Evidence status: conditional support People form trustworthiness impressions from faces very quickly, and judges often show some agreement. A 2022 meta-analysis found modest statistical associations between appearance-based impressions and several criteria of “actual” trustworthiness, but its authors concluded that the accuracy was unlikely to be practically useful. Agreement, small correlations and reliable person judgment must remain separate. This is a version 0.1 research draft, with evidence reviewed to August 2026.
How Everyday Psychology Takes Shape · Season Two: How Do We Read Other People? · Article 3
Place two identification photographs side by side. One face may immediately seem warm and dependable; the other cold and suspicious. The judgment arrives even when we know that a photograph contains no behavioural history. It feels less like an inference than like seeing a colour.
That speed gives facial impressions an unusual authority. We say, “This person looks trustworthy,” and quietly drop the word looks. A fact about the perceiver’s response becomes a moral claim about the person being perceived.
Research confirms that humans make rapid facial evaluations with partially shared structure. It does not justify turning those evaluations into dependable measures for recruitment, lending, criminal justice, partner selection or security screening. The gap between those claims is the subject of this article.
What are we seeing in a face?
Research by Alexander Todorov and colleagues shows that people can form trait impressions after extremely brief exposure. Additional viewing often increases confidence and stabilises a judgment rather than creating an entirely new dimensional structure. In 2008, Nikolaas Oosterhof and Todorov proposed that many social evaluations of faces could be organised along two broad dimensions. One resembled a positive–negative or approach–avoidance evaluation, commonly labelled trustworthiness; another concerned dominance.
A “trustworthy-looking” face partly carries traces of emotional expression. Features resembling slight happiness tend to increase positive evaluation, while structures resembling anger tend to reduce it. Even when researchers use nominally neutral faces, the visual system can interpret relatively stable facial shape as if it were a temporary expression.
This account helps explain agreement among observers without establishing the character of the target. Many perceivers can use the same cue and converge on the same error. In measurement terms, agreement speaks to one aspect of reliability. Validity asks whether the rating corresponds to the property it claims to measure.
Recent work also emphasises substantial idiosyncrasy. Aggregate models can show which facial configurations produce average ratings while concealing variation among perceivers. People bring different learning histories, cultural associations and experiences with particular faces. “Consensus” is never the whole judgment.
What does modest accuracy mean?
In 2022, Yap Zhi Foo and colleagues published a meta-analysis of accuracy in facial trustworthiness impressions. They separated two questions. Across faces, were targets who looked more trustworthy also rated or measured as more trustworthy by external criteria? Across perceivers, were some observers better than others at distinguishing targets? The pooled association was approximately r = .14 at the face level and r = .27 at the perceiver level. Both were statistically above zero. The authors explicitly concluded that this modest accuracy was unlikely to have practical utility.
Why is “better than zero” insufficient? Real decisions classify individuals. A modest correlation permits extensive overlap and error. A trustworthy-looking person may be dishonest; a stern-looking person may be highly dependable. Base rates, noisy criteria, selective reporting and the particular domain of judgment further alter predictive performance.
The criterion called “actual trustworthiness” is not a single natural substance. Studies may use returned money in a trust game, informant reports, honesty tasks, aggression or records of sexual unfaithfulness. These outcomes are not morally or behaviourally interchangeable. Combining them into the claim that faces reveal good and bad character erases what was measured.
The meta-analysis also identified a predominantly Western literature, limited exchange between research clusters and problems in reporting clarity. Small effects require especially serious external validation because they are easy to select and exaggerate in application.
How a facial impression can manufacture its own evidence
First impressions change not only evaluation but conduct. If someone looks trustworthy, we may be warmer, disclose more information and offer more opportunities to cooperate. If the person looks suspicious, we remain distant and ask sharper questions. The other person responds to that treatment, and the original impression can seem confirmed.
This is a possible feedback process, not a claim that every outcome is self-fulfilling. A person with deceptive intent will also produce relevant behaviour, and caution can genuinely protect us. The important point is that later interaction is no longer independent of the first judgment. Without a record of the sequence, “he became defensive in response to my suspicion” can be remembered as “I saw his nature immediately”.
Institutional use magnifies the problem. Human recruiters already need safeguards against appearance bias. If a machine-learning system is trained on thousands of human ratings of who looks trustworthy, what it primarily learns is an automated version of a social impression. Reproducing rater consensus accurately does not show that the system discovered honesty in the target. Using such a score for consequential decisions would resemble physiognomy in technical form.
A legitimate place for first impressions
A first impression can function as an internal alert, provided it is labelled as a hypothesis. If a situation feels unsafe, a person need not prove malicious intent before leaving. Personal safety permits low-cost precaution. Protecting oneself, however, is different from declaring another person untrustworthy in a way that affects opportunity or reputation. The first manages risk; the second makes a claim about someone.
Where cooperation is required, stronger evidence comes from behaviour. Are commitments honoured? Can information be checked? How does the person handle error? Are conflicts of interest disclosed? Important transactions should also rely on contracts, separation of authority, audit and reversible stages. Good institutions do not require miraculous judges of character. They allow trust to be built gradually and breach to be detected and repaired.
A useful mental sentence is: “This face produces distrust in me, but I do not yet know whether this person is untrustworthy.” The first clause respects a real response; the second stops the response from impersonating a measurement. We can then ask whether expression, resemblance to a past person, a group stereotype or observed conduct is driving the judgment.
This distinction matters even when the impression later turns out to be right. A correct guess does not retroactively make the method reliable. If a coin lands heads once, the outcome cannot prove an ability to predict coins.
Are we seeing a person or our evaluative system?
The philosophical significance of face research is not that vision is unreal. Faces carry genuine information about age, expression, gaze and some current states. Social life cannot perform a background check before every interaction. Rapid evaluation has functions.
The limit is different. A system that helps allocate approach and avoidance need not be qualified to judge moral character. A heuristic shaped by learning or evolution can direct attention without acquiring the legitimacy to deny employment, credit or liberty. The greater the consequence, the stronger the evidence that should be demanded.
The version 0.1 conclusion is clear. Faces strongly influence whom we experience as trustworthy. Across selected research criteria, impressions show small average associations with outcomes, but those associations are far too weak to judge an individual. A first impression can prompt the collection of evidence; it cannot complete the case. Above all, institutions should not train a score on the proposition “people agree that this face looks trustworthy” and then allow the score to exercise power over the person pictured.
Primary research sources
- Oosterhof, N. N., & Todorov, A. (2008), “The Functional Basis of Face Evaluation”.
- Todorov, A., Olivola, C. Y., Dotsch, R., & Mende-Siedlecki, P. (2015), “Social Attributions from Faces”.
- Foo, Y. Z., Sutherland, C. A. M., Burton, N. S., Nakagawa, S., & Rhodes, G. (2022), “Accuracy in Facial Trustworthiness Impressions”.
- Todorov, A., Face Value: The Irresistible Influence of First Impressions.
Series navigation: How Everyday Psychology Takes Shape — Season Two overview
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.