Appearance
10.1 — How You Already Read People, and How Often You Are Wrong
Show someone a photograph of a stranger's face for one tenth of a second and ask how trustworthy the person looks.
They can answer. And their answer agrees closely with the answer they give when allowed to look for as long as they want, and with the answers other people give about the same face.
Alexander Todorov's research established this repeatedly. Judgements of trustworthiness, competence, dominance and likeability from faces are formed in well under a second, they are highly consistent between people, and they are held with confidence.
And they have little relationship to what the person is actually like.
That combination — fast, confident, consistent, and largely unrelated to the truth — is the central fact of this whole Part. It is not a reason to stop reading people. It is a reason to know what your reading is actually made of.
What you are actually doing
Face reading is not one process. It is at least four, running simultaneously and reporting as a single impression.
Emotion detection. Reading someone's current state from their face, voice and posture. This works reasonably well, particularly for people you know, and it is the part of your ability that is genuinely good.
Trait inference. Concluding what kind of person they are from their appearance. This works badly, and it is the part that feels most like perception.
Situation reading. Working out what is going on for them from context and behaviour over time. This can be excellent, and it is a skill rather than an instinct (10.9).
And your own projection. Attributing to them what you would be feeling. Unreliable, and invisible from the inside.
Most people run all four and cannot tell which one produced the conclusion. Separating them is the single most useful move in this Part, because the first and third deserve confidence and the second and fourth do not.
Why trait inference from faces fails
The mechanism is well understood and it is worth knowing, because knowing it interrupts the process.
Faces resembling an emotional expression get read as having that trait. A face whose neutral resting state has slightly downturned corners is read as unfriendly. A face with raised inner eyebrows is read as anxious. This is emotion detection misfiring on structure rather than expression — the person is not feeling anything; their face is simply shaped that way.
Faces resembling someone you know inherit that person's associations, entirely outside awareness.
Attractiveness leaks into everything. The halo effect: attractive people are rated as more intelligent, more competent, kinder and more trustworthy, in study after study. The effect is large, it operates without your consent, and the underlying correlations with actual ability are negligible.
And babyish features — large eyes, round face, small chin — produce judgements of honesty, warmth and naivety. This has real consequences: research on legal outcomes finds associations between facial maturity and how defendants are treated.
None of these have any predictive value about the person. They are structural features being read as psychological ones.
The confidence problem
Here is the part that makes it dangerous rather than merely inaccurate.
Confidence in reading people is largely uncorrelated with accuracy. This shows up in the lie-detection literature most starkly (10.6), where people who are certain do no better than people who are unsure, and where professionals with decades of experience perform close to chance while being far more confident than novices.
Why experience does not fix it. From 2.5, accurate intuition requires many repetitions with fast, clear feedback. Reading people supplies almost no feedback. You decide someone is untrustworthy, you avoid them, and you never find out. You decide someone is competent, hire them, and if they fail you attribute it to circumstances.
In the absence of feedback, what accumulates is not skill. It is confidence, because each unchallenged judgement feels like a hit.
This is worth sitting with if you consider yourself a good judge of character. Almost everybody does, the trait cannot be uniformly distributed that way, and the belief itself is not evidence.
What you are actually good at
The picture is not all negative, and being accurate about your own accuracy means naming the parts that work.
Reading current emotional state, in people you know well. Genuinely good. You can detect that your partner is upset from something you could not articulate, and you are usually right.
Detecting that something is off, without knowing what. The general "something is wrong here" signal has decent hit rates, and it is worth treating as a prompt to look rather than as a conclusion.
Reading behaviour over time. Patterns across many interactions carry real signal (10.11), unlike single impressions.
And reading situations, as distinct from people. What is happening in this room, who has power here, what everybody wants — this is frequently accurate and it is a learnable skill.
The two errors that do the most damage
Attributing behaviour to character rather than situation (5.8). The fundamental attribution error. He is unreliable, rather than he has a sick parent at home.
And stopping at the first read. The impression forms in a second and then everything afterwards is fitted to it (1.4). The first impression does not merely come first — it filters everything subsequent, which is why it takes disproportionate evidence to overturn.
What this Part will do
10.2 to 10.4 go through the channels — face, body, voice and words — and say what each one actually carries, which is different from what popular books claim.
10.5 draws the distinction that fixes most errors: reading a state against reading a character.
10.6 covers lie detection, where the gap between confidence and accuracy is widest.
10.7 and 10.8 cover first impressions and the question of when to judge somebody at all.
10.9 gives the disciplined method that replaces the hunch — working out someone's situation from evidence.
10.10 covers how people are fooled by those who do this deliberately.
And 10.11 covers the thing that actually works: building an accurate model of one specific person over time.
What to do with this
Separate the four processes when you form an impression. Am I reading their state, inferring their character, reading the situation, or projecting? The first and third deserve weight; the second and fourth do not.
Notice the face-structure effect specifically. When someone strikes you as cold, unfriendly or shifty on sight, the honest first hypothesis is that their face is shaped that way, not that you have detected something.
Downgrade your confidence, deliberately, once. Think of a person you decided about quickly and turned out to be wrong about. Everybody has one. Hold it as the calibration point rather than as an exception.
And start collecting feedback, which is the only thing that converts experience into skill (3.5). Write down a prediction about someone — how they will handle a deadline, whether they will follow through, how they will react — and check it later. Twenty of these will tell you more about your own accuracy than twenty years of unchecked impressions.
Next: 10.2 takes the channel you rely on most — the face — and gives the honest account of what it carries, including the argument currently running in the research about whether expressions mean the same thing everywhere.