August 16, 2026 · Asanka Herath · 13 min read

Hot vs. Cold Cognition

#AI

cold cognition

/kōld käɡˈniSHən/

noun reasoning that is deliberate, rule-based, and largely indifferent to the reasoner’s emotional state.

hot cognition Abelson, 1963

/hät käɡˈniSHən/

noun judgment that is entangled with feeling — a sense of rightness or wrongness arrived at before, or instead of, any explicit reasoning.¹

Ask someone why they didn’t trust a person the moment they walked into the room, and you’ll get an answer eventually. But the distrust came first. The reasons are a story told afterward, to a mind that already knew.²

That’s hot cognition. It’s not sloppy thinking, and it’s not the opposite of intelligence. It’s judgment that is inseparable from feeling, built up over a lifetime of encounters that never got written down anywhere, least of all in language.

Cold cognition is the other kind. Given the rules, follow them. Given the data, compute the answer. It doesn’t need you to have lived any particular life. It just needs you to apply the procedure correctly.

Most of what we’ve automated so far is cold cognition. Spreadsheets, compilers, search indexes, chess engines — all of it is procedure applied faithfully, at a scale and speed no person can match.

Large language models complicate the picture because they can now produce something that reads like judgment. Ask one whether a business plan is convincing, whether a sentence is cruel, whether a stranger’s argument is trustworthy, and it will answer with what looks exactly like a vibe.

But it isn’t one, not in the sense that matters. It’s statistical mimicry dressed up like genuine compassion.

A model’s “sense” of a thing is a statistical residue of what millions of people have said about similar things. It’s an average taken over other people’s hot cognition, compressed into weights. When it renders a judgment, it is not consulting a felt history of its own. It is predicting what a person — many people, blended — would say next.

That can be extraordinarily useful. It can also be right more often than any single person, in the way that a crowd’s guess at the weight of an ox beats most individual guesses.³ But being a good aggregate of other people’s judgment is not the same thing as being a source of judgment.

The distinction matters because we are in a hurry to automate everything we can, and cold and hot cognition don’t fail the same way when automated carelessly.

Automate cold cognition and the worst case is a bug. Automate hot cognition — hand over the vibe check, the trust call, the sense of whether something is off — and you’ve quietly replaced a person’s lived judgment with a blended echo of everyone else’s, laundered to look like a heartfelt sentiment.

None of this means AI has no place in judgment-adjacent work. It means the place is augmentation, not substitution: surface the pattern, flag the precedent, do the cold-cognition legwork that judgment depends on — and leave the actual call, the part that has to come from a life actually lived, to the human who’s still standing there when the consequences come home to roost.

Hot vs. cold cognition in literature

The terms are old, but the argument keeps getting re-litigated with new tools.

Robert Abelson’s 1963 chapter is where “hot cognition” enters the record, and it’s worth pausing on the context: he was building computer simulations of personality — specifically, of how people revise their beliefs — and trying to model how emotionally loaded input distorts that revision compared to neutral, algorithmic updating Abelson, 1963. Janet Metcalfe and Walter Mischel gave the framework its fullest psychological treatment in 1999, modeling self-control as a contest between a “hot,” impulsive, reflexive system and a “cool,” slow, strategic one — the paper most commonly cited when people invoke hot/cold cognition today Metcalfe & Mischel, 1999.

The more directly relevant work is recent and empirical rather than theoretical. A team at Sapienza University of Rome benchmarked six large language models against expert ratings (NewsGuard, Media Bias/Fact Check) and human evaluators on judgments of news source credibility. The models often matched expert conclusions, but the paper argues the match is largely coincidental: models lean on lexical association and statistical priors rather than the contextual, normative reasoning experts actually use to get there. They call the resulting mismatch between apparent and actual process “epistemia” — the illusion of knowledge that surface plausibility creates Loru et al., 2025. A companion commentary sharpened the framing further, describing the models’ output as “counterfeit judgments”: convincing enough to pass as the real thing, but produced by a different process entirely Perc, 2025. That’s close to a direct, empirical version of the claim this post is making by argument alone — that a model’s vibe is a simulation of judgment, not a source of it.

Computer simulation of 'hot cognition' - Computer Simulation of Personality. Abelson, R. P. 1963. pp. 277-302 (Google Scholar) (Google Search)

The paper that named hot cognition. Abelson builds a computer simulation of belief processing in which the material being processed is affectively charged rather than neutral, and shows the machinery working to preserve an existing attitude — finding grounds to reject what threatens it — rather than updating on what it is given.

Vox Populi - Nature. doi.org Galton, F. 1907. pp. 450-451 (Google Scholar) (Google Search)

Galton tabulates the entries in a weight-judging competition at a West of England fat stock show, where roughly eight hundred people guessed the dressed weight of an ox, and finds the middlemost estimate within a fraction of a percent of the true figure. He offers it as evidence about the trustworthiness of a democratic verdict; it is now cited as the origin of the wisdom of crowds.

The Emotional Dog and Its Rational Tail: A Social Intuitionist Approach to Moral Judgment - Psychological Review. doi.org Haidt, J. 2001. pp. 814-834 (Google Scholar) (Google Search)

Moral judgment arrives first, as fast intuition, and the reasoning follows to justify it. Haidt's social intuitionist model casts explicit moral reasoning as largely post hoc — the rational tail wagged by the emotional dog — and argues its real work is persuading other people rather than reaching the conclusion in the first place.

The simulation of judgment in LLMs - PNAS. doi.org Loru, E., Nudo, J., Quattrociocchi, W. 2025. (Google Scholar) (Google Search)

Six LLMs and human non-experts are put through the same structured evaluation procedure — select criteria, retrieve content, justify a verdict — and benchmarked against expert news-reliability ratings. The outputs align; the criteria visibly guiding them do not, with the models leaning on lexical association and statistical prior where people reason from context. The authors name the failure mode epistemia: the illusion of knowledge that arises when surface plausibility stands in for verification.

A Hot/Cool-System Analysis of Delay of Gratification: Dynamics of Willpower - Psychological Review. doi.org Metcalfe, J., Mischel, W. 1999. pp. 3-19 (Google Scholar) (Google Search)

A two-system account of willpower: a cool, slow, strategic “know” system and a hot, fast, stimulus-driven “go” system. Delay of gratification holds when attention is arranged so the cool system keeps control, and collapses when the hot system takes over — which is why what the tempting thing is made to represent matters more than how badly it is wanted.

Telling more than we can know: Verbal reports on mental processes - Psychological Review. doi.org Nisbett, R. E., Wilson, T. D. 1977. pp. 231-259 (Google Scholar) (Google Search)

A review of studies in which people confidently explain behaviour that was in fact driven by something they never noticed, and fail to report influences that demonstrably moved them. Nisbett and Wilson conclude that introspective reports are not privileged access to mental process but plausible theories about it, drawn from the same stock of explanation an onlooker would reach for.

Counterfeit judgments in large language models - PNAS. doi.org Perc, M. 2025. (Google Scholar) (Google Search)

A commentary on Loru et al. in the same volume, putting their result under the heading of counterfeit judgment: an output carrying the form of a judgment without the process that would make it one, and passing as genuine because what gets inspected is the verdict rather than what produced it.

Feeling and Thinking: Preferences Need No Inferences - American Psychologist. doi.org Zajonc, R. B. 1980. pp. 151-175 (Google Scholar) (Google Search)

Zajonc argues that affective reactions can be faster than, and independent of, the cognitive appraisals supposed to produce them. Preference behaves as a system of its own rather than as a product of judgment: liking can be established before, and without, recognising the thing liked.

The idea that affect can precede and operate independently of reasoned inference is most associated with Robert Zajonc’s 1980 paper “Feeling and Thinking: Preferences Need No Inferences,” which argued against the then-dominant view that emotion is downstream of cognitive appraisal Zajonc, 1980.arrow_upward
This is an established, if still debated, finding rather than a rhetorical flourish. Nisbett and Wilson’s classic 1977 studies found that people routinely report confident reasons for judgments that could not have actually driven them, with no awareness that they’re confabulating Nisbett & Wilson, 1977. Jonathan Haidt’s social intuitionist model builds on this directly, arguing that moral judgment in particular is typically an intuition first, with reasoning arriving afterward mainly to construct a justification for a conclusion already reached Haidt, 2001. The open question in the field isn’t whether this happens, but how often reasoning is purely post hoc versus genuinely capable of revising the initial intuition.arrow_upward
Francis Galton’s 1907 note in Nature, “Vox Populi,” describes a county-fair contest to guess the weight of an ox: the median of 787 guesses came within 1% of the true weight, even though almost no individual guess did Galton, 1907.arrow_upward

This is an established, if still debated, finding rather than a rhetorical flourish. Nisbett and Wilson’s classic 1977 studies found that people routinely report confident reasons for judgments that could not have actually driven them, with no awareness that they’re confabulating Nisbett & Wilson, 1977. Jonathan Haidt’s social intuitionist model builds on this directly, arguing that moral judgment in particular is typically an intuition first, with reasoning arriving afterward mainly to construct a justification for a conclusion already reached Haidt, 2001. The open question in the field isn’t whether this happens, but how often reasoning is purely post hoc versus genuinely capable of revising the initial intuition.