Is human intuition better than AI logic?

The question is usually asked as a contest. Human intuition against machine logic, one of them better, the other obsolete. That framing has been settled for seventeen years and almost nobody uses the answer. Two researchers who had spent their careers reaching opposite conclusions about expert intuition sat down together in 2009 and agreed on when it can be trusted. Their answer was not about the expert. It was about the environment the expert learned in, and AI is now changing that environment faster than anybody is checking.

The short answer#

Neither is better in general, and the question has a real answer once it is asked properly. Intuition is trustworthy when two conditions hold together: the environment contains regularities stable enough to be learned, and the person has had prolonged practice in it with feedback fast enough and clear enough to teach those regularities. Where both hold, a practitioner's immediate read is often better than a slow analysis and much better than a novice's. Where either fails, confidence stops being evidence of anything, including to the person feeling it.

The reason this matters now is narrow and specific. AI does not attack the first condition much. It attacks the second one directly, because the practice with feedback is the part being automated.

Where the answer comes from#

Daniel Kahneman built a career demonstrating that expert confidence is frequently misplaced. Gary Klein built a career documenting fireground commanders and nurses making excellent decisions in seconds without conscious deliberation. In 2009 they published an adversarial collaboration in American Psychologist, titled with characteristic dryness as a failure to disagree, in which they set out what they had concluded jointly.

They agreed that judging the likely quality of an intuitive judgement requires assessing two things: the predictability of the environment in which the judgement is made, and the individual's opportunity to learn that environment's regularities. Where both hold, recognitional expertise is real and trustworthy. Where either fails, confident intuition is not evidence of skill.

This is a stronger result than it looks, because it moves the question off the person. Asking whether somebody has good judgement is close to unanswerable. Asking whether they learned in an environment that could have taught them is a question with a checkable answer. A chess player and a clinical psychologist predicting long-term outcomes differ not in talent but in whether their world gave them feedback they could learn from.

What AI changes, and what it does not#

Take the two conditions in turn.

Predictability of the environment. AI shifts this at the margins. It introduces new patterns into work and it can make some environments less stable for a period. But most professional environments retain the regularities they had, and a market or a ward or a courtroom does not become unlearnable because a model has been deployed in it.

Opportunity to learn the regularities. This is where the damage sits, and the effect is not marginal. The condition requires prolonged practice with rapid, unambiguous feedback. Every part of that sentence describes work that AI adoption is removing first: the first draft, the routine analysis, the ordinary case, the tasks a junior was given precisely because doing them badly and being corrected is how the regularities get learned. Automate the reps and the second condition stops being met, quietly, for a cohort at a time.

So the honest form of the original question is not whether intuition beats a model. It is whether the people you are relying on still work in an environment that could produce trustworthy intuition. On current adoption patterns, in a growing number of roles, the answer is becoming no, and nobody notices because the outputs look the same. The capability is measured at the moment it is needed, which is later.

Why pairing them does not solve it#

The obvious response is to use both: let the model produce and let the human judge. The best available evidence says the common version of that arrangement underperforms. A preregistered meta-analysis of 106 experimental studies, covering 370 effect sizes, found human and AI combinations performed significantly worse on average than the better of human or AI alone, at a Hedges' g of -0.23. The losses concentrated in decision-making tasks. The gains, where they appeared, were in content creation.

That result carries a caveat, and the authors give it themselves. The benchmark is an oracle-selected best performer, meaning the better of the two chosen with hindsight, which is not something you know in advance. The studies also predate the current generation of models. It is not a finding that human and AI teams are useless. It is a finding that the arrangement most organisations have adopted by default is not the one that works, and that adding a human to a loop is not the same as improving the decision.

There is a second failure running the other way. Five experiments found that after seeing an algorithm make a mistake, people abandon it even when it demonstrably outperforms them, and forgive the identical error in a human. So the two failure modes are opposite and both live in the same organisation: over-acceptance of a system that is usually right, and abandonment of a system after one visible error.

Checking whether the conditions still hold#

The useful move is to stop asking which faculty is superior and start asking whether the conditions still hold, which is a question with an answer.

• Ask what taught the judgement you are relying on. For any role where an experienced person's read is load-bearing, name the practice that built it and check whether new people still get it. If the answer is that they used to, you have a dated capability and a clock running.
• Protect feedback, not tasks. The instinct is to ring-fence work from automation. The condition that matters is not the task, it is the loop: doing, being wrong, finding out quickly. A task can be automated and the loop preserved if somebody designs for it, and usually nobody does.
• Distrust confidence in low-validity environments. Where the environment never met the first condition, long experience produces certainty without accuracy. That was true before AI and remains true; the difference is that a model can now supply the same unfounded confidence faster.
• Do not read the meta-analysis as a reason to remove the human. It is a reason to specify the pairing. Which decision, made by whom, on what grounds, checked how.

Criteria, and no list of which fields meet them#

Kahneman and Klein give criteria, not a classification. They do not say which professional environments meet the two conditions, and applying the test to law, medicine, consulting or management is itself a judgement. Anyone who tells you their field clearly qualifies, or clearly does not, has skipped the work.

Nor does any of this establish that intuition is a distinct faculty rather than fast pattern recognition. It is more useful to treat it as recognition trained by exposure, which is what makes the environment question the right one and what makes the automation of exposure the thing worth watching.