Thin-slicing
Judgements of a person formed from a few seconds of their behaviour predict certain outcomes about as well as judgements formed from several minutes.
Judgements of a person formed from a few seconds of their behaviour predict certain outcomes about as well as judgements formed from several minutes.
The claim is about exposure length, not about intuition being wise. Ambady and Rosenthal asked whether observers who see under five minutes of somebody's expressive behaviour can predict an outcome measured elsewhere, and whether seeing more of it helps. Their answer was that a short look does about as well as a longer one. Two things decide what that is worth.
The first is the criterion: predicting whether a person is lying is a different task from predicting how their students will rate them at the end of term. The second is who is doing the predicting. Almost every headline correlation in this literature comes from averaging many observers into one score, and then correlating that average with an outcome across a handful of targets. It describes what a crowd of strangers knows collectively, not what one stranger knows.
Repeats disagree, or the effect holds in some conditions and not others. The other 248.
Mixed Short exposures keep predicting things above chance, and the individual results range from clean replications to a documented miss, with the causal reading of the famous election finding contested.
The formal direct replication in this family is the Reproducibility Project: Psychology attempt on Rule and Ambady's 2008 finding that ratings of chief executives' faces track their companies' profits. Crossref, OpenAlex and Europe PMC turned up no other on 21 September 2026. Its registered criterion was the partial correlation between rated leadership and profits. The original gave r(41) = .30 from a starting set of 50 chief executives; the replication gave r(41) = .27, p = .086.
That misses the criterion while sitting almost on top of the original estimate. The replication report is more careful than the label. The paper's other main correlation, between rated power and profits, came back at r(41) = .37, p = .01, against .36 in the original. The report says its own criterion was arbitrary, because the authors held up neither correlation as the important one, and it warns readers off the replicated-or-not reading in this case.
The project's own note explains why nobody can fix this by running more participants: the targets are the unit of analysis, so power is set by the faces, not by the number of undergraduates rating them. Elsewhere the record is genuinely split. Ballew and Todorov reproduced the election result for governors. Pre-election competence judgements called 68.6% of 35 gubernatorial races and 72.4% of Senate races in 2006.
Telling people to deliberate cut accuracy to 60.0%, which was not different from chance. A 100-millisecond exposure predicted 59.6% of 89 gubernatorial races, significant only at the .10 level. Atkinson, Enos and Hill then attacked the causal reading. Scoring the faces of 972 congressional candidates, they found that parties field better-looking challengers where they think they can win.
Moving a race from safe to tossup on the Cook scale raised predicted challenger facial competence by about 0.9 of a point in the House and 0.8 in the Senate, close to a full standard deviation. Controlling for that, the direct effect of a face on one vote is 1 to 3 points for partisans and about 4 for independents.
Across all 99 Senate elections they studied, the challenger's face was never the difference between winning and losing. Cox, Devine, Bischmann and Hyde ran five experiments on inferring sexual orientation and concluded that orientation is not visible in the face, and that apparent face-based accuracy comes from a third-variable confound. Furley, Wunderlich, Grosshaus and Schweizer tested the wisdom of crowds on track and field. Averaged judgements of athletic performance correlated .44 to .94 with the real results, while the mean single judge managed .18 to .64.
The lecturer judged in the first minuteillustration
Twenty students watch a minute of a new lecturer with the sound off and rate how warm and organised she seems. Their pooled rating tracks the end-of-term evaluations about as closely as a pooled rating taken from five minutes of the same recording.
The panel's opening impressionillustration
Four interviewers each form a view during the first exchange and write it down before the questions begin. The average of those four notes predicts the panel's final ranking about as well as the average taken after an hour, though no single interviewer's note does.
The shortened auditionillustration
A choir hears eight bars from each singer rather than a whole piece, and the order a panel produces from the short excerpts matches the order it produces from the full pieces. The extra minutes buy confidence in the ranking rather than a different one.
Cases marked as illustrations describe nobody in particular. They are there to make the pattern recognisable, not to report that a particular thing happened.
The 1992 meta-analysis reports an unweighted mean r of .39 and a weighted mean of .41, with a 95% confidence interval of .34 to .48. The distribution is wide: the median result is .30, the lower quartile .22, the lowest .10 and the highest .87. Effect sizes for six exposure bands ran from .24 to .45, and a linear contrast on exposure length was not significant, which is the basis for the claim that longer looks add nothing.
The band for clips under 30 seconds contained a single result, r = .33. Median sample size across studies was 32 targets, with a range of 2 to 271, and a median of 37 judges each. The best-known demonstration is Ambady and Rosenthal's 1993 teacher study. Nine female judges watched three silent 10-second clips of each of 13 college teaching fellows.
Their averaged ratings on 15 dimensions correlated .76 with the teachers' end-of-semester student evaluations, on 11 degrees of freedom. A second sample of 13 high-school teachers gave .68 against a principal's ratings. Rated physical attractiveness correlated .32 with the college criterion and -.18 with the school criterion, neither significant. Cutting the clips to 5 and 2 seconds gave combined correlations of .46 and .54, against .72 at 10 seconds.
The authors note an unexplained reversal in which 2-second clips beat 5-second ones for the college sample. The paper prints a combined effect size of .59 across all six sub-studies. On faces rather than behaviour, Todorov, Mandisodza, Goren and Hall had participants pick the more competent-looking of two candidates. The one picked won 71.6% of 95 US Senate races from 2000 to 2004 and 66.8% of 600 House races from 2002 and 2004.
Competence ratings correlated .44 with the difference in vote share across the Senate races. A later study in the same paper, run on the 2002 Senate races alone, separated the traits. There the competence factor correlated .58 with the vote difference, trust -.09 and likability -.17, and neither of the last two was significant. A 1-second exposure predicted 67.6% of Senate races.
6 primary research4 replication and reanalysis3 commentary and critique
Their paper "Thin slices of expressive behavior as predictors of interpersonal consequences: A meta-analysis" gathered 38 independent results from 44 studies, none resting on more than 300 seconds of observation. Ratings had to be tied to a clearly defined external behavioural criterion or to expert ratings; self-report and peer ratings were excluded as criteria.
The claim, and the replication this entry reads its figures from.
The target samples are small and the entry's correlations should be read with that in mind. Both teacher studies rest on 13 teachers, so a correlation of .76 carries 11 degrees of freedom. The Reproducibility Project figures start from 50 faces, and both the original and the replication partial correlations carry 41 degrees of freedom. No confidence intervals are printed for the 1993 correlations or the replication because neither source reports them, and computing them here would produce a number no source contains.
The criterion also decides what the accuracy means. Ambady and Rosenthal's inclusion rule accepted objective behavioural outcomes and expert ratings, and their highest effect size by type of outcome was social psychology at .47, ahead of clinical at .41 and deception detection at .31. Predicting a student evaluation is predicting another set of impressions, which is a weaker claim than predicting how well somebody teaches.
One source here is named rather than quoted. Murphy and colleagues' five studies on one-minute slices are cited for their reported conclusion of some loss in predictive validity, and of little loss once the first two slices are combined, for five of their six behaviours. Their coefficients were not obtained. The only open record, at EPFL's Infoscience, is marked metadata-only and holds no file, although Unpaywall still lists it as an open submitted version.
The Reproducibility Project's own spreadsheet lists Rule and Ambady (2008) under the wrong journal, volume and pages, and gives it the subject headings of a different paper by the same authors; the citation here is taken from Crossref instead.
Each of these is named in this entry's own prose as something it gets taken for. Where the two came out differently, that is worth knowing before you quote either.
The pooled r of .39 is read as showing that snap judgements of people are accurate. It shows that averages of many raters' snap judgements correlate with a criterion across a small set of targets. The median study in that same meta-analysis returns .30, with a lower quartile of .22. The finding that exposure length does not matter is also stretched past its evidence: the band for clips under half a minute held one result.
Willis and Todorov's 100-millisecond study is routinely cited as proof that impressions formed in a tenth of a second are correct. It measured agreement, correlating judgements made at 100, 500 and 1,000 milliseconds against judgements made with no time limit, and reports no accuracy criterion at all. Entry no. 506 sets out that experiment in full. The election work is the most over-read.
Todorov and colleagues wrote in the paper itself that there is no good evidence that trait inferences from facial appearance are accurate. Their design is correlational, and candidate selection is the gap Atkinson, Enos and Hill filled. Popular retellings, above all Malcolm Gladwell's, treat this literature as a licence to trust a first impression. The studies measure what a pooled crowd of raters predicts, and they do not license any single reader's hunch. Ballew and Todorov's deliberation condition is not evidence that thinking makes people worse judges of people in general.
Every claim on this page was held against these sources on 2026-09-21. Nothing here is written from memory. Written and checked by Krishna Chagti; how entries are written and corrected. Found a mistake? Report it, with the sentence and a source.