Thin Slices: What a Four-Minute Judgment Actually Predicts

People form stable judgments of others from extremely brief exposure — seconds, not minutes — and those judgments correlate with longer-term ratings well above chance.¹ The finding is genuinely robust and it has been misapplied for thirty years, because the interesting question is not whether snap judgments predict something. It is what they predict.

What thin slices are good at

They are good at reading things that are visible in seconds because they are expressive: warmth, confidence, energy, whether someone is comfortable being watched. Judgments of this kind from very short clips track longer evaluations closely, which makes sense — the signal is fully present in the sample.

They are correspondingly poor at anything that only reveals itself over time: whether the person finishes things, whether they are accurate under pressure, whether they say "I don't know" when they don't know. None of that is in the sample. It cannot be read quickly because it is not there to read.

The interview measures interview performance. Interview performance is a skill, and it is not the skill you are hiring for unless the job is being interviewed.

The consequence for hiring

The book-length version of this argument is Noise, by Kahneman, Sibony and Sunstein: wherever the same judgment is made repeatedly by people, the variation between judges is larger than anyone believes, and structure is the only reliable correction. Applied to hiring, that is the entire case for asking every candidate the same questions.

This is why unstructured interviews perform so poorly as predictors of job performance while feeling so persuasive. The persuasion comes from the thin slice — you form an accurate read of how the person comes across, then generalise it to competence, where it does not transfer.²

The comparative evidence on selection methods is unusually clear for a social-science question, and the ordering has been stable across decades of meta-analysis: work samples and structured interviews substantially outperform unstructured conversation, and unstructured conversation barely outperforms nothing.³ The specific coefficients are worth checking and have been revised, but the direction has not moved.

The practical translation for a small company, where a bad hire is not a statistic but a quarter of the payroll:

1. Ask every candidate the same questions in the same order. Nearly all the benefit of "structured" is here, and it costs nothing. Comparison is impossible when every conversation went somewhere different.
2. Score immediately after each answer, not after the meeting. A single strong moment otherwise recolours the entire memory of the conversation.
3. Use a paid work sample. Two hours of the actual work, paid, beats every interview technique available. For an operational role, one shift shadowed. The sample tests the thing you are buying.
4. Interview last, not first. Once you have seen the work, the conversation is about how the person thinks and whether you can work with them — which is a question the interview is legitimately good at.

Where the snap judgment is the right instrument

Not everything about a hire is competence. If the role is client-facing, then warmth, presence and how someone lands in the first minute are part of the job — and there the thin slice is measuring exactly the right thing. The error is not trusting the fast judgment; it is trusting it about the wrong variable.

The same applies on the other side of the table. A prospective client is forming a thin-slice judgment of you in the first minutes of a meeting, and that judgment is accurate about your warmth and confidence and uninformative about your competence. Which is a reasonable argument for letting the work sample do the talking in sales, too.

The uncomfortable part

Fast judgments of expressive traits are accurate and they carry the demographic biases of the observer. Both things are true, and the second is why structure is not merely a performance improvement — it is the main defence you have against reproducing your own preferences and calling the result merit.