Industry still asking, does AI have gaydar? Should it?
Opposition is growing in the EU to facial recognition algorithms billed as being able to accurately assign genders and sexual identities to digital images.
A 2019 report in the journal Proceeding of the ACM on Human-Computer Interaction found that gender classification is common in face analysis software. And in all cases, gender is presented strictly as a binary feature.
Advocacy group Access Now is organizing a campaign, along with facial recognition skeptics in Reclaim Your Face, to ban biometric mass surveillance in the EU’s public places, particularly attempts to identify and surveil LGBT+ people.
It is not simply that the software might be inaccurate, although that is a major concern. In fact, the few researchers who have tried to make the case that computers can be programmed with accurate sexual orientation insight have found some positive results.
A 2019 article by The Register, found that a pair of AI systems were able to demonstrate better accuracy than chance, and in some perspectives, demonstrated surprisingly high accuracy. (More studies would be required to draw solid conclusions.)
Access Now claims that merely programming a computer to do so about non-binary people, reinforces harmful outdated stereotypes. Even fuzzy logic cannot sufficiently identify people reliably on a spectrum as nuanced as this. A recent IPVM report meanwhile found that most demographic analysis systems judge gender solely or almost entirely on hair length.
It is more than an academic matter. There remain at least eight nations that punish homosexuality with the death penalty. Many more whip victims of reactionary governments and societies. Seventy nations have national laws outlawing same-sex conduct, according to Human Rights Watch.
Indeed, 11 U.S. states still prohibit consensual same-sex conduct despite the U.S. Supreme Court ruling such regulation unconstitutional almost a decade ago.
If nothing else, politicians could use just the existence of gender- and sexuality-divining biometric algorithms in their surveillance systems to cow populations.