The Face-Inversion Effect in Autism Is Half as Large

What the research found

In the experiment, a photograph of a face is rotated by 180 degrees and the participant is asked whose face it is. Accuracy falls and responses slow, in almost everyone, and the drop is among the most reliable results in the psychology of perception. Every feature survives the rotation, and only the arrangement of those features in space has been disturbed.

That is what makes the manipulation informative. Recognition that depends on arrangement, on the eyes sitting above the nose at a particular spacing, suffers badly when the arrangement is scrambled, while recognition that proceeds feature by feature suffers far less. Psychologists call the first kind configural or holistic processing, and the inversion penalty is the standard way of estimating how much of it a person uses.

Jason Griffin and colleagues at the Yale Child Study Center pooled every published study that had run this comparison in autistic and neurotypical participants, in JAMA Psychiatry in 2023. Four databases searched through August 2022 yielded 38 eligible articles and 122 effect sizes, drawn from 1764 participants: 899 autistic and 865 neurotypical. Sample ages averaged 18.2 years, samples were 82.6 percent male, and the individual studies ran from children of four and five to adults in their thirties.

Results are reported in Hedges g, which measures a difference in standard deviations. Neurotypical participants showed the expected inversion penalty at g = 0.95. Autistic participants showed a penalty of g = 0.52, with the whole posterior distribution favouring upright faces over inverted ones. Between the groups the difference was g = -0.41, on a 95 percent credible interval from -0.63 to -0.18.

Both halves of that result carry weight. Autistic participants recognised upright faces more successfully than inverted ones, consistently across the pooled studies, which puts configural processing plainly in operation. Their advantage reached roughly half the magnitude of the neurotypical advantage, which puts the degree of specialisation lower.

The gap widened when the task was emotion. Studies asking participants to identify an expression produced a group difference of g = -0.79, interval -1.22 to -0.29; studies asking them to identify a person produced g = -0.33, interval -0.58 to -0.06. Electrical measures recorded from the scalp within a fifth of a second of a face appearing gave a group difference of g = -0.28 on an interval that included zero, the least clear separation in the analysis.

Variation among the 122 effect sizes was considerable, and 30 of the 38 studies included no non-face comparison, leaving the specificity of the effect to faces unresolved. Sample age predicted nothing about the size of the difference, which appeared in the studies of four-year-olds and in the studies of adults in their thirties.

What it means for you and your child

Configural processing is what allows a face to be identified at a glance and at the far end of a corridor. Featural processing locates individual parts, the eyes or the hairline, and assembles an identity from them. Both routes arrive at recognition, and the pooled evidence indicates that autistic participants, on average, lean on the second route more heavily than neurotypical participants do.

An account of that kind predicts recognition that holds when features stay put and becomes effortful when they change. A teacher encountered in a supermarket rather than a classroom, or a cousin who has cut his hair, would be the harder case. That prediction was not tested in these studies, and it is an extension of the mechanism rather than a finding of it.

The largest group difference in the whole analysis came from tasks asking what a face was feeling. Identifying who someone is can be done at leisure; reading an expression happens inside a second, while the conversation carries on without waiting. The authors note that the diagnostic criteria for autism concern the reading of emotion in faces, which is why they treat this moderator as the clinically meaningful one.

The pooled figures are averages across studies that varied a great deal among themselves, and no individual child is described by one of them. Their value lies in settling a question the smaller studies had left open for thirty years, since excluding the outlying results and adjusting for publication bias moved the estimate only to g = -0.45 and g = -0.44.

The inversion effect is a difference score, computed within each person, between performance on upright faces and performance on inverted ones. It describes the shape of an individual’s advantage and reports nothing about how accurate that person was overall, and nothing whatever about effort or interest in other people.

These 38 studies describe a face system tuned with a different emphasis, operating at every age the literature sampled, and consistent enough in direction to be quantified. That is a statement about perceptual style. They measured photographs on a screen, and the corridor and the supermarket await a study of their own.