Intrinsic vs. Optional Signals: Is there a difference?
A few days ago, I came across this article. It’s a response to the so-called “AI Gaydar” paper by Wang and Kosinski, wherein the authors claim that they were able to train an facial classifier to identify the sexual orientation of the data subjects using only a photo of their faces.
A lot’s been said about the original paper by Wang and Kosinski, and I’m not sure I have much to add to it. What I’d like to talk about is actually the first article I mentioned, the one by Blaise Agüera y Arcas, Alexander Todorov, and Margaret Mitchell. In this article, Agüeras y Arcas et. al. are specifically concerned with refuting the claim that the identifiability Wang and Kosinski claim is somehow due to particular biological characteristics. Indeed, it seems that there is strong evidence that people present themselves differently depending on their sexual orientation, and that these differences, rather than any immutable physical characteristics are the best explanation for the Wang and Kosinski’s finding.
Or rather, self-presentation is another competing explanation for why the model was relatively effective. Jumping from an effective classifier to an explanation of why that classifier is effective isn’t a particularly sound jump. Deep neural networks, and indeed most other ML models, define patterns which more often than not lead to correct classification. They are a description, not a conclusion. They’re like that by design, in fact the field of ML grew out of a realization that conventional statistics was too concerned with causality and mechanism to be useful in tasks where we didn’t or couldn’t figure out a particular mechanism. To determine if there’s a particular explanation for a model, you need statistics.
That said, Agüeras y Arcas’ et. al. alternative explanation does seem much more likely, especially given the breakdown of the dividing questions. It’s not quite a rigorous study, but it doesn’t have to be to show that the conclusion Wang and Kosinski reach from their data is perhaps less than well-founded.
It’s easy to take away from Agüeras y Arcas’ et. al. that this research shouldn’t be of any concern. However, that’s not really what they’re saying. The signals they cite, while not absolutely inherent (certainly if you have a beard, it can be shaved off, you can take your selfies at whatever angle you want), are presumably part of a system of signs necessary for the subjects to perform parts of their social roles. If someone wants to signal femininity, they’ll need to adopt at least some detectably feminine characteristics. So while technically an individual may be able to subvert such a system, it may make other aspects of their life unteneble.
To the authors’ credit, they do acknowledge this at the end, but this acknowledgement is in tension with their claims at the beginning that they’re confronting “claims that AI face recognition reveals deep character traits”. Their argument seems to be that because the characteristics that are actually being trained on are not “deep”, neither this nor future similar systems are as powerful as one might have thought after reading the headlines. Despite the fact that the signals being measured appear to have no actual basis in biology, the fact remains that they are still signals. It doesn’t really matter if an ML model isn’t able to see into who I truly am, being able to see what sort of social roles I enact is more than enough.