Stereotyping is a big interest of social psychologists, and it's long established that we make judgments about the personality of strangers based on their appearance. These judgments are reliable - that is to say, that people tend to agree about what personality a face represents, at a level higher than chance would allow. An interesting issue with these kinds of phenomena is how they get started: is there a 'kernel of truth' that tipped judgments one way or the other, resulting in the far more sweeping and gross generalisations that now exist? We should bear in mind that in previous times, the idea that elements of personality persist in the face wasn't just a scientific question, it was a presupposition: The art of physiognomy, or reading faces, was employed as far back as the ancient Greeks, as recently as Schopenhauer (see here for some of his thoughts on the issue) and has been popular between the two. In the present day, people still attribute the same kinds of traits to the same faces, and perhaps more surprising, they explicitly believe that the appearance of the face reveals personality.

However, it is up to the research to show that this reliability among viewers of a face has any correspondence to the genuine trait exhibited. I should note that one doesn't need to be a hardline nativist (someone certain that traits are innate, genetically determined and fairly resistant to change) to consider this hypothesis. Activity affects physiology, so in principle personality, by affecting your activities, could affect how you look. And how you look may shape your personality. So is there anything to this? A kernel of truth, a whole lot of truth, or no truth to speak of?

Previous research showed that people were better at chance at at least some of the personality judgments they made about faces. However this was marred by the common use of full-face photographs, which provide cues such as haircuts, jewellery and the like. Even a cropped photo may contain scars and other unique markers. This also doesn't tell you very much about whether there are types of face that correspond to types of personality.

Penton-Voak's work addressed this using interesting methods and providing exciting results. The first part of the method was fairly standard: they took 300 students and photographed their faces with neutral expressions, and asked them to complete a 40-item questionnaire which was designed to tap into the 5 factors of personality. Then, 100 participants rated the faces on each characteristic - e.g. agreeableness - one at a time. As expected, these rating were reliable. Also, they correlated with some of the questionnaire measures: Extraversion, and also male Neuroticism and Openness to expression.

So far, so expected. Here comes the geek bit: concentrate. For each personality factor, Penton-Voak took the faces of the people who rated themselves as highest on that factor (the top ten percent) and used them to create a personality composite. This was achieved by overlaying each face, preserving their commonalities but gradually smoothing out the differences to arrive at some kind of 'platonic ideal' of the face for that factor. The same was done with the bottom 10 percent on each scale. Would candidates just presented with this averaged information still be able to discern which was the agreeable and which the unagreeable ideal?

Turns out they could, for the following dimensions: Agreeableness, Extraversion, and Neuroticism. So, if anything, the composites stripped of individual social cues such as distinctive haircuts, allowed better perception of the personality trait that underlay them. Penton-Voak suggests that this may be because if the face does carry personality information, a single face is carrying information from multiple, independent sources (as there are proposed to be 5 independent personality measures) and this muddies the waters, as we find it difficult to ignore information and tend to incorporate it into our judgments, even when it isn't helping.

They also made another kind of composite: the product of those faces high or lowly rated on a dimension by the viewers. It would seem trivial to say that these composites continued to be rated as high or low on the scale, just as their components were, but remember that this suggests that some component of personality stereotyping has roots in across-face averages, rather than picking and operating on idiosyncracies. Moreover, using a form of statistical investigation termed discriminant analysis, they revealed that underlying those judgments made upon women was a reliance on their level of perceived attractiveness - their agreement was based not on a shared impression of what constitutes agreeableness in a female face, but a shared supposition that attractive faces are more agreeable. This is termed a halo effect - where perception of one feature determines another one.

In what for me is the most exciting aspect of the research, Penton-Voak took this yet further. If these composites really were ideals, then it should be possible to apply them to fresh data and in effect, produce a transformation along that dimension. He took his strongest candidates - the high and low Agreeableness composites - and applied them to new faces, morphing them towards those common aspects the composite held. These faces, both unknown and famous, were then rated on all dimensions. As hoped for, those morphed towards the high-Agreeable composite were judged more agreeable, and those towards the low judged less. But the most impressive finding was that there was no effect of the morphing on any of the other personality dimensions - they were not rated as more or less extraverted, neurotic etc. It was totally specific to one - in effect, isolating the 'agreeableness' transformation.

This seems compelling evidence that this method has locked on to a proto-face structure that communicates information about one personality dimension only, and promises a method of interrogating whether multiple personality dimensions can be communicated this way, and whether they map onto the classic Big-5 measures. Penton-Voak thinks not - not all the distinctive personality differences that are researched in people will be coded in any sense in the face - but thinks he can find at least one or two more using this method. It also holds out the possibility that, in images at least, we may be able to morph our personality.

I can't resist spelling out the applications of this kind of work in true Mind Hacks fashion...actually, I'll let Ian do it for me:

...the computer graphic faces generated by this project will be useful in an applied setting, as they will allow controlled alterations of perceived personality in faces. The use of computer generated characters (avatars) that successfully elicit personality judgements of the designer’s (or user’s) choosing may increase the usability of, and satisfaction with, computer interfaces. With the increasing use of avatars in many forms of human computer interaction, the current project has value as a first step towards the principled use of facial characteristics in computer graphic avatar design...

As a coda, I should add that the possibility I held out above, that it's personality that is presently morphing our faces is stronger than mere conjecture; studies from the 80s show that couples grow more alike over time in physical appearance, and it's no great step to say that those that smile together wrinkle together in the same way.