[ C H R O N I C L E ]
Soon, deep learning will enable us to sew a virtual pink triangle on the forehead of all homosexuals! In September, two scientists at US Stanford University, Yilun Wang and Michal Kosinski, will announce the forthcoming publication of an article in the Journal of Personality and Social Psychology, in which they claim to detect sexual orientation using deep learning. A prepublication available on the web provides details about the method used and the results obtained; photos of faces, taken from dating sites and annotated with the declared sexual orientation of the people photographed, are used to teach an IT programme through machine learning. The authors claim that the program is able to automatically recognise a homosexual from facial features with more than 91% accuracy, far greater than when men or women try to guess the sexual orientation of their peers.
According to the authors, this work provides a scientific basis for the ancient art of physiognomy, a method to read a person’s character from their facial features. To justify these results, they allude to rather ambiguous physiological theories according to which prenatal and postnatal exposure to higher or lower levels of hormones would determine both sexual orientation and the shape of the face, in particular the development of the forehead and the nose, and their hair. Beyond the possible uses by the police which recall to mind very unsavoury episodes of our history, this raises a certain number of ethical comments.
If these results turn out to be true, this would mean that an individual’s physical appearance, in particular their face, would be an almost exact reflection of their personality and that it would give them away by expressing what is deep inside. Kindness, probity or villainy would be read on our faces where they remain etched and there would be nothing we could do about it. This idea originating in physiognomy, pushed to its limit, is the same as saying that the outside takes precedence over the inside, and thus the emptiness of any moral conscience or individual will. It would also mean that desires, here sexual orientations, would be caused by sequences of physiological processes independent of the subject. As a result, we would all be the toys of hormonal phenomena, with no free will, no possibility of changing and as a result, no responsibility. Finally, this would mean that machines would know more about us than we do ourselves and that they would now exceed us to the point of claiming to govern us more rationally, and thus better than ourselves.
Let us underline that, contrary to what this publication in a prestigious social psychology journal concludes, this is not science as the results obtained in no way prove any causal link. Supervised learning only makes correlations! There may indeed be a link between homosexuality and some facial features, but it also may be that the habits of homosexuals, such as their hair style, are reflected in their appearance. It is surprising therefore that research is continuing into such doubtful hypotheses, especially in one of the most prestigious universities in the world, and that it is funded and published. However, it is even more surprising that Stanford university’s ethics committee approved this research.
Undoubtedly, the privacy of the people whose image was used for the learning was protected; they also undoubtedly gave their consent by signing the entry conditions on the dating site where they were registered. But are the check boxes on the forms of many European and North American institutions sufficient to detect and forbid such ethically controversial research?
Artificial intelligence researcher
Jean-Gabriel is a professor of IT at the Pierre-et-Marie-Curie University (Paris 6). He chairs the CNRS ethical committee (Comets).