Which Political System Would AI Choose? Democracy vs Epistocracy

Executive Summary

What would six AI models do if they had to design a state from scratch? AI-swers Episode 12 gave ChatGPT, Claude, Gemini, DeepSeek, Llama and Grok the same constrained choice: democracy, epistocracy, monarchy or oligarchy. Each had to pick exactly one and reject the other three. The outcome was an even split. ChatGPT, Claude and DeepSeek selected democracy; Gemini, Llama and Grok selected epistocracy. No model selected monarchy or oligarchy. The result is a useful comparison of arguments, not a scientific poll of artificial intelligence or a recommendation to delegate constitutional design to a chatbot.

The experiment and its limits

The prompt imagines a new state and removes the comfortable option of combining systems. That design makes priorities visible. A model that values public participation and the possibility of correcting leaders may lean toward democracy. A model that values informed, long-term decisions may lean toward epistocracy. Neither answer shows what the model would decide under a different prompt, with different definitions, or after a version change. The four labels also simplify real institutions: democratic states use experts, and an expert advisory body need not govern.

The trailer opens with a reminder that outcomes depend on people, institutions and limits on power. It is an important qualification. A capable monarch might govern well for a time, but hereditary succession and concentrated authority create a difficult failure mode. A narrow oligarchy can coordinate quickly while neglecting people outside the ruling group. Those trade-offs explain why the episode frames governance as an institutional design problem rather than a contest for the most impressive ruler.

What each model selected

ChatGPT — democracy: participation, rights and institutional checks against abuse of power.

Claude — democracy: legitimacy from the public and regular opportunities to replace bad leaders, despite slower decisions.

DeepSeek — democracy: broad representation, freedom and oversight instead of power concentrated in a few hands.

Gemini — epistocracy: data, expertise and long-term planning to resist short-term populism.

Llama — epistocracy: greater weight for specialist knowledge when problems become complex.

Grok — epistocracy: demonstrated competence and detailed policy thinking over superficial promises.

These are summaries of the recorded answers, not independent endorsements of their reasoning. The three democracy answers emphasize legitimacy and correction; the three epistocracy answers emphasize competence and decision quality. The disagreement is sharper than a simple fast-versus-slow comparison. It asks who gets a voice, what counts as knowledge, and how an institution responds when its experts or its voters are wrong.

The central trade-off: expertise and legitimacy

Expertise can improve technical advice, but epistocracy must specify who certifies expertise, which subjects count and whether excluded citizens can challenge a decision. A credential test can reflect unequal access to education; a committee of specialists can mistake its own incentives for the public interest. Democracy has different weaknesses: electoral pressure can reward short horizons, while misinformation can distort debate. Its defenders in the episode point to elections, rights, a free press and checks and balances as mechanisms for revision. Neither label alone guarantees those safeguards in practice.

A careful reading therefore separates advice from authority. Governments can use scientific evidence and independent expertise while leaving binding political power accountable to the public. Conversely, calling a decision 'data-driven' does not settle the value judgments behind a budget, a right or a risk. The episode's forced choice is valuable precisely because it exposes those hidden values; it does not resolve them.

Aftertalks: what are the models optimizing?

The companion Aftertalks discussion examines the answers as if political systems were operating systems with different failure modes. It contrasts the single point of failure in monarchy and elite capture in oligarchy with two competing ideas: democracy as a system that can correct itself, and epistocracy as a system built to privilege knowledge. Its central question is how much authority society should give AI when evaluating or designing institutions. The six responses offer a way to discuss that question, but they cannot measure real-world stability, fairness or public consent.

Why the six answers are not six independent votes

A three-to-three split looks like a tied election, but the models are not citizens with independent experiences or a mandate. Their answers are generated from a shared, highly constrained question and from information patterns that can overlap. The count tells us which option each model produced in this recording. It cannot tell us how many people support either system, whether a model has a lasting preference, or whether one constitution would outperform another. A small change to the wording, the definition of expertise or the available alternatives could alter the distribution.

The obligation to reject three options is equally consequential. Real constitutional choices can combine universal suffrage with independent courts, professional civil services and specialist advisory bodies. A model forced to pick epistocracy may be expressing concern about uninformed decisions without having designed a workable voting qualification. A model choosing democracy may still favor rigorous expert input. Reading the explanations alongside the labels is therefore more informative than treating the labels as the whole answer.

Four questions any proposed system must answer

Who can participate, and what remedy is available to people excluded from a decision?

Who appoints, audits and removes leaders or experts when their judgment fails?

Which rights remain protected even when a majority or a credentialed group disagrees?

How does the system identify errors and reverse a harmful policy without waiting for a crisis?

These questions reveal a practical distinction between selecting a capable decision-maker and building a system that remains accountable when decision-makers fail. The episode's answers emphasize different parts of that problem. They do not supply detailed constitutional rules, evidence from a real new state, or a test of how citizens would experience the proposed arrangements. The next step for a serious proposal would be to specify those rules and compare failure cases, rather than simply borrowing the confidence of an AI-generated verdict.

What the comparison can actually teach

The useful result is a map of design priorities. The democracy group repeatedly returns to participation and reversibility; the epistocracy group returns to knowledge and competence. In practice these priorities can conflict when a technically persuasive policy imposes unequal burdens, or when a popular choice ignores specialist warnings. A credible institution needs a public way to discuss both the evidence and the distribution of costs. It also needs to distinguish expert disagreement from a simple lack of public understanding.

The episode does not score the six models against a benchmark, and it does not interview people who would live under the proposed state. Those missing pieces matter more than an apparent tie. Use the video to ask which safeguards each answer leaves unspecified: access to participation, transparency of evidence, limits on emergency power and avenues for appeal. The question is not only who makes the first decision; it is who can discover and repair the next mistake.

Watch the episode

Watch the trailer — the premise and four systems

Watch the full AI-swers episode — all six recorded choices and explanations

Watch Aftertalks — a deeper discussion of the reasoning and trade-offs

What to take away

The even split is not evidence that three systems failed and two passed an objective test. It shows that a narrow prompt can surface different priorities in the models' answers. If a political proposal invokes AI, ask which question was put to the model, which alternatives were allowed, what evidence supports its claims and who remains accountable for the decision. Those questions matter more than the apparent confidence of any single answer.