What Body Language Actually Predicts, and What It Does Not

The 93% figure is where this whole genre went wrong

Almost every list of body-language tells rests, somewhere in its lineage, on one number: 7% of communication is the words, 38% the tone, 55% the body. It is the most quoted statistic in the field, and it does not say what it is used to say.

It comes from two papers published in 1967, one by Albert Mehrabian with Morton Wiener and one with Susan Ferris, and the three percentages were assembled afterwards from two separate experiments. In the first, listeners heard single words spoken in positive, neutral or negative tones and judged the speaker's attitude. In the second, they weighed a voice recording against a photograph of a face. Neither study involved a conversation. Neither measured how much information a channel carries. Both asked a much narrower question: when the word and the tone disagree about a feeling, which one does the listener believe?

The answer was the tone, and that is the entire finding. Mehrabian spent decades saying so. Applied to communication in general, to a lecture or a negotiation or a first date, the figure is not an exaggeration of a real result. It is a category error, and we take it apart in more detail in the piece on texting, where the same misquote does the same damage to how people read a message on a screen.

It matters here because 7-38-55 is what licenses the promise underneath the genre: that the body is a wide, high-bandwidth channel you can learn to read off another person. The two experiments behind the number establish something far smaller. Most of what follows is what happened when researchers went and tested the larger claim directly.

The second assumption: that the body leaks what the mouth is hiding

The premise of every decoder is that people manage their words better than their bodies, so watching the body gets you behind the performance. That is a testable claim, and the place it has been tested hardest is deception.

Bond and DePaulo pooled 206 documents and 24,483 judges asked to sort statements into lies and truths. Average accuracy came out at 54%, against a 50% baseline. Judges called 47% of lies deceptive and 61% of truths non-deceptive, which is a bias toward believing people rather than an ability to see through them. Confidence barely tracked accuracy.

The reason shows up in the companion meta-analysis. DePaulo and colleagues went through 1,338 estimates of 158 candidate cues to deception drawn from 120 independent samples. A few held up weakly: liars were slightly less forthcoming, slightly tenser, told slightly less compelling stories. Many of the cues people actually watch for showed no discernible link at all. There is no equivalent of a nose that grows.

The fair counterweight, because none of this says the body is noise. Ambady and Rosenthal's meta-analysis of thin slices took 38 studies in which judgements were based on under 300 seconds of expressive behaviour and related to a real external criterion. The mean effect size was r = .39, and longer exposure did not reliably beat short exposure. Something real is being transmitted, and it arrives fast.

Put the two results side by side and the honest shape of the channel appears. Nonverbal behaviour carries a genuine signal about broad impressions, formed in seconds, that staring longer does not improve. What it does not do is grant privileged access to what a person is deliberately withholding. Those are different claims, and this genre runs almost entirely on the second one.

What survived the largest synthesis, and what did not

The direct question, which behaviours actually track attraction, has been answered at scale. Montoya, Kershaw and Prosser pooled 309 effect sizes from 54 studies and 5,422 participants. The overall association between enacted behaviour and reported attraction was z = .20.

Eleven behaviours reached significance. On the indirect side: gaze duration and frequency, smiling, laughter, mimicry, head nodding. On the direct side: physical proximity, the amount of talking, the duration of talking. Two more were specific to speed dating. Direct behaviours related to attraction more strongly than indirect ones, which is worth holding on to, because this genre is almost entirely about the indirect ones.

The half that never gets quoted is the null list. In the University of Dayton's summary of the paper, Montoya names the behaviours that showed no evidence of a relation to liking: hair flips, eyebrow raises, gestures, head tilts, primping, open body posture, and leaning in.

Three of those were presented as tells on this page until today.

And there is the interpretation the author gives of his own result, which is the most useful sentence in the area. The eleven behaviours that survive are trust-and-rapport behaviours: when we like someone, we act in ways to get them to trust us, and the same set turns up between doctors and patients, parents and children, salespeople and customers. They are evidence of liking. Nothing in them separates the kind of liking, which is the whole problem laid out in what the evidence supports about reading someone's interest and the reason flirting and ordinary friendliness are so hard to tell apart.

Foot direction: this page got it wrong, and here is the correction

Until today, the first cue on this page was where a person's feet point. The claim was that feet escape conscious management, so a foot vector aimed at you across an evening is the body voting honestly. It is one of the most confidently repeated claims in the genre, and we repeated it.

It is not in the literature.

Montoya, Kershaw and Prosser coded every behaviour that had been measured against attraction across 54 studies. Foot direction is not among the eleven that worked, and it is not among the seven that failed, because it is not among the behaviours anyone had tested. Hall and Xing coded 36 verbal and nonverbal behaviours in recorded first conversations. Foot orientation is not there either. The claim traces to popular body-language books and to the sites that copy them, and it has been passed along for long enough that it now reads as established.

Absence of evidence is not proof the cue is empty. Feet plausibly are less monitored than faces. But an untested guess and a meta-analytic result are not the same kind of thing, and this page presented the guess as the stronger of the two. That is precisely what our editorial policy commits us not to do, so the section is gone rather than quietly softened.

The general lesson is worth more than the specific retraction. When a cue is stated with unusual confidence and no citation, the confidence is often doing the work the evidence cannot. That test would have caught this one, and it will catch most of what you read on this subject elsewhere.

The pupil and the eyebrow flash were never attraction cues

Two of the most-repeated involuntary tells come from real research that found something other than what it is cited for.

Eckhard Hess and James Polt published pupil measurements in Science in 1960 and reported that pupil size tracked the interest value of the pictures people were shown. Real finding, real effect. The trouble is what interest turned out to mean. Bradley, Miccoli, Escrig and Lang measured pupil size while people viewed emotional images and found it indexed emotional arousal and sympathetic activation: pupils dilated for unpleasant pictures as well as pleasant ones. The pupil is an arousal and effort gauge, not a valence gauge, and certainly not a romance gauge. It also responds far more strongly to ambient light than to anything psychological, which makes the classic observation in a dim bar close to uninterpretable.

The eyebrow flash has a cleaner story and an equally unhelpful conclusion. Grammer, Schiefenhövel, Schleidt, Lorenz and Eibl-Eibesfeldt coded 255 brow raises filmed in unstaged interactions across three cultures: the Eipo in New Guinea, the Yanomami in Venezuela, and Germans in Austria. The movement has a strikingly constant form, an onset of roughly 100 milliseconds, and it is longest during openings, most often accompanied by a smile. That is a greeting signal. It says the other person has registered you and is opening an interaction. Montoya's synthesis puts eyebrow raises squarely in the no-evidence column for liking.

The older version of this page attributed the eyebrow flash to Hess, who did not study it, and read pupil dilation as attraction, which it does not measure. Both claims are corrected here, and the pupil is covered from the other direction in the piece on reading interest.

Mirroring, and which way the arrow points

Mirroring is the single most repeated cue in the genre, and it has been tested against interest more directly than any other.

Grammer, Kruck and Magnusson filmed opposite-sex encounters and state in their own abstract that they were unable to demonstrate a relation between synchronisation defined as movement echo or position mirroring and the subjective experience of pleasure and interest. What they did find was a hierarchically patterned synchronisation, pair-specific and constant in its time structure, detectable only with a purpose-built search algorithm. That is a real phenomenon. It is not something you can read off a person across a table, and it was not related to how interested either of them felt.

Montoya's meta-analysis does list mimicry among the behaviours associated with attraction across studies, so there is something there statistically. Both results can be true: a small average association across thousands of participants is not a cue with diagnostic value for one evening.

The more interesting problem is the direction of the arrow. Chartrand and Bargh's chameleon-effect experiments had confederates deliberately mirror participants' posture and mannerisms, and participants who were mirrored liked the confederate more and found the interaction smoother, without noticing anything had happened. Hove and Risen found the same for imposed synchrony: manipulated coordination increased affiliation. In both cases the mimicry came first and the liking followed.

So when you notice someone matching your posture, the honest reading is that you may be watching rapport being built rather than a preference being revealed, and that whatever warmth you feel a minute later is partly a product of the thing you thought you were measuring. This section replaces a correction added on 31 August, when this page still said the opposite of what Grammer found.

Eye contact does something to you while you are reading it

Gaze is the one cue with an unambiguous place on Montoya's positive list: duration and frequency of eye contact both track reported attraction. It is also the cue where the causal problem is sharpest.

Kellerman, Lewis and Laird ran two studies in which strangers were asked to hold unbroken mutual gaze for two minutes. In the first, 96 participants were assigned to gaze at the other person's eyes or hands in various combinations, or to count the other person's blinks. Those gazing into eyes that gazed back reported markedly higher affection than anyone else. In the second, with 72 participants, the mutual-gaze pairs showed significant increases in passionate love, dispositional love and liking for a person they had just met.

Two minutes. Strangers. No conversation.

That result does not make sustained eye contact meaningless as a signal. It makes it a signal that is partly manufacturing what you are using it to detect. If you hold someone's gaze across a table for a while and then consult your own feelings for confirmation, some of what you find was produced in the last ninety seconds.

This is why we keep saying the cluster over weeks beats the moment. A pattern that survives several separate occasions, in different rooms and different moods, is carrying information a single charged evening cannot, because the single evening is partly an experiment you are running on yourself. It also matters what you brought into the room: anxious attachment inflates ambiguous signals and avoidant attachment discounts them, which is the difference our attachment style test scores in about two minutes, and which love and anxious attachment takes apart in more detail.

The best-measured channel is the one almost nobody writes about

If you want the single strongest piece of evidence that people signal attraction nonverbally without meaning to, it is not in any of the visual cues. It is in the voice.

Leongómez and colleagues recorded 110 speakers, English and Czech, responding out loud to twelve silent video clips of potential partners and competitors, previously rated for attractiveness by an independent panel. That produced 1,304 free-speech recordings. Men varied their pitch most when responding to attractive women, and reached a lower minimum pitch with them. Women's pitch variability was more sensitive to attractive same-sex competitors. The pattern was strikingly similar across two languages from different branches of Indo-European.

Then the part that makes it matter. Naive listeners from the other language group, who could not understand a word, were played pairs of recordings by the same speaker and asked which was more attractive. They chose the speech aimed at attractive targets. The preference survived low-pass filtering at 400 Hz, which strips the recording of intelligibility and leaves pitch. And it was specific: when the same listeners rated the same recordings for friendliness instead, the effect did not hold.

That is unusual in this literature. The modulation is unconscious, it is detectable by a complete stranger with no access to the content, and it separates a mating context from general warmth, which is exactly the separation the visual cues fail to make.

The caveats are real. It is a hypothetical dating scenario with students in two European languages, and the authors call for confirmation elsewhere. But the honest ranking of channels puts a change in someone's voice above almost everything the genre tells you to watch for. The older piece on subtle signs is the one most in need of that correction.

Posture: one real experiment, about photographs, in a shaky field

Open posture and leaning in are both on Montoya's no-evidence list. There is one prominent result that appears to cut the other way, and it is worth reading precisely.

Vacharkulksemsuk and colleagues coded postural expansiveness on a seven-point scale in 144 speed dates among 24 people. Each one-unit increase in expansiveness was associated with 76% higher odds of a yes from the partner. The authors flag the obvious problem themselves: they could not tell whether people expanded because they were being liked or were liked because they expanded. So they ran a second study, creating twelve dating-app profiles from six confederates, each photographed once expansively and once contractively, and collected around 3,000 responses. Expansive photographs were 27% more likely to draw a yes.

Read that carefully and it is a finding about photographs. The causal half of the study manipulated a still image, not live behaviour, and the effect it produced was modest. The live half rests on 24 people and cannot establish direction. It is a reasonable argument for how you sit in a profile picture, which we also touch on in the first-date piece. It is not a decoder for the person across from you.

The surrounding literature deserves a warning label. Carney, Cuddy and Yap's power-posing study was one of the most publicised findings in psychology; Ranehill and colleagues ran a larger replication and found no hormonal or risk-taking effects. Körner and Schütz, reviewing the whole area of power poses and body postures, describe a field where self-directed effects have largely failed to hold up while effects on observers have fared better. Expansive posture may well change how you are seen. What it does not do is tell you what someone else feels.

There is no single signature. There are at least five.

The deepest problem with a list of tells is not that individual items are wrong. It is that the list assumes attraction has one behavioural signature.

Hall and Xing put 51 pairs of opposite-sex strangers into 10 to 12 minute conversations, had four independent coders score 36 verbal and nonverbal behaviours, and then asked each person how physically attracted they had been. The behavioural correlates of attraction differed systematically by the person's flirting style. Physical flirts showed more conversational fluency. Traditional female flirts showed more demure behaviour while traditional male flirts were more assertive and open. Sincere flirts fidgeted, teased and got distracted less, and smiled more. Playful flirts did the obviously flirtatious things. And polite flirts, who are genuinely attracted, showed more reserved and distancing behaviour.

Read the decoder against that last group and it returns a confident, wrong answer. A polite flirt is doing the opposite of the checklist while feeling the thing the checklist claims to detect.

This is where the evidence actually leaves us, and it is smaller and more useful than a decoder. There is a real channel that carries broad impressions quickly. There is a short list of behaviours that reliably accompany liking of every kind, and a longer list of famous cues that accompany nothing. There is at least one channel, the voice, that does better than the rest. And there is no measured behaviour that distinguishes attraction from warmth, no reliable way to detect what a person is deliberately not showing you, and no signature that holds across different people.

Which leaves observation doing what observation can do, and a question doing the rest. Ambiguity is not an answer and it is not an invitation; the only thing that resolves it is asking once, plainly, in a way that makes no easy to say. More in this vein under body language, and the same instrument turned on a case where it fails hardest in why people keep ending up in the friend zone, where your own attraction predicts your reading of theirs about three times better than their actual feelings do. The same discipline applied to a much more consequential case is in the psychology of jealousy, where misreading behaviour has a cost of its own.