Francis Fukuyama declared liberal democracy the final form of human government in 1992. Thirty-four years later, 72 percent of the world’s population lives under autocracy, Freedom House has documented eighteen consecutive years of global democratic decline, and autocracies outnumber democracies for the first time since 2002. The thesis of this essay is not that democracy should be abandoned. It is that democracy should be examined with the same structural rigour we would apply to any other system — and that refusing to do so is not loyalty. It is negligence dressed as piety.

A system property, not a value judgment

Democracy structurally maximises a specific kind of variance: periodic leadership turnover, policy reversals driven by electoral cycles, legitimacy tethered to shifting public sentiment. This is not a criticism. It is a description. The Paris Climate Agreement has been joined, withdrawn from, rejoined, and withdrawn from again across four consecutive administrations. The Mexico City Policy has toggled with every change of party since Reagan. The Iran nuclear deal was negotiated, abandoned, tentatively revived, and abandoned again. Each reversal was produced by the system operating exactly as intended.

The instinct is to compare this with autocratic stability and declare one preferable. That comparison is less straightforward than either side admits. Autocracies do minimise visible political process variance — leadership continuity, long planning horizons, insulation from popular mood. But on outcomes, the picture inverts sharply. Between 1990 and 2009, autocratic country-decades showed an economic growth variance of 23.1, against 5.5 for democratic ones — roughly four times the volatility. Both the best and worst performers in terms of growth rates are more likely to be autocracies. The stability is real at the surface level. The outcomes underneath are a roulette wheel.

So the honest framing is not that democracy is unstable and autocracy is stable. It is that democracy distributes instability visibly and regularly through its own procedures, while autocracy concentrates it invisibly until it arrives catastrophically. Neither pattern is safe. But only one of them is examined critically in mainstream political discourse.

The only system that contains its own off switch

Karl Popper identified the paradox decades ago: democracy permits the majority to decide that a tyrant should rule. Taken to its logical conclusion, freedom negates freedom, tolerance negates tolerance, and democracy negates democracy. These are not hypothetical extrapolations. They are the historical record.

Weimar Germany did not fail to prevent Hitler. It enabled him through its own constitutional procedures. Article 48 — the emergency powers provision — was invoked sixty-three times by President Ebert and sixty times by President Hindenburg before Hitler ever touched it. The Enabling Act passed by a vote of 441 to 94, formally meeting the requirements for constitutional amendment. The Weimar Constitution was never repealed. It simply ceased to matter.

Hugo Chávez was democratically elected in 1998. His first act was to demand a constitutional convention — a procedure found nowhere in the existing constitution. The Supreme Court, under significant political pressure, allowed it. One chief justice reportedly stated that the court had committed suicide rather than wait to be killed by the Assembly. Chávez’s 60 percent popular vote translated to over 90 percent of assembly seats. The assembly’s first act was to suspend Congress and the Supreme Court. The entire process was, at every step, procedurally legitimate.

Viktor Orbán’s Fidesz won 52.7 percent of Hungary’s vote in 2010 but received 68 percent of parliamentary seats — enough for a constitutional supermajority. The bar for writing a new constitution required a four-fifths majority, but the rule itself could be overturned by a two-thirds majority. Fidesz did exactly that. A new constitution was drafted behind closed doors, debated for nine days, and passed on a party-line vote. Every action was legal under the provisions of the system it was dismantling.

Erdoğan’s 2017 referendum passed 51 to 49, under a state of emergency, with opposition politicians imprisoned and media controlled. The Venice Commission called it a threat to democracy. The system’s own mechanisms produced the result.

The counterargument — articulated carefully by Manjeet Ramgotra — is that democracy does not truly vote itself out. By the time the decisive vote arrives, democratic institutions have already been hollowed from within. This is a useful correction to the timeline, but it does not change the structural point. It refines it. Democracy does not contain a guaranteed mechanism for its own preservation. It relies on cultural norms it does not generate, institutional strength it does not automatically rebuild, economic surplus it does not ensure, and restraint it cannot compel. When those external inputs erode, the system has no internal failsafe. Every other governance system can only be destroyed from outside or by internal coup. Democracy can be destroyed by its own operating procedure, provided someone hollows it out patiently enough.

The median and its ceiling

A decision system designed around median preference has, by design, capped its outcomes at median quality. This is not a failure. It is the intended output. Kenneth Arrow proved mathematically in 1950 that no rank-order electoral system can simultaneously satisfy a set of reasonable fairness criteria. Bryan Caplan documented four systematic biases — anti-market, anti-foreign, make-work, and pessimistic — that voters carry into the booth not because they are stupid but because delusional political beliefs are free: there is no practical cost to holding them and no practical benefit to correcting them.

The strongest epistemic defence of democracy — Hélène Landemore’s argument that cognitive diversity in deliberation tracks truth better than expert panels — is genuinely compelling under its own conditions. The Condorcet Jury Theorem shows that as the number of independent, better-than-random voters increases, the probability of a correct collective answer approaches certainty. The difficulty is that the conditions are precisely what democracy does not ensure. Voters are not independent — media, social pressure, and algorithmic curation see to that. They are not sincere — strategic voting is rational. And they are not reliably better than random on complex policy questions that require domain expertise to evaluate.

Every field that has produced exceptional results — science, engineering, medicine, technology — explicitly rejects democratic consensus as a method for determining truth or direction. The implicit assumption that governance should work differently from every other domain of human excellence has never been justified. It has only been asserted so often that most people can no longer distinguish the assertion from an argument.

The prerequisites democracy cannot generate

Seymour Martin Lipset established the relationship in 1959: the more well-to-do a nation, the greater the chances it will sustain democracy. Przeworski sharpened this into something more precise: wealth does not create democracy, but no democracy with a per capita GDP above Argentina’s 1975 level has ever collapsed. Below that threshold, the twenty-year survival probability drops to roughly one in three.

Democracy quietly requires basic education, spare cognitive bandwidth, relative economic security, trust that institutions will not collapse overnight, and the ability to lose politically without existential fear. Robert Putnam documented the decline of every one of these inputs in the United States — membership in religious groups, unions, PTAs, veterans’ organisations, fraternal groups — attributing 25 percent to electronic entertainment and 50 percent to generational change. The civic infrastructure democracy depends on is eroding in the wealthiest democracy in history.

The far-right surge across Europe is the predictable output. Seven EU member states now have far-right parties in government. The AfD won 20.8 percent of Germany’s national vote in February 2025. The Freedom Party took 29.1 percent in Austria. In the former industrial heartlands of eastern Germany, the combination of welfare populism and anti-immigration messaging is not an aberration. It is democracy transmitting economic despair into political power — which is exactly what it was designed to do. The system does not filter for quality. It transmits whatever the population is.

What the body count actually says

The most devastating counterargument to any structural critique of democracy is not theoretical. It is 170 million dead.

R.J. Rummel calculated that governments killed 170 million of their own people during the first 88 years of the twentieth century — nearly 4.5 times the 38.5 million battle deaths from all wars in the same period. Fifteen regimes killed more than a million each. Democratic regimes accounted for between 1 and 2.2 percent of the total. Amartya Sen established that no substantial famine has ever occurred in any independent democratic country with a relatively free press. The Great Leap Forward alone killed an estimated 30 million — the deadliest famine in human history, produced by exactly the kind of centralised, unaccountable decision-making that an uncritical reading of this essay might seem to endorse.

It would be dishonest to soften this. Every alternative to democracy that has been tried at scale has produced body counts that democracies have not. The democratic peace — the observation that democracies virtually never go to war with each other — remains statistically robust across 140 years of data. Autocratic economic growth variance runs four times higher than democratic variance. The gamble is not symmetrical.

Any serious structural critique of democracy must hold this in one hand while holding the system’s design limitations in the other. The fact that alternatives have been catastrophically worse is not evidence that democracy is structurally sound. It is evidence that the alternatives tried so far have been worse — which is a different and far less comforting claim. The challenge is not to replace democracy with something from the twentieth century’s catalogue of horrors. It is to ask whether the catalogue is complete.

South Korea provided the strongest recent evidence for democratic resilience in December 2024, when President Yoon declared martial law and lawmakers annulled it within two and a half hours. Yoon was impeached, arrested, and convicted on insurrection charges — the system’s antibodies working in real time. It is worth noting, simultaneously, that the system was hours away from a very different outcome, that the coup failed partly because of incompetent military mobilisation rather than institutional design, and that the episode demonstrated both the system’s capacity for self-correction and its continued vulnerability to the attempt.

The capacity that comes first

The pattern across every durable governance transition is consistent enough to be uncomfortable. Britain built a professional civil service through the Northcote-Trevelyan reforms of 1854, decades before mass suffrage. Prussia’s bureaucratic capacity preceded democracy by over a century. South Korea built its industrial and institutional base under Park Chung-hee’s authoritarian rule between 1961 and 1979 — and the middle class his economic reforms created was precisely what demanded and sustained democratisation afterward. Singapore’s GDP per capita became one of the highest in the world under Lee Kuan Yew’s managed authoritarianism.

The real axis of governance is not democracy versus autocracy. It is institutionalised power versus personalised power. China is not successful because it is autocratic — it is successful because it built bureaucratic, fiscal, and infrastructural capacity across decades. Venezuela is not a failure of autocracy — it is a failure of state capacity, a kleptocracy that pillaged its own institutions until nothing structural remained. North Korea is not evidence that autocracy fails. It is evidence that personalised, dynastically inherited power without institutional constraint fails. Comparing China to North Korea as evidence about autocracy is selection bias, not analysis.

Samuel Huntington stated the uncomfortable sequencing truth in 1968: men may have order without liberty, but they cannot have liberty without order. Authority has to exist before it can be limited. The countries that democratised successfully built capacity first. The countries where democracy was introduced before capacity — Iraq, Libya, much of post-colonial Africa — did not produce liberal democracy. They produced instability, followed by either autocratic reversion or state failure.

Acemoglu and Robinson’s Nobel Prize-winning work complicates this, and the complication deserves its full weight. Their framework argues that inclusive institutions — which correlate with democratic governance — consistently outperform extractive ones over the long term. Countries that switch to democracy experience a 20 percent GDP increase over 25 years. The correlation is robust. But it is also worth noting that their framework is about inclusive versus extractive, not democratic versus autocratic in the narrow sense. Singapore has extractive political institutions and inclusive economic ones. The framework may support the capacity argument as much as the democracy argument.

The framework that is coming whether the framework is ready or not

Every governance system humanity has produced emerged from a specific set of coordination problems under specific material conditions. Tribal governance solved coordination at the scale of kinship groups. City-states solved it at the scale of agricultural surplus. Empires, feudalism, constitutional monarchy, representative democracy — each was a response to a coordination problem created by changes in technology, population, and economic structure.

AGI changes the coordination problem. Not incrementally. Categorically. Once decision-making, modelling, forecasting, and optimisation can be performed by non-human systems at speeds and scales no human institution can match, the question of who steers civilisation stops being answerable within the existing vocabulary. Dario Amodei has noted that there is no strong reason to believe AI will preferentially advance democracy and peace, and that AI-powered authoritarianism is too terrible to contemplate. Both observations are correct simultaneously — and both point at the same conclusion: the governance categories designed for a world of human cognitive limits are about to encounter a world without them.

The post-AGI governance question will not be democracy versus autocracy. It will be: who sets the objectives, and how reversible are those decisions? Consent may shift from continuous voting to one-time constitutional encoding. Autonomy may shift from political voice to guaranteed exit, privacy, and non-interference. Human rights may shift from social norms to hard constraints encoded into systems no human can override. Robin Hanson’s futarchy — vote on values, bet on beliefs — and Glen Weyl’s quadratic voting are early sketches of what governance might look like when the bottleneck is no longer information aggregation but objective-setting. They are imperfect sketches. They are also evidence that the design space is larger than the binary we have been offered.

The mythology and what it costs

Churchill’s famous line — democracy is the worst form of government except for all the others that have been tried — is quoted so often it has acquired the force of settled argument. It is worth noting that Churchill was quoting someone else, that the remark was explicitly hedged, and that the operative phrase is “that have been tried.” The sentence is not a defence of democracy. It is an admission that the catalogue of alternatives is short, and that shortness is being mistaken for completeness.

Universal suffrage is between 60 and 130 years old, depending on the country. Athenian democracy lasted 180 years with interruptions. The evidence base for democracy as a permanent form of governance does not exist because the experiment has not run long enough to generate it. This is not a reason to discard democracy. It is a reason to stop treating it as a destination rather than a phase — one that may be succeeded by something better, or something worse, depending entirely on whether the transition is managed by people who have examined the system honestly or by people who refused to.

The honest position is not to abandon democratic values. It is to stop mythologising the system that currently houses them. Mythologising a fragile system is the surest way to prevent it from adapting — and a system that cannot adapt to the century approaching has already decided how it ends.