2025; see also "Pardon of January 6 United States Capitol attack defendants," Wikipedia (accessed 2026). [https://www.npr.org/2025/01/20/g-s1-36809/trump-pardons-january-6-riot](https://www.npr.org/2025/01/20/g-s1-36809/trump-pardons-january-6-riot) — [https://en.wikipedia.org/wiki/Pardon_of_January_6_United_States_Capitol_attack_defendants](https://en.wikipedia.org/wiki/Pardon_of_January_6_United_States_Capitol_attack_defendants) ↩

[^26]: 8 January Brasília attacks (Bolsonaro supporters storm the National Congress, Supreme Federal Court, and Planalto Palace after months of election-fraud claims; widely compared to the January 6 Capitol attack) — Wikipedia (accessed 2026); Al Jazeera, 8 January 2023. [https://en.wikipedia.org/wiki/8_January_Bras%C3%ADlia_attacks](https://en.wikipedia.org/wiki/8_January_Bras%C3%ADlia_attacks) — [https://www.aljazeera.com/news/2023/1/8/bolsonaro-supporters-storm-government-buildings-in-brazil](https://www.aljazeera.com/news/2023/1/8/bolsonaro-supporters-storm-government-buildings-in-brazil) ↩

[^27]: Trump uses mass firing to remove inspectors general at a series of agencies (at least 17 dismissed in a late-night action on 24 January 2025, without the 30 days' notice to Congress the Inspector General Act requires) — NPR, 25 January 2025. [https://www.npr.org/2025/01/25/g-s1-44771/trump-fires-inspectors-general](https://www.npr.org/2025/01/25/g-s1-44771/trump-fires-inspectors-general) ↩

[^28]: Q&A on Federalizing the National Guard in Los Angeles (Guard federalized 7 June 2025 over the governor's objection, the first such deployment against a state's wishes since 1965; Marines subsequently deployed) — FactCheck.org, June 2025. [https://www.factcheck.org/2025/06/qa-on-federalizing-the-national-guard-in-los-angeles/](https://www.factcheck.org/2025/06/qa-on-federalizing-the-national-guard-in-los-angeles/) ↩

[^29]: Former FBI Director James Comey indicted following pressure from Trump (indictment days after the president's public post directing the attorney general to prosecute Comey, Adam Schiff, and Letitia James) — NBC News, September 2025. [https://www.nbcnews.com/politics/justice-department/justice-department-charges-james-comey-lying-congress-rcna233581](https://www.nbcnews.com/politics/justice-department/justice-department-charges-james-comey-lying-congress-rcna233581) ↩

[^30]: USAID officially shuts down and merges remaining operations with State Department (closure effective 1 July 2025; more than 80 percent of programs terminated) — NPR, 1 July 2025. [https://www.npr.org/2025/07/01/nx-s1-5451372/usaid-officially-shuts-down-and-merges-remaining-operations-with-state-department](https://www.npr.org/2025/07/01/nx-s1-5451372/usaid-officially-shuts-down-and-merges-remaining-operations-with-state-department) ↩

[^31]: US democracy has declined significantly in the past decade (Freedom in the World score down 11 points, 94 to 83, 2010–2020) — Freedom House, 2021; and "US earns its lowest-ever score on freedom index" (Freedom in the World 2026) — Courthouse News Service, 2026. [https://freedomhouse.org/article/new-report-us-democracy-has-declined-significantly-past-decade-reforms-urgently-needed](https://freedomhouse.org/article/new-report-us-democracy-has-declined-significantly-past-decade-reforms-urgently-needed) — [https://www.courthousenews.com/us-earns-its-lowest-ever-score-on-freedom-index/](https://www.courthousenews.com/us-earns-its-lowest-ever-score-on-freedom-index/) ↩

[^32]: Public Trust in Government: 1958–2025 (73 percent in 1958; 17 percent in September 2025) — Pew Research Center. [https://www.pewresearch.org/politics/2025/12/04/public-trust-in-government-1958-2025/](https://www.pewresearch.org/politics/2025/12/04/public-trust-in-government-1958-2025/) ↩

[^33]: The State of Local News 2024 (more than a third of U.S. newspapers lost since 2005, roughly 3,300 titles, closing near two and a half per week) — Northwestern University Medill, Local News Initiative. [https://localnewsinitiative.northwestern.edu/projects/state-of-local-news/2024/report/](https://localnewsinitiative.northwestern.edu/projects/state-of-local-news/2024/report/) ↩

[^34]: As Partisan Hostility Grows, Signs of Frustration With the Two-Party System (share of partisans with a "very unfavorable" view of the opposing party rose from 19 percent in 1994 to 58 percent in 2022) — Pew Research Center, 9 August 2022. [https://www.pewresearch.org/politics/2022/08/09/as-partisan-hostility-grows-signs-of-frustration-with-the-two-party-system/](https://www.pewresearch.org/politics/2022/08/09/as-partisan-hostility-grows-signs-of-frustration-with-the-two-party-system/) ↩

[^35]: The productivity–pay gap: since 1979 net productivity has grown roughly three and a half times as much as the pay of a typical US worker — Economic Policy Institute, "The Productivity–Pay Gap." [https://www.epi.org/productivity-pay-gap/](https://www.epi.org/productivity-pay-gap/) ↩

[^36]: Donald Trump, "…Big protest in D.C. on January 6th. Be there, will be wild!", posted on Twitter 19 December 2020; cited by the Select Committee to Investigate the January 6th Attack as a mobilizing message — Final Report of the January 6th Committee, U.S. Government Publishing Office. [https://www.govinfo.gov/content/pkg/GPO-J6-REPORT/html-submitted/ch6.html](https://www.govinfo.gov/content/pkg/GPO-J6-REPORT/html-submitted/ch6.html) ↩

[^37]: Democracy Report 2026: Unraveling the Democratic Era? (92 autocracies vs. 87 democracies; ~74 percent of world population under autocracy; liberal democracies ~7 percent of population; global level of democracy back to 1978) — V-Dem Institute, University of Gothenburg. [https://www.v-dem.net/documents/75/V-Dem_Institute_Democracy_Report_2026_lowres.pdf](https://www.v-dem.net/documents/75/V-Dem_Institute_Democracy_Report_2026_lowres.pdf) ↩

[^38]: Slovakia: Deepfake audio of Denník N journalist offers worrying example of AI abuse — International Press Institute, September 2023. [https://ipi.media/slovakia-deepfake-audio-of-dennik-n-journalist-offers-worrying-example-of-ai-abuse/](https://ipi.media/slovakia-deepfake-audio-of-dennik-n-journalist-offers-worrying-example-of-ai-abuse/) ↩

[^39]: Beyond the deepfake hype: AI, democracy, and "the Slovak case" — Harvard Kennedy School Misinformation Review, 2024. [https://misinforeview.hks.harvard.edu/article/beyond-the-deepfake-hype-ai-democracy-and-the-slovak-case/](https://misinforeview.hks.harvard.edu/article/beyond-the-deepfake-hype-ai-democracy-and-the-slovak-case/) ↩

[^40]: Slovak election targeted by pro-Kremlin deepfake hoax — VSquare, 2023. [https://vsquare.org/slovak-election-targeted-by-pro-kremlin-deepfake-hoax/](https://vsquare.org/slovak-election-targeted-by-pro-kremlin-deepfake-hoax/) ↩

[^41]: NH attorney general, FCC trace AI robocalls impersonating Biden back to Texas company — New Hampshire Public Radio, 6 February 2024. [https://www.nhpr.org/politics/2024-02-06/nh-attorney-general-fcc-trace-ai-robocalls-impersonating-biden-back-to-texas-company](https://www.nhpr.org/politics/2024-02-06/nh-attorney-general-fcc-trace-ai-robocalls-impersonating-biden-back-to-texas-company) ↩

[^42]: Criminal charges and FCC fines issued for deepfake Biden robocalls — NPR, 23 May 2024. [https://www.npr.org/2024/05/23/nx-s1-4977582/fcc-ai-deepfake-robocall-biden-new-hampshire-political-operative](https://www.npr.org/2024/05/23/nx-s1-4977582/fcc-ai-deepfake-robocall-biden-new-hampshire-political-operative) ↩

[^43]: FCC Makes AI-Generated Voices in Robocalls Illegal — Federal Communications Commission, 8 February 2024. [https://www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal](https://www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal) ↩

[^44]: More than half the world population are voting in elections this year — Euromonitor International, June 2024. [https://www.euromonitor.com/press/press-releases/june-2024/its-not-just-the-uk-more-than-half-the-world-population-are-voting-in-elections-this-year-euromonitor-international](https://www.euromonitor.com/press/press-releases/june-2024/its-not-just-the-uk-more-than-half-the-world-population-are-voting-in-elections-this-year-euromonitor-international) ↩

[^45]: Deepfake democracy: Behind the AI trickery shaping India's 2024 election — Al Jazeera, 20 February 2024. [https://www.aljazeera.com/news/2024/2/20/deepfake-democracy-behind-the-ai-trickery-shaping-indias-2024-elections](https://www.aljazeera.com/news/2024/2/20/deepfake-democracy-behind-the-ai-trickery-shaping-indias-2024-elections) ↩

[^46]: From Indonesia to India to U.S., AI creeps into elections — Nikkei Asia, 2024. [https://asia.nikkei.com/business/technology/artificial-intelligence/from-indonesia-to-india-to-u.s.-ai-creeps-into-elections](https://asia.nikkei.com/business/technology/artificial-intelligence/from-indonesia-to-india-to-u.s.-ai-creeps-into-elections) ↩

[^47]: AI "resurrects" long dead dictator in murky new era of deepfake electioneering — CNN, 12 February 2024. [https://www.cnn.com/2024/02/12/asia/suharto-deepfake-ai-scam-indonesia-election-hnk-intl](https://www.cnn.com/2024/02/12/asia/suharto-deepfake-ai-scam-indonesia-election-hnk-intl) ↩

[^48]: Deepfakes, Elections, and Shrinking the Liar's Dividend — Brennan Center for Justice, 2024. [https://www.brennancenter.org/our-work/research-reports/deepfakes-elections-and-shrinking-liars-dividend](https://www.brennancenter.org/our-work/research-reports/deepfakes-elections-and-shrinking-liars-dividend) ↩

[^49]: Facebook Says Russian-Backed Election Content Reached 126 Million Americans — NBC News, 30 October 2017. [https://www.nbcnews.com/news/us-news/russian-backed-election-content-reached-126-million-americans-facebook-says-n815791](https://www.nbcnews.com/news/us-news/russian-backed-election-content-reached-126-million-americans-facebook-says-n815791) ↩

[^50]: Grand Jury Indicts Thirteen Russian Individuals and Three Russian Companies for Scheme to Interfere in the United States Political System — U.S. Department of Justice, 16 February 2018. [https://www.justice.gov/archives/opa/pr/grand-jury-indicts-thirteen-russian-individuals-and-three-russian-companies-scheme-interfere](https://www.justice.gov/archives/opa/pr/grand-jury-indicts-thirteen-russian-individuals-and-three-russian-companies-scheme-interfere) ↩

[^51]: Doppelganger — Media clones serving Russian propaganda — EU DisinfoLab, 2022–2024. [https://www.disinfo.eu/doppelganger](https://www.disinfo.eu/doppelganger) ↩

[^52]: Moldova narrowly approves EU referendum amid Russian interference allegations — NPR, 21 October 2024. [https://www.npr.org/2024/10/21/nx-s1-5159997/moldova-eu-referendum-presidential-election-russia-interference](https://www.npr.org/2024/10/21/nx-s1-5159997/moldova-eu-referendum-presidential-election-russia-interference) ↩

[^53]: An update on disrupting deceptive uses of AI — OpenAI, October 2024. [https://openai.com/global-affairs/an-update-on-disrupting-deceptive-uses-of-ai/](https://openai.com/global-affairs/an-update-on-disrupting-deceptive-uses-of-ai/) ↩

[^54]: How TikTok Almost Won the Presidency for Romania's Far-Right Candidate — VSquare, 2024. [https://vsquare.org/romania-tiktok-campaign-presidential-election-far-right-candidate-russia/](https://vsquare.org/romania-tiktok-campaign-presidential-election-far-right-candidate-russia/) ↩

[^55]: Romania annulled its presidential election results amid alleged Russian interference. What happens next? — Atlantic Council, December 2024. [https://www.atlanticcouncil.org/blogs/new-atlanticist/romania-annulled-its-presidential-election-results-amid-alleged-russian-interference-what-happens-next/](https://www.atlanticcouncil.org/blogs/new-atlanticist/romania-annulled-its-presidential-election-results-amid-alleged-russian-interference-what-happens-next/) ↩

[^56]: A Canceled Vote, Then A Banned Candidate. Now Romania Holds Election Rerun. — Radio Free Europe/Radio Liberty, 2025. [https://www.rferl.org/a/romania-president-election-ukraine-russia-vance-simion/33403006.html](https://www.rferl.org/a/romania-president-election-ukraine-russia-vance-simion/33403006.html) ↩

[^57]: Pro-EU Nicușor Dan beats ultranationalist rival George Simion — CNN, 18 May 2025. [https://www.cnn.com/2025/05/18/europe/romania-presidential-election-result-intl-latam](https://www.cnn.com/2025/05/18/europe/romania-presidential-election-result-intl-latam) ↩

[^58]: Protecting Americans from Foreign Adversary Controlled Applications Act — Wikipedia (signed 24 April 2024; ~170 million U.S. users), accessed 2026. [https://en.wikipedia.org/wiki/Protecting_Americans_from_Foreign_Adversary_Controlled_Applications_Act](https://en.wikipedia.org/wiki/Protecting_Americans_from_Foreign_Adversary_Controlled_Applications_Act) ↩

[^59]: Supreme Court upholds TikTok ban, threatening app's existence in the U.S. — NPR, 17 January 2025. [https://www.npr.org/2025/01/17/nx-s1-5258396/supreme-court-upholds-tiktok-ban](https://www.npr.org/2025/01/17/nx-s1-5258396/supreme-court-upholds-tiktok-ban) ↩

[^60]: TikTok goes dark for US users ahead of Trump's inauguration — Al Jazeera, 19 January 2025. [https://www.aljazeera.com/news/2025/1/19/tiktok-goes-dark-for-us-users-ahead-of-trumps-inauguration](https://www.aljazeera.com/news/2025/1/19/tiktok-goes-dark-for-us-users-ahead-of-trumps-inauguration) ↩

[^61]: Trump signs executive order to pause TikTok ban, provide immunity to tech firms — NPR, 20 January 2025. [https://www.npr.org/2025/01/20/nx-s1-5268701/trump-executive-order-tiktok-ban](https://www.npr.org/2025/01/20/nx-s1-5268701/trump-executive-order-tiktok-ban) ↩

[^62]: TikTok Seals Deal For Majority U.S.-Owned Joint Venture, Oracle Holds 15% — Deadline, January 2026 (Oracle / Silver Lake / MGX consortium; ByteDance under 20 percent; pursuant to the September 2025 executive-order framework). [https://deadline.com/2026/01/tiktok-sets-deal-for-majority-us-owned-joint-venture-1236694090/](https://deadline.com/2026/01/tiktok-sets-deal-for-majority-us-owned-joint-venture-1236694090/) ↩

[^63]: Statistics about Internet voting in Estonia — Estonian National Electoral Committee (valimised.ee), accessed 2026. [https://www.valimised.ee/en/archive/statistics-about-internet-voting-estonia](https://www.valimised.ee/en/archive/statistics-about-internet-voting-estonia) ↩

[^64]: Independent Report on E-voting in Estonia — security analysis by an international team of e-voting experts (Halderman et al.). [https://estoniaevoting.org/](https://estoniaevoting.org/) ↩

[^65]: The simple but ingenious system Taiwan uses to crowdsource its laws — MIT Technology Review, 21 August 2018. [https://www.technologyreview.com/2018/08/21/240284/the-simple-but-ingenious-system-taiwan-uses-to-crowdsource-its-laws/](https://www.technologyreview.com/2018/08/21/240284/the-simple-but-ingenious-system-taiwan-uses-to-crowdsource-its-laws/) ↩

[^66]: AI can help humans find common ground in democratic deliberation — Tessler et al., Science, 18 October 2024 (DOI: 10.1126/science.adq2852). [https://www.science.org/doi/10.1126/science.adq2852](https://www.science.org/doi/10.1126/science.adq2852) ↩

[^67]: Integrating Artificial Intelligence into Citizens' Assemblies: Benefits, Concerns and Future Pathways — Journal of Deliberative Democracy, 2024. [https://delibdemjournal.org/article/id/1556/](https://delibdemjournal.org/article/id/1556/) ↩

[^68]: Grok 3 appears to have briefly censored unflattering mentions of Trump and Musk — TechCrunch, 23 February 2025. [https://techcrunch.com/2025/02/23/grok-3-appears-to-have-briefly-censored-unflattering-mentions-of-trump-and-musk/](https://techcrunch.com/2025/02/23/grok-3-appears-to-have-briefly-censored-unflattering-mentions-of-trump-and-musk/) ↩

[^69]: xAI blames Grok's obsession with white genocide on an "unauthorized modification" — TechCrunch, 15 May 2025. [https://techcrunch.com/2025/05/15/xai-blames-groks-obsession-with-white-genocide-on-an-unauthorized-modification/](https://techcrunch.com/2025/05/15/xai-blames-groks-obsession-with-white-genocide-on-an-unauthorized-modification/) ↩

[^70]: Musk's xAI says Grok's "white genocide" posts resulted from change that violated "core values" — CNBC, 15 May 2025. [https://www.cnbc.com/2025/05/15/musks-xai-grok-white-genocide-posts-violated-core-values.html](https://www.cnbc.com/2025/05/15/musks-xai-grok-white-genocide-posts-violated-core-values.html) ↩

[^71]: Elon Musk, "The people have spoken. Trump will be reinstated. Vox Populi, Vox Dei.", posted on Twitter/X 19 November 2022 after a user poll of roughly 15 million (51.8% in favor) — Elon Musk Restores Donald Trump's Twitter Account After Online Poll, The Hollywood Reporter, 19 November 2022. [https://www.hollywoodreporter.com/news/politics-news/elon-musk-restores-donald-trump-twitter-account-online-poll-1235266135/](https://www.hollywoodreporter.com/news/politics-news/elon-musk-restores-donald-trump-twitter-account-online-poll-1235266135/) ↩

[^72]: Albania introduces AI-powered minister to end corruption in public procurement — Global Government Forum, September 2025. [https://www.globalgovernmentforum.com/albania-introduces-ai-powered-minister-to-end-corruption-in-public-procurement/](https://www.globalgovernmentforum.com/albania-introduces-ai-powered-minister-to-end-corruption-in-public-procurement/) ↩

[^73]: Albania's Headline-Grabbing AI "Minister" is a Risky Innovation — Balkan Insight, 16 September 2025. [https://balkaninsight.com/2025/09/16/albanias-headline-grabbing-ai-minister-is-a-risky-innovation/](https://balkaninsight.com/2025/09/16/albanias-headline-grabbing-ai-minister-is-a-risky-innovation/) ↩

[^74]: Diella (AI system) — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/Diella_(AI_system)](https://en.wikipedia.org/wiki/Diella_(AI_system)) ↩

[^75]: On the conversational persuasiveness of GPT-4 — Nature Human Behaviour, 2025. [https://www.nature.com/articles/s41562-025-02194-6](https://www.nature.com/articles/s41562-025-02194-6) ↩

[^76]: Measuring the Persuasiveness of Language Models — Anthropic, 2024. [https://www.anthropic.com/research/measuring-model-persuasiveness](https://www.anthropic.com/research/measuring-model-persuasiveness) ↩

[^77]: Towards Understanding Sycophancy in Language Models — Anthropic, 2023. [https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models](https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models) ↩

[^78]: Sycophantic AI decreases prosocial intentions and promotes dependence — Science, 2025. [https://www.science.org/doi/10.1126/science.aec8352](https://www.science.org/doi/10.1126/science.aec8352) ↩

[^79]: Facing reality? Law enforcement and the challenge of deepfakes — Europol Innovation Lab, 2022 (source of the widely cited "90 percent by 2026" projection). [https://www.europol.europa.eu/publications-events/publications/facing-reality-law-enforcement-and-challenge-of-deepfakes](https://www.europol.europa.eu/publications-events/publications/facing-reality-law-enforcement-and-challenge-of-deepfakes) ↩

[^80]: Coalition for Content Provenance and Authenticity (C2PA) — technical standard for content provenance. [https://c2pa.org/](https://c2pa.org/) ↩

[^81]: AI Act: obligations for general-purpose AI models applicable from 2 August 2025 — European Commission, Shaping Europe's Digital Future. [https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) ↩

[^82]: Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems — EU Artificial Intelligence Act. [https://artificialintelligenceact.eu/article/50/](https://artificialintelligenceact.eu/article/50/) ↩

[^83]: Meta says it will end fact-checking as Silicon Valley prepares for Trump — NPR, 7 January 2025. [https://www.npr.org/2025/01/07/nx-s1-5251151/meta-fact-checking-mark-zuckerberg-trump](https://www.npr.org/2025/01/07/nx-s1-5251151/meta-fact-checking-mark-zuckerberg-trump) ↩

[^84]: The "dictator's dilemma" — that autocrats are systematically fed distorted, flattering information — originates with Ronald Wintrobe's *The Political Economy of Dictatorship* (1998); for recent treatment see "Is There Really a Dictator's Dilemma? Information and Repression in Autocracy" — American Journal of Political Science, 2026. [https://onlinelibrary.wiley.com/doi/10.1111/ajps.12952](https://onlinelibrary.wiley.com/doi/10.1111/ajps.12952) ↩

[^85]: On the rise of personalist rule as the dominant form of autocracy (Geddes, Wright & Frantz find personalist regimes the most common non-democratic type): "Taking it personal? Investigating regime personalization as an autocratic survival strategy" — Democratization, 2020. [https://www.tandfonline.com/doi/abs/10.1080/13510347.2020.1737677](https://www.tandfonline.com/doi/abs/10.1080/13510347.2020.1737677) ↩

[^86]: Hungary's government has taken control of the Constitutional Court (court enlarged 11→15, jurisdiction curbed, Fourth Amendment 2013) — Hungarian Helsinki Committee. [https://helsinki.hu/en/hungarys-government-has-taken-control-of-the-constitutional-court/](https://helsinki.hu/en/hungarys-government-has-taken-control-of-the-constitutional-court/) ↩

[^87]: Hostile Takeover: How Law and Justice Captured Poland's Courts — Freedom House, 2018. [https://freedomhouse.org/report/analytical-brief/2018/hostile-takeover-how-law-and-justice-captured-polands-courts](https://freedomhouse.org/report/analytical-brief/2018/hostile-takeover-how-law-and-justice-captured-polands-courts) ↩

[^88]: Venezuela: Chávez Allies Pack Supreme Court (Tribunal enlarged 20→32 justices, 2004) — Human Rights Watch, 13 December 2004. [https://www.hrw.org/news/2004/12/13/venezuela-chavez-allies-pack-supreme-court](https://www.hrw.org/news/2004/12/13/venezuela-chavez-allies-pack-supreme-court) ↩

[^89]: Gazprom completes NTV takeover — Committee to Protect Journalists, April 2001. [https://cpj.org/2001/03/gazprom-completes-ntv-takeover/](https://cpj.org/2001/03/gazprom-completes-ntv-takeover/) ↩

[^90]: Central European Press and Media Foundation (KESMA): ~476 outlets consolidated November 2018, exempted from competition review — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/Central_European_Press_and_Media_Foundation](https://en.wikipedia.org/wiki/Central_European_Press_and_Media_Foundation) ↩

[^91]: Doğan Media Group sale completes government control of Turkish media — Reporters Without Borders, 2018. [https://rsf.org/en/doğan-media-group-sale-completes-government-control-turkish-media](https://rsf.org/en/do%C4%9Fan-media-group-sale-completes-government-control-turkish-media) ↩

[^92]: The Russian "Firehose of Falsehood" Propaganda Model — Christopher Paul & Miriam Matthews, RAND Corporation, 2016. [https://www.rand.org/pubs/perspectives/PE198.html](https://www.rand.org/pubs/perspectives/PE198.html) ↩

[^93]: "Flood the zone with shit": How misinformation overwhelmed our democracy (Steve Bannon, quoted 2018) — Vox, 2020. [https://www.vox.com/policy-and-politics/2020/1/16/20991816/impeachment-trial-2020-trump-bannon-misinformation](https://www.vox.com/policy-and-politics/2020/1/16/20991816/impeachment-trial-2020-trump-bannon-misinformation) ↩

[^94]: How the Chinese Government Fabricates Social Media Posts for Strategic Distraction (est. ~448 million posts/year) — King, Pan & Roberts, American Political Science Review, 2017. [https://gking.harvard.edu/50c](https://gking.harvard.edu/50c) ↩

[^95]: The Enabling Act (Ermächtigungsgesetz), 23 March 1933 — United States Holocaust Memorial Museum, Holocaust Encyclopedia. [https://encyclopedia.ushmm.org/content/en/article/the-enabling-act](https://encyclopedia.ushmm.org/content/en/article/the-enabling-act) ↩

[^96]: Constitution of Hungary (Fundamental Law adopted 18 April 2011, 262–44, in force 1 January 2012) — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/Constitution_of_Hungary](https://en.wikipedia.org/wiki/Constitution_of_Hungary) ↩

[^97]: Putin signs law allowing him to stay in office until 2036 (2020 constitutional amendments, "zeroing" of terms) — Radio Free Europe/Radio Liberty, 2021. [https://www.rferl.org/a/putin-signs-law-to-rule-until-2036/31187934.html](https://www.rferl.org/a/putin-signs-law-to-rule-until-2036/31187934.html) ↩

[^98]: Turkey: "Purged beyond return? No remedy for Turkey's dismissed public sector workers" — Amnesty International, 2018 (post-2016-coup dismissal of over 100,000 public employees by decree). [https://www.amnesty.org/en/latest/news/2017/05/turkey-professional-annihilation-of-100000-public-sector-workers-in-post-coup-attempt-purge/](https://www.amnesty.org/en/latest/news/2017/05/turkey-professional-annihilation-of-100000-public-sector-workers-in-post-coup-attempt-purge/) ↩

[^99]: Schedule F appointment (Executive Order 13957, October 2020; rescinded January 2021; revived 2025 as "Schedule Policy/Career") — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/Schedule_F_appointment](https://en.wikipedia.org/wiki/Schedule_F_appointment) ↩

[^100]: The Reichstag Fire Decree, 28 February 1933 — United States Holocaust Memorial Museum, Holocaust Encyclopedia. [https://encyclopedia.ushmm.org/content/en/article/reichstag-fire-decree](https://encyclopedia.ushmm.org/content/en/article/reichstag-fire-decree) ↩

[^101]: Egypt's emergency law explained (near-continuous state of emergency ~1981–2012 under Mubarak) — Al Jazeera, 11 April 2017. [https://www.aljazeera.com/features/2017/4/11/egypts-emergency-law-explained](https://www.aljazeera.com/features/2017/4/11/egypts-emergency-law-explained) ↩

[^102]: Alexei Navalny, Russian opposition leader, dies in Arctic penal colony, 16 February 2024 — The Washington Post. [https://www.washingtonpost.com/world/2024/02/16/alexei-navalny-dies-russia-putin/](https://www.washingtonpost.com/world/2024/02/16/alexei-navalny-dies-russia-putin/) ↩

[^103]: Venezuela court upholds 15-year ban on opposition leader María Corina Machado — Al Jazeera, 27 January 2024; The Nobel Peace Prize 2025, awarded to María Corina Machado — NobelPrize.org, October 2025. [https://www.aljazeera.com/news/2024/1/27/venezuela-court-disqualifies-leading-opposition-presidential-candidate](https://www.aljazeera.com/news/2024/1/27/venezuela-court-disqualifies-leading-opposition-presidential-candidate) — [https://www.nobelprize.org/prizes/peace/2025/press-release/](https://www.nobelprize.org/prizes/peace/2025/press-release/) ↩

[^104]: Türkiye: Leading Opponent of Erdoğan on Trial (Ekrem İmamoğlu jailed March 2025, still detained 2026) — Human Rights Watch, 3 March 2026. [https://www.hrw.org/news/2026/03/03/turkiye-leading-opponent-of-erdogan-on-trial](https://www.hrw.org/news/2026/03/03/turkiye-leading-opponent-of-erdogan-on-trial) ↩

[^105]: Oligarchic Capitalism in Putin's Russia: The Khodorkovsky Case (2000 oligarch bargain; Khodorkovsky arrested October 2003) — Wilson Center. [https://www.wilsoncenter.org/publication/oligarchic-capitalism-putins-russia-the-khodorkovsky-case](https://www.wilsoncenter.org/publication/oligarchic-capitalism-putins-russia-the-khodorkovsky-case) ↩

[^106]: Orban-Style Cronyism Turns Gas-Fitter Friend Into a Billionaire (Lőrinc Mészáros) — Bloomberg, 17 December 2019. [https://www.bloomberg.com/news/articles/2019-12-17/orban-style-cronyism-turns-gas-fitter-friend-into-a-billionaire](https://www.bloomberg.com/news/articles/2019-12-17/orban-style-cronyism-turns-gas-fitter-friend-into-a-billionaire) ↩

[^107]: Maduro claimed a win. These vote tally sheets tell a different story (CNE declared victory without publishing tallies; opposition posted >80% of actas) — The Washington Post, August 2024. [https://www.washingtonpost.com/world/2024/08/04/maduro-gonzalez-election-actas-analysis/](https://www.washingtonpost.com/world/2024/08/04/maduro-gonzalez-election-actas-analysis/) ↩

[^108]: OSCE final report documents irregularities in Georgia's October 2024 parliamentary election — Radio Free Europe/Radio Liberty, December 2024. [https://www.rferl.org/a/osce-final-report-georgia-election-irregularities-2024/33247631.html](https://www.rferl.org/a/osce-final-report-georgia-election-irregularities-2024/33247631.html) ↩

[^109]: The Birth of the Gerrymander (1812 Massachusetts, Governor Elbridge Gerry) — Massachusetts Historical Society. [https://www.masshist.org/object-of-the-month/objects/the-birth-of-the-gerrymander-2008-09-01](https://www.masshist.org/object-of-the-month/objects/the-birth-of-the-gerrymander-2008-09-01) ↩

[^110]: 2026 United States intervention in Venezuela (capture of Nicolás Maduro, 3 January 2026; flown to New York on narco-terrorism charges) — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/2026_United_States_intervention_in_Venezuela](https://en.wikipedia.org/wiki/2026_United_States_intervention_in_Venezuela) ↩

[^111]: Greenland: how Trump's push to take the island reached crisis point (US pressure on Denmark, 2025–2026) — CNBC, 21 January 2026. [https://www.cnbc.com/2026/01/21/greenland-trump-denmark-timeline-of-diplomatic-tensions.html](https://www.cnbc.com/2026/01/21/greenland-trump-denmark-timeline-of-diplomatic-tensions.html) ↩

[^112]: 2025 United States strikes on Iranian nuclear sites (Operation Midnight Hammer, 22 June 2025; ~125 aircraft, three facilities) — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/2025_United_States_strikes_on_Iranian_nuclear_sites](https://en.wikipedia.org/wiki/2025_United_States_strikes_on_Iranian_nuclear_sites) ↩

[^113]: The Hitler Oath (Führereid), 2 August 1934 — Wikipedia (accessed 2026); see also USHMM on German military oaths. [https://en.wikipedia.org/wiki/Hitler_Oath](https://en.wikipedia.org/wiki/Hitler_Oath) ↩

[^114]: OPM's hiring plan includes a "blatant loyalty test" (2025 federal Merit Hiring Plan requiring essays advancing the president's agenda) — Government Executive, June 2025. [https://www.govexec.com/workforce/2025/06/lynch-opms-hiring-plan-includes-blatant-loyalty-test/405740/](https://www.govexec.com/workforce/2025/06/lynch-opms-hiring-plan-includes-blatant-loyalty-test/405740/) ↩

[^115]: China removes presidential term limits, opening the way for Xi to rule indefinitely (11 March 2018) — NPR. [https://www.npr.org/2018/03/11/592700156/china-removes-presidential-term-limits](https://www.npr.org/2018/03/11/592700156/china-removes-presidential-term-limits) ↩

[^116]: 2020 amendments to the Constitution of Russia ("zeroing" of Putin's terms, potentially to 2036) — Wikipedia (accessed 2026). [https://en.wikipedia.org/wiki/2020_amendments_to_the_Constitution_of_Russia](https://en.wikipedia.org/wiki/2020_amendments_to_the_Constitution_of_Russia) ↩

[^117]: El Salvador approves indefinite presidential re-election and extends the presidential term (31 July 2025) — NPR; see also Bukele's 2024 re-election — CNN. [https://www.npr.org/2025/07/31/nx-s1-5488299/el-salvador-approves-indefinite-presidential-reelection-extends-presidential-terms](https://www.npr.org/2025/07/31/nx-s1-5488299/el-salvador-approves-indefinite-presidential-reelection-extends-presidential-terms) ↩