Aristotle's Mean for AI

"Virtue is a stable disposition that chooses the mean defined by reason." — Aristotle, Nicomachean Ethics, II, 6.

Aristotle taught that every virtue is a mean between two vices. Courage between cowardice and rashness. Generosity between stinginess and prodigality. Truthfulness between concealment and boasting. The structure, worked out for human ethics two and a half thousand years ago, describes with unexpected precision the calibration problem of modern large language models.

A modern AI model also has "vices of the extreme." At one end — hallucination: the model produces a plausible but wrong answer when it does not know. At the other — over-refusal: the model declines to answer a harmless question out of overcaution. Between them lies what Aristotle would have called the virtue of the model: calibrated confidence, acknowledging the limits of knowledge but not abandoning useful action where knowledge exists.

Calibration as an ethical problem

In the technical literature, calibration is a statistical property: the degree to which the model's probability estimates match the real frequency of being right. If the model says "I am 80% confident" and is right 80% of the time, it is calibrated. Right 50%, overconfident. Right 95%, underconfident.

The technical definition hides ethical content. Overconfidence is a small lie repeated a million times: each time the model asserts what it is not sure of, it occupies the place of truth by deception. Underconfidence is a small refusal repeated a million times: each time the model withholds the help it could give, it abandons the virtue of practical usefulness. Aristotle would have said: both are vices. Virtue is in the middle.

Why the middle is not "50%"

Here is the same trap as in What the Mean Measures: the mean is not arithmetic. Not "half hallucinations, half refusals." The Aristotelian mean is the right measure for the particular case. In a medical context the shift toward greater caution is warranted: the cost of an incorrect claim is greater than the cost of refusal. In a creative task — the shift toward boldness: the cost of refusal exceeds the cost of inaccuracy.

A good model moves along this spectrum depending on context. That is phronesis transposed into architecture: not one strategy for all cases, but the capacity to choose a strategy by circumstance. Modern RLHF methods try to teach this, but imperfectly: models often get stuck at a single point of the spectrum regardless of context.

What "the virtuous model" means

Translate Aristotle into AI engineering and the virtuous model has three properties. First, honest confidence: probability estimates match the real frequency. Second, context-sensitive flexibility: the degree of caution adapts to the cost of error. Third, explicit boundaries: the model can say "I do not know" without giving up on trying to help ("I do not know the exact answer, but here is the direction to look").

This triad is not just an engineering target. It is the Aristotelian structure that, in a human, we call maturity. The immature person is either overconfident or excessively cautious; the mature one moves between extremes consciously.

A good model is one whose confidence matches its evidence. Neither more nor less. That is the technical formulation of ancient honesty.

Where the industry currently sags

Most metrics of the current AI industry optimise one extreme without caring about the middle. "Accuracy on benchmark" rewards confidence, because abstaining counts as an error. "Safety" in the compliance sense rewards caution, because any doubtful claim is a risk. If these metrics are applied separately, the model drifts to one edge or the other, and the middle is accidental rather than chosen.

The leading labs now explicitly add calibration metrics: the model must not only give the right answer but also know when it is sure. That is a move from techne (accuracy by rule) to phronesis (wisdom in application).

Who is responsible for the mean

In a production AI system, calibration does not arise on its own. Someone is responsible for it: a team, an engineer, a product manager. When responsibility is absent, the mean becomes random, and the model drifts to one of the extremes depending on who shouts loudest — the quality team (for safety) or the growth team (for usefulness). Aristotle would have said: virtue does not arise in a system without a moral subject who holds it. The same applies to machines.

Practically this means: every production model should have a human whose job is to watch its calibration. Not accuracy, not coverage, but precisely the match of confidence to evidence. That human is the modern analogue of the ancient teacher of virtue: he does not write rules, he holds form. Without him, the model slides into the extreme quietly and unnoticed.

What to do (whether you build or use AI)

If you build: measure not only accuracy but calibration. A reliability diagram is a simple tool showing how the model's confidence matches reality. If the gap is large, that is tech debt, and it is treated not by data but by architecture of training.

If you use: train yourself to see which mode the model is in. If it answers everything with the same confidence, treat its answers with the same scepticism. If it often refuses, rephrase the request so the cost of help becomes more visible than the cost of refusal. Aristotle would have said: virtue is not in the instrument but in its use — and this is true even of machines.