First Impression Research: What 50 Years of Science Actually Says

The 100-Millisecond Study (Willis & Todorov, 2006)

Willis and Todorov's experiment was simple in design and definitive in result. Participants viewed faces under three conditions: 100ms exposure, 500ms, or unlimited time. After each exposure, raters scored the face on five traits: attractiveness, likability, trustworthiness, competence, and aggressiveness.

The judgments formed under 100ms exposure correlated with judgments formed without time pressure at r=0.5 to r=0.7 across traits. Trustworthiness showed the highest correlation (about 0.73). Increasing exposure from 100ms to 500ms to unlimited mostly increased rater confidence without changing the content of the judgment. The rapid read was the read.

Subsequent replications across diverse populations (American, European, East Asian) confirmed the timing finding. The trait-content of the rapid judgment varies across cultures somewhat — for instance, the visual cues weighting trustworthiness ratings shift with cultural context — but the timing remains stable. Wherever it has been tested, the snap judgment forms in roughly 100ms.

Ekman 1971: Why Facial Emotion Reads Cross Cultures

Before Willis & Todorov could establish how fast face judgments happen, Paul Ekman had to establish that the inputs were universal. His 1971 work on cross-cultural emotion recognition is one of the foundational findings of modern face research. Ekman tested whether facial expressions of basic emotions — happiness, sadness, anger, fear, surprise, disgust — were recognized consistently across cultures.

His critical study was with the Fore people of Papua New Guinea, an isolated population in the 1960s with essentially zero exposure to Western media. Ekman showed them photographs of Western faces displaying the six basic emotions and asked them to match the expression to a brief story (e.g., "his friend has died" for sadness). Match rates were consistently above chance and aligned with Western rater patterns. The same expressions read the same way to populations with no shared visual culture.

This established that facial emotion recognition has a biological substrate rather than being purely learned. The implication for first-impression research: when an observer reads a face within 100ms, they're using a perceptual system that is largely shared across humans. Different cultures will weight features somewhat differently, but the underlying "read this expression" machinery is common hardware.

Ekman's later FACS (Facial Action Coding System) provided the granular framework for analyzing which muscle activations drive which perceived expressions. This is the toolkit that downstream face-perception researchers (Todorov included) built on to identify which specific facial geometric cues drive specific trait inferences.

Thin-Slice Studies: 30 Seconds Predicts Outcomes

Ambady and Rosenthal's 1993 paper "Half a minute" extended the idea that rapid judgments are accurate into ecological domains. They showed undergraduates 30-second silent video clips of college instructors teaching, then asked them to rate the instructors on traits like warmth and competence. The ratings predicted end-of-semester student evaluations of the same instructors with surprising accuracy.

Thirty seconds of silent video, generated from raters who never met the instructors, predicted a semester of student evaluations from raters who spent months in their classes. The accuracy didn't drop even when video was cut to 6 seconds. The thin-slice judgments worked because the same cues an observer uses to form a 100ms face read also drive longer-term impressions of teaching effectiveness.

This generalizes. Rule and Ambady showed that brief exposures to CEO faces predicted real company financial outcomes — observers rating CEOs as "powerful" or "competent" based on a face alone correlated meaningfully with actual firm performance. Antonakis and Dalgas 2009 ran the famous Swiss-kids-predicting-French-elections study, where children aged 5-13 viewed pairs of competing French parliamentary candidates and chose "who would be a better captain" — their picks matched actual election winners 71 percent of the time.

These studies cut two ways. They show first-impression judgments are real and consequential — observers in high-stakes domains are using exactly the same rapid-read machinery as the lab participants. They also show that observer bias is built into outcomes. Whether the elected official is actually competent or just face-competent, voters use the same input. This bias is the thing first-impression optimization tries to work with.

What Facial Features Drive Which Trait Judgments

Todorov's lab after the 2006 paper spent years mapping which specific geometric features drive which trait inferences. Their reverse-correlation studies use computer-generated faces to isolate the geometric variation that maximally moves each trait dimension. The findings:

• Trustworthiness is driven primarily by the eye region (upturned mouth corners, slight brow lift, wider eye openings). High-trust faces look like they're mid-smile even at neutral.
• Dominance is driven by lower-face structure — squared jaw, lower brow ridge, wider facial width-to-height ratio. Carre & McCormick 2008 work on FWHR ties into this stream.
• Competence blends both — eye/mouth cues from trust, plus jaw/brow cues from dominance. High-competence faces typically score moderately on both dimensions.
• Attractiveness draws on bilateral symmetry, mildly positive canthal tilt, sexually-dimorphic features, skin quality, and averageness in different weights depending on the rater pool and sex of the rated face.

Different photo contexts privilege different traits. LinkedIn headshots weight trust and competence over attractiveness. Dating apps weight attractiveness and some warmth signals. Founder portraits weight competence and dominance. Optimizing the wrong dimension for the context is one of the most common photo-strategy mistakes.

What the Research Doesn't Say

A few clarifications, because online discussion of this research routinely oversells the findings:

The 100ms judgment isn't accurate in the sense of correctly inferring the person's actual traits. It's consistent — different observers rate the same face similarly and that rating is stable over time. Whether the rating maps to truth depends on the trait. Trustworthiness ratings from a face do not strongly predict whether the person is actually trustworthy. They predict whether other observers will trust them, which then drives outcomes through social channels independent of actual trustworthiness.

Effect sizes vary across studies. The Hönekopp 2006 symmetry meta-analysis reported r=0.15-0.25 for symmetry-attractiveness. The Willis & Todorov 100ms correlations were r=0.5-0.7 with unlimited-time ratings of the same face — strong by social-science standards. The Ambady thin-slice predictions of teaching ratings were also strong (r~0.5+ in original samples). Recent replication-crisis-era studies have generally come in with smaller effect sizes than originals, so treat headline numbers from individual papers with skepticism and rely on convergent patterns across well-replicated work.

Photo optimization for first-impression dimensions doesn't change who you are — it changes the rapid-read input observers get. For high-stakes contexts (dating profile, professional headshot, founder portrait, election photo), this is the leverage point. The judgment is going to form in 100ms either way. The question is what those 100ms get to work with.