How Fast We Judge a Face

You meet someone, and before either of you has said a word, some part of your brain has already decided whether they seem trustworthy, competent, or a little intimidating. This is not a slow, deliberate read of their character. It happens in about the time it takes light to cross a room.

The 100-millisecond experiment

In 2006, Janine Willis and Alexander Todorov showed participants photos of unfamiliar faces for exactly 100 milliseconds, 500 milliseconds, 1000 milliseconds, or as long as they wanted, and asked them to rate each face on trustworthiness, competence, likability, aggressiveness and attractiveness. The striking result was not that snap judgments happened - everyone expects that - but how little additional viewing time changed them. Ratings made after just 100 milliseconds correlated strongly with ratings the same people made with unlimited time to look. Giving people longer mostly made them more confident in the judgment they had already made in that first fraction of a second, rather than leading them to revise it.

Thin slices, outsized predictions

This is not an isolated oddity of faces. In 1992, Nalini Ambady and Robert Rosenthal pooled together a wide range of studies in which observers judged people from very brief samples of behaviour - a few seconds to a couple of minutes of video, sometimes with the sound turned off - and compared those snap judgments to outcomes measured much later, such as end-of-semester teacher evaluations or ratings of physicians' bedside manner made after full appointments. Across this body of work, brief "thin slice" judgments predicted those later outcomes about as well as much longer observation did. Whatever pattern-matching the social brain runs on a face or a few seconds of behaviour, it is packing in a surprising amount of information, or at least a surprising amount of the information that later, slower judgments end up relying on too.

None of this is really specific to faces. It fits a broader pattern in how snap social judgments seem to work: a first read is not so much a rough draft waiting to be corrected by more information as it is a conclusion that additional time mostly gets spent justifying. That is worth knowing before trusting your gut about a stranger you have just met - and an even more useful thing to keep in mind before trusting it about how a stranger has just read you.

Widely shared, and not particularly accurate

Here the story turns more cautionary. A 2015 review by Alexander Todorov, Christopher Olivola, Ron Dotsch and Peter Mende-Siedlecki, drawing together a large body of research on face-based judgment, found that people agree with each other remarkably well about which faces look trustworthy or competent - strangers rating the same photographs tend to converge on similar scores, and agreement holds up reasonably well even across different cultural backgrounds. What that same review found much less evidence for is accuracy: how trustworthy or competent a face looks tracks poorly, if at all, with how trustworthy or competent that person actually turns out to be. The consensus is real. What it is a consensus about is mostly a shared reaction to features - jawlines, brow position, resting mouth shape, the kind of face that resembles an angry or a relaxed expression - not a shared window into anyone's character.

This gap between agreement and accuracy is not a harmless curiosity. The same review notes that face-based impressions of competence and trustworthiness have been linked to outcomes that carry real weight - including which political candidates voters favour and how severely defendants are treated - not because the face reveals something true about the person, but because other people act on the impression as though it were true.

Part of why the agreement is so strong and so fast is that it does not depend on knowing anything about the person at all. Nikolaas Oosterhof and Alexander Todorov built a computer model of what makes a face look trustworthy by systematically varying facial features and asking people to rate the results, and found that trustworthy-looking faces structurally resemble a relaxed, mildly happy expression, while untrustworthy-looking ones resemble a face on the verge of anger - even on a face that is, in reality, doing neither. This works on completely unfamiliar, computer-generated faces with no name, biography or history attached, which is a strong hint that the reaction is about the geometry of a face echoing an emotional expression, not about anything the observer actually knows or has correctly inferred about the character behind it.

What this means for you

The snap judgment you form of a stranger, and the one they form of you, is close to finished before either of you has had a chance to speak - and it resists revision less because it is well-informed and more because looking longer mainly cements it. That is worth remembering on both sides of the encounter. When the stakes are real - hiring someone, judging a first date, weighing up a new colleague - it is worth deliberately building in more information before treating your gut read as settled, since the read itself was formed too fast to have used much. And if you are self-conscious about how your own face comes across to people who do not know you, it can help to remember what strangers are actually reacting to in that first instant: a rough, largely involuntary template built from your resting features, not a genuine readout of your character. It says something real about how first impressions work. It says very little about who you actually are.