Does a Face Mean the Same Everywhere?

In 1967 Paul Ekman sat in the highlands of New Guinea, members of the Fore in front of him and a stack of photographs of Western models' faces in his hand. He told a short story — “this is a man whose child has died” — and had the matching face picked out. The Fore chose correctly above chance, and out of this scene came the textbook sentence of the following fifty years: emotions have faces, and the faces are understood everywhere.

Two details of the scene stayed in small print for a long time. First, the choice was given in advance — the Fore could only choose among the pictures lying on the table. Second, one pair did not work: fear and surprise were reliably confused, and the fear story had to be specifically extended before the confusion could be damped down. Universality was a question of procedure from the start. It merely took half a century before anyone made the procedure itself the object.

The basic-emotions programme

Out of the fieldwork grew a research programme: a small number of evolved base programmes — joy, sadness, anger, fear, disgust, surprise — each with its own facial expression, its own physiology, universally readable. The Facial Action Coding System, which Ekman presented with Wallace Friesen in 1978, measured the facial muscles involved and made expression countable for the first time. From the programme grew lie-detection trainings, airport security schemes and a textbook consensus that held for decades. In 2009 Time counted Ekman among the world’s most influential people.

The historical background explains why the finding landed as it did. In 1884 William James, whose file lies in this archive, had reversed the order:

The bodily response comes first, the feeling is its perception. Wundt sorted experience instead into dimensions: pleasure and displeasure, excitement and calm. Ekman’s six kinds seemed to settle the quarrel in favour of discrete natural kinds — and to do so with data from the field rather than with introspection.

What the forced-choice format produces

The methodological objection came in 1994 from James A. Russell, long before the big debate. He read through the cross-cultural studies and showed how much of the agreement sits in the arrangement: posed photographs of models who produce the expression in exaggerated form; a handful of prescribed answer words; participants who see the same faces repeatedly and in doing so learn which categories are on offer at all. With six answer options the guessing rate is already close to 17 per cent, and anyone who can confidently rule out two pictures raises it further without any knowledge of emotion.

Lisa Feldman Barrett has built this line into a counter-theory. Her starting point remains the measurement: whoever hands participants six words and six faces builds the answer structure they then find — take the word lists away and let people describe freely, and the agreement between cultures drops sharply. When Himba in Namibia were allowed to sort the expression photographs freely, what emerged instead of six clean clusters were piles by visible behaviour; anger, disgust and sadness landed together. In free labelling of Western emotional vocalisations, what mainly remained was pleasant against unpleasant and, more weakly, calm against aroused: Wundt’s dimensions, not Ekman’s kinds. Another research group put the Western fear face to adolescents on the Trobriand Islands; the majority read a threat display in it — not someone who is afraid, but someone who makes others afraid.

What the body does not yield

The second attack came from physiology. If six base programmes exist, they should show up in the body. A meta-analysis by Siegel and colleagues evaluated more than 200 studies with over 8,400 people in 2018 and found no reliable autonomic fingerprints: heart rate, skin conductance and respiration do not cleanly separate anger from fear. The same holds for the brain — no amygdala is “the fear centre”; it is a structure involved in a great deal and one that can be absent in some cases without fear disappearing.

Barrett’s counter-proposal, the theory of constructed emotion, builds feelings from two ingredients: core affect — the bodily baseline of valence and arousal — plus concepts that a culture supplies and a brain applies predictively. Anger, on this account, is not a circuit but a learned category over bodily states in situations; real the way money is real, through shared convention.

The collision became practical when applications were at stake. The large collective review of 2019, in which Barrett was a leading author, certified for emotion recognition from faces what the replication crisis has certified for many applications: the mapping from expression to feeling is on average weak, context-dependent and culturally variable. A furrowed brow sometimes means anger, often concentration, occasionally nothing. For software that sorts applicants or passengers by their faces this is a devastating finding; the billion-dollar field of “emotion AI” stands on a premise its own basic research no longer carries. The airport programme SPOT, for whose inspector training Ekman was under contract in 2007 for a good million dollars, had already been cleared away in 2013 by the US Congress’s auditors: the available evidence did not support the claim that dangerous persons can be picked out by behavioural cues — the report recommended cutting the funding. That is expressly not the same as refuted, but unsupported; the difference is exactly the one this archive otherwise defends.

A word of fairness towards the older programme: it was built to be testable, and that is no small thing. Ekman put photo sets, coding systems and predictions on the table that fifty years of research could work against. The construction theory owes its ammunition to the data the basic-emotions programme produced in the first place. A programme that makes its own refutation possible has worked even when the refutation arrives.

Where the quarrel stands today

The honest balance is unevenly distributed. Well supported: some expressions are read similarly above chance, far above chance and far below the perfection of the textbook charts. Context, culture and method shift the values massively. Fixed bodily fingerprints of individual emotions are missing. And the dimensional structure of valence and arousal surfaces in practically every dataset.

A meta-analysis by Hillary Anger Elfenbein and Nalini Ambady measured, in 2002, the intermediate position most specialists now hold. Across many cross-cultural comparisons the hit rates lay well above chance — and at the same time were systematically higher when expression and observer came from the same cultural group. The basic-emotions camp has taken this home advantage on board: six programmes became families with cultural “dialects”; hardly anyone still defends the strong version. The construction theory in turn has its own testing problem. It is so flexible that critics ask which observation it rules out — the Popper argument this journal has already seen deployed against dream interpretation.

The history of the German language hands the construction side its prettiest everyday argument in passing. Words like Schadenfreude, Fernweh or Torschlusspanik sort bodily states into feelings for which other languages hold no category — and whoever knows the word knows the feeling more precisely. That categories sharpen experience is measurable in the small: trainings that differentiate emotion vocabulary demonstrably change how finely people report their own states and how they regulate them. No proof of the grand theory, but a pointer to where its ingredient “concept” actually does its work.

What remains

The answer to the question in the title is: partly. A face is read above chance everywhere and nowhere as unambiguously as the six textbook tiles suggest — and how far the hit rate falls short of unambiguity depends on the procedure, not on the emotion. What is settled is the direction of the finding, not its size. What remains contested is the explanation: whether natural kinds with cultural dialects or constructions with biological ingredients stand at the end will be decided by types of data only now coming into being — field recordings instead of posed photographs, free description instead of an answer list, courses over time instead of snapshots.

In everyday life a plain rule follows: a face is a hint, not a verdict. Anyone who wants to know how someone is doing asks — and wherever asking is possible, reading faces is the worse method. For software that sorts job interviews, classrooms or security checkpoints by faces, the same sentence holds with added force, because there nobody can object. And for the history of the field one point remains: the quarrel has landed back with James, where it started. Barrett’s predictive brain, categorising bodily states, is a modern version of his sentence that the feeling is the perception of the bodily response — enriched by what James lacked: concepts as an ingredient, prediction as a mechanism. The chapter the textbooks carried as settled had never, in a hundred and forty years, been decided.