Kahneman and Klein (2009): Two Conditions for Trusting Intuition── Intuition can be trusted only where the environment is regular and we have learned from feedback; the strength of our certainty is no guide.

Anyone who uses generative AI at work knows the moment: you read a draft the AI produced and feel that something is off, without being able to say why. How far can we trust an intuition we cannot explain? Only as far as a domain that is regular and in which we have learned from feedback; how certain we feel tells us nothing about that.

01The range of trustworthy intuition is set by environment and learning, not certainty

That answer rests on a joint conclusion reached by two psychologists. In 2009 Daniel Kahneman, who had spent his career on the errors of intuition, and Gary Klein, who had spent his on the reliability of expert intuition, wrote a paper together. They had long disagreed, and the paper set out to see how far they could agree. They concluded that the quality of an intuition should be judged by two conditions.

In the same paper they wrote that subjective confidence is not a reliable indicator of whether a judgement is accurate. Feeling strongly that something is so and being right about it are two separate matters.

This conclusion applies directly to the work of reading AI output. When a generative AI draft makes you uneasy, or when it reassures you that it is probably fine, the strength of the feeling should not decide what you do. What should decide is whether you have learned from feedback in that domain. Unease is a reason to check. Comfort is not a reason to skip checking.

The sections that follow work through what an unexplained intuition is, where it meets AI output, why certainty misleads, what increases reliance on AI, and what remains unknown.

02Knowledge beyond words exists, and the body signals risk before awareness

Before treating intuition as conditional, it is worth asking whether an intuition we cannot explain knows anything at all.

In The Tacit Dimension, the philosopher Michael Polanyi rebuilt his account of knowledge from a simple fact: we can know more than we can tell. We can recognise a face without being able to say how we recognised it. What we know and what we can explain do not coincide.

Neuroscience adds evidence. In 1997 Antoine Bechara, Antonio Damasio and colleagues gave healthy participants a card-gambling task in which some decks paid off and others lost money. Participants began choosing advantageously before they realised which strategy worked best, and showed skin conductance responses before drawing from the losing decks.

The order matters. Bodily response came first, then advantageous choice, then awareness of the strategy. Judgement was already moving before it could be explained. An intuition that cannot be explained is not empty for that reason.

The experiment also shows something else: intuition grew in a task that offered a chance to learn. Every card drawn returned a gain or a loss. Section 04 comes back to that condition.

03Intuition meets AI output in workflows in which AI proposes and a person reviews

If unexplained intuition is real, the place where it meets AI output is the step at which AI makes a proposal and a person decides whether to accept or reject it.

That step has been studied at length in medicine. A systematic review by Goddard and colleagues, published in 2012, selected 74 papers from 13,821 and examined how often over-reliance on automation occurs in clinical decision support, what strengthens it and what weakens it.

Where the conditions in the middle column hold, you can later find out whether your unease was right. Where the right-hand column applies, AI output is adopted without being checked, and you never learn whether your unease was justified.

Work in which AI drafts and a person reads and decides has spread well beyond medicine, into document review, translation and checking summaries. Each time, the choice between following your own intuition and following the AI's output becomes a single decision. The review, however, covered clinical decision support. It cannot tell us whether reliance occurs as often in other kinds of work. What it does show is that other work contains steps with the same structure.

04Certainty misleads because it forms even without any chance to learn

Deciding each time whether to follow intuition, we are tempted to lean on how certain we feel. Kahneman and Klein rejected that, for reasons that lie in how intuition develops.

According to them, intuitive skill develops when the environment is sufficiently predictable and a person has had a long opportunity to learn its regularities. The judgements of fire-ground commanders, which Klein studied, meet this condition. Similar situations recur and the results of decisions come back clearly.

The trouble is that certainty forms just as readily where those conditions are missing. My reading of the paper is that certainty depends on how easily an answer turns into a coherent story, not on whether that story is backed by regularities in the world.

So the strength of certainty alone cannot separate skill from mistaken belief. To separate them, you have to look outside the feeling. Is the domain regular? Have I built up experience in it where right and wrong answers came back to me?

Applied to reading AI output, two kinds of certainty are possible: “this draft is wrong” and “this draft is fine”. Neither, on its own, says whether it is right.

05Over-reliance on AI grows with trust, load and time pressure and eases with accountability

If certainty is no guide, the next question is under what conditions a person's judgement gives way to AI output, and what restores it. Goddard's review lists the factors.

On the side that increases reliance, the review listed trust in automation and confidence in one's own judgement, workload, task complexity and time constraints. When people are busy, facing hard problems and short of time, they are more likely to follow AI output.

On the side that reduces it, the review listed training, emphasising the user's accountability, where advice is placed on the screen, attaching confidence levels to output, and presenting output as information rather than as a recommendation. Most of these depend less on individual resolve than on how work is organised and how screens are designed.

A caution is needed. A 2010 paper by Parasuraman and Manzey found that automation bias occurs in both novices and experts and cannot be prevented by training or instructions. Training appears among the factors that reduce reliance, but it is not enough by itself.

06Unease is a reason to check; comfort is never a reason to skip checking

From the factors that raise and lower reliance, three things follow that should change how intuition is handled at work.

First, the strength of certainty does not show that an intuition is right. Kahneman and Klein concluded that certainty is not a reliable guide to accuracy, and, as section 04 showed, certainty arises whether or not the domain is regular. Strong certainty is not a reason to skip checking.

Second, unexplained unease can arrive before awareness. In Bechara's experiment, bodily responses preceded awareness of the strategy. Being unable to give a reason is not a reason to discard the unease. Even without a reason, record where you felt it and check that spot against primary sources.

Third, reliance on automation occurs among experienced people too, as Parasuraman and Manzey reported. On the regulatory side, Article 14(4)(b) of the EU AI Act requires that people overseeing high-risk AI remain aware of the tendency to rely automatically or excessively on its output. Years of experience are not an exemption from checking.

Taken together, the three points turn the handling of unease into a cycle. Record the unease with its location and time, check it against evidence, return the result, right or wrong, to yourself, and use it in the next judgement. Done this way, each moment of unease becomes one of the chances to learn described in Section 04.

07Whether reading AI output trains intuition depends on whether feedback comes back

Recording unease leaves one question: whether those records can build intuition over time, and what has not yet been established.

Applying Kahneman and Klein's conditions, reading AI drafts can become intuitive skill only if two things hold. The AI's errors must have some regularity, and whether your unease was right or wrong must come back to you afterwards.

On the first, the way a model errs may change each time it is updated. I do not know whether any one pattern of error lasts long. On the second, where drafts are corrected and used without keeping a record of what was corrected, no feedback returns.

Whether that feedback actually returns in workplaces that use AI, and whether intuition grows when it does, has not yet been established by research. Everything in this section is inference from the two conditions.

Because it has not been established, the record matters. If you keep your moments of unease and whether they proved right, you can later count for yourself where your intuition holds and where it fails.

An intuition we cannot explain may be trusted only within domains that are regular and where we have learned from feedback. Facing AI output, record unease as a reason to check, and never treat comfort as a reason to skip it.

Whether to trust intuition is not settled once. By counting hits and misses over time, you can check afterwards how far your own intuition reaches.