Mark Simpson feels the gay bumps and lumps that betray the ‘gay brain’

Everyone loves Spotting The Gay. It’s a fun game the whole family can play.

Scientists seem to love it too. There have been a number of studies in the last decade or so that claim to demonstrate that ‘gaydar’ – the name implying microwave-accuracy we like to give to gut-instinct gayspotting – exists.

Though the results are not exactly stunning: only a bit better than chance, even after excluding bisexuality and pretending the world is neatly and helpfully divided half and half into heteros and homos. 10% more than chance is certainly ‘statistically significant’ in a laboratory, but not really the kind of accuracy you would trust to land your jumbo jet.

In fact, a series of intriguing recent studies by William Cox and his colleagues at the University of Wisconsin-Madison have suggested that even that accuracy, modest as it is, has been massively overestimated because a major flaw in the mathematics of previous research. Given the low incidence of homosexuality in the general population, 60% accuracy in the laboratory where subjects are 50% gay/straight translates into 93% inaccuracy in a real world in which 95% of the population are straight.

You really wouldn’t want ‘gaydar’ anywhere near Air Traffic Control.

Moreover, their research claims to debunk the whole concept of gaydar as nothing more than a legitimising myth for stereotyping. They argue that stereotyping is not, in general, an accurate way to judge sexuality and showed the more you validate the stereotyping the more people jump to conclusions.

Participants in one of their studies were asked to judge whether men were gay or straight based on whether they had interests that related to gay stereotypes, like fashion, shopping or theatre. Others had interests related to straight stereotypes, like sports, hunting or cars. Those who were told gaydar is real stereotyped much more than the control group and participants stereotyped much less when they had been told that gaydar is just another term for stereotyping.

I’d add here that in a sense inaccuracy is the point of ‘gaydar’. The ‘false positives’ are what it’s really targeting – it’s a system of surveillance. Gay stereotypes, particularly in regard to male homosexuality, are not so much aimed at gays, as at men in general – gayspotting is actually about spotting and discouraging male gender non-conformity. That 60% ‘accuracy’ claimed in previous studies means that for every 100 people, there will be 38 straight men incorrectly labelled gay, but only three gay people correctly labelled. Result!

Cox et al conducted another, slightly hair-raising experiment to show how stereotypes work. They had participants administer electric shocks to a male subject in the other room: ‘Participants learned only one thing about this other person, either that he was gay or simply liked shopping.’

Prejudiced people tended to refrain from shocking the man who was confirmed as gay, but delivered extremely high levels of shocks to the man who liked shopping. The researchers conclude that this experiment shows how stereotyping can give people opportunities to express prejudices without fear of reprisal. You can imagine them explaining: But, but… I didn’t electrocute The Gay!.

I would go further. I would say that it indicates how non-gay men who like things that are ‘gay’ are nowadays frequently safer targets for prejudice than (out) gay men – and as I say, homophobia, while particularly bad of course for gay men, is mostly about controlling the behaviour of non-homo men, of which there are a much larger number. So Cox et al’s shocking experiment was in effect a study in (American) metrophobia – I’m gonna fry that metrosexual faggot!

It’s not just straight people who like stereotypes, however. Gays can be even keener on gayspotting, and even more convinced of their unerring accuracy. Especially if the subject is hot or famous. And if they’re both then, my dear, there’s no question. Sometimes it almost looks like a re-enactment of their own childhood – though this time around they’re the ones shouting GAY!! (See much of Perez Hilton’s career.) Though in tests, les-gays, for all their skill at reading ‘gayness’ are only marginally better than straight people at gayspotting. Which is to say, only slightly less rubbish.

I know my gaydar’s rubbish – and I’m always the last to know. But even that knowledge doesn’t stop me from making snap judgements about strangers that are: a) entirely subjective, and b) really none of my fucking business. I can’t help but wonder if the fascination with spotting something is directly related to one’s actual incompetence. After all, if you were any good you would probably get bored with the game very quickly indeed.

And people are anything but bored with gaydar. My good chum Martin Karaffa first drew my attention to the flurry of publicity surrounding the recent paper, ‘Deep Neural Networks Can Detect Sexual Orientation From Faces’ (Yilun Wang, Michal Kosinski). The attention the paper received was not surprising, bringing together as it did dating profiles, Artificial Intelligence and ‘gaydar’. COOL!

‘AI Can Tell If You’re Gay From a Photo, And It’s Terrifying’, screamed one typical headline.  Typically false, that is. AI can’t tell from a photo whether you’re gay or not, and the authors of the paper don’t really claim to have achieved that. Additionally – and I know this will come as a terrible disappointment to many – they are very sceptical about the very concept of ‘gaydar’ as most people understand it, and in fact provide further evidence of its rather severe wonkiness (more on that later).

For their study they seem to have harvested profile photos of (white) men and women from an (unnamed) dating website, labelled them ‘gay’ or ‘straight’ according to the gender of the partner they said they were looking for, and then taught a computer to Spot The Gay – or rather, choose from pairs of photos of men or women where one was ‘gay’ and one was ‘straight’, which one was more likely to be ‘gay’. (I’m using scare quotes here because by the authors’ own admission, they themselves assigned the sexuality of the people in those photos – and, naturally, excluded bisexuality as a category.)

Under these carefully controlled conditions the AI guessed right 81% of the time for men and 71% of the time for women. Which sounds impressive, and is certainly significantly better than previous human judges studies, but it’s important however to remember again that the starting point for each pair of photos is 50% – the probability of guessing correctly by chance.

Some have critiqued the reporting of the paper and the paper itself  – particularly the way it seems to resurrect the once-fashionable 19th Century pseudoscience of physiognomy and phrenology, where an individual’s character is supposedly ‘written’ into someone’s appearance – the size of their nose, the bumps on their head.  The equally old-fashioned (but recently re-fashioned) biological theory it calls on, in this case the theory that homosexuality is the result of endrocrinal abnormalities in the womb, has also been questioned.

According to prenatal hormonal theory (PHT), gay men are gay because they didn’t get enough testosterone in utero and are thus ‘feminine’ or improperly masculinised. Lesbians are lesbian because they got too much and have thus been butched. In a sense, homosexuality doesn’t really exist – just hormonally-confused heterosexuality: masculine goes for feminine and that universally. (Though it seems to me this theory fails to account for how gay men can be attracted to other gay men – who would have to have the same ‘feminine’ orientation, psychology and appearance and thus be doubly or triply unsuitable love-objects.)

The paper’s authors Wang & Kosinski describe PHT as ‘widely-accepted’, which seems like something of an overstatement. Besides, the most recent PHT study rather embarrassingly failed to find that gay men were exposed to less testosterone than straight men – falling back on speculation that gay men might be ‘less sensitive’ to testosterone. (It did claim to find however that lesbians were exposed to more testosterone than straight women.)

What Wang & Kosinski say they did for their study was train their AI to look for ‘feminine’ features and behaviours in men and ‘masculine’ features and behaviours in women. Hence their claim that their findings help confirm the PHT theory of sexual orientation – and obviously they believe that this is a ‘good thing’.

Our results provide strong support for the [prenatal hormone theory], which argues that same-gender sexual orientation stems from the underexposure of male fetuses and overexposure of female fetuses to prenatal androgens responsible for the sexual differentiation of faces, preferences, and behavior

As a lay person it appears to me that for a number of reasons they’re wrong to claim confirmation of PHT, but were absolutely right to imagine they would be thanked for this claim in an era when essentialist ideas about sexuality are in vogue – as much as, paradoxically, essentialist ideas about gender are not. You can apparently choose your gender now, but your sexuality absolutely has to choose you.

So Kosinski himself, responding to criticisms of his study from GLAAD and others who described it as ‘junk science’ claimed it supported LGBT rights:

“It’s a great argument against all of those religious groups and other demagogues who say, ‘Why don’t you just change or just conform?’ You can’t stop, because you’re born this way,”

And he has a point, though perhaps not the one he thought he was making. In the US, where this study was conducted and where the culture wars rage (and thanks largely to social media are increasingly being exported to the UK, whether we want them or not) the God/Darwin made me this way tendency – what we might call the Gaga Doctrine – has dominated liberal and ‘progressive’ discourse for some time. Largely because it seems to represent an irrefutable and ‘immutable’ answer to the religious right moralism of ‘it’s a choice!’

But in addition to being, as Kosinski’s deployment of the argument demonstrates, not actually very progressive (‘I can’t help myself!’) but rather a reaction to reactionary ideas, this determinist approach is anyway very likely also doomed to failure. Establishing a predisposition for same sexiness – the most you could reasonably hope for, and even then not in all cases – is not enough for the Gaga Doctrine. Because with predisposition, as the word suggests, experience is decisive. The whole point of being ‘born that way’ is to get rid of all of that messy stuff called ‘life’.

And Kosinski isn’t the one resurrecting dodgy physiognomy, it’s already been galvanised Dr Frankenstein style – by ‘progressive’ scientists who want to in effect Spot The Gay in the womb in order to prove that homosexuality is congenital – a ‘third sex’. Often scientists who are themselves gay and on a quest to prove they were ‘born that way’, baby. (Kosinski has just had more publicity than most – and possibly been more clumsy than most; and since he is not gay himself is easier to attack.)

So tenuous claims about gay hair whorls, gay drivers, gay index fingers, gay genes and gay brains are to be welcomed because they reveal God’s amniotic plan for sodomy – a kind of gay creationism. Likewise ‘gaydar’ studies that ‘out’ the inborn gay essence in gay faces. Hence the public support for the study provided by a prominent gay scientist, fervent believer in ‘gay brains’ and author of a book called ‘Born Gay’.

To a non-believer like me – it offends my amour propre to think my sexuality is in effect nothing to do with me – probably the most interesting aspect of the ‘AI gaydar’ study is that it goes out of its way to point out how poor actual humans are at gayspotting, despite the allegedly decisive effect of that allegedly faulty amniotic fluid. In other words, the AI gaydar study torpedoes gaydar as it is understood by pretty much everyone.

They point out that the ‘gender atypical’ nature of gay men and lesbian faces they claim to have shown are only an average aggregation of ‘subtle’ differences – and not applicable to all gay men and lesbians. Most relevantly, and most disappointingly for many, they state: ‘Our results in no way indicate that sexual orientation can be determined from faces by humans.’

They even ran their own large human judges study – asking men and women from the US to choose between two photos of two men (and two women), one gay/lesbian and one straight – selecting the one more likely to be gay/lesbian. 50,000 pairs were used for each gender. They achieved an accuracy of 61% for men and 54% for women – ‘comparable with the accuracy obtained in the previous studies, which range from approximately 55 to 65%’.

They conclude that this ‘confirms that humans are rather inaccurate when distinguishing between facial images of gay and homosexual [sic] individuals.’

Regarding their own AI study, which of course had significantly better results, they allow this criticism:

‘It is possible, however, that facial images posted on a dating website are particularly revealing of sexual orientation. Perhaps the users selected the photos that their desired partners might find the most appealing.’

No shit, Sherlock.

In fact, part of the reason why they ran the human judges study was to try and prove that there was nothing especially ‘revealing’ about those pics – the fact that the human viewers scored similar (poor) rates to earlier studies demonstrates this, they believe.

Along with many other critics, I’m not so sure. Photos posted for a same-sex date/hook-up on a (presumably mostly hetero) dating site are almost by definition presenting you in a certain way, even if the human eye wasn’t (consciously) picking it up. Otherwise, why post them? They are self-selected photos uploaded by self-selecting homos: out enough to put their faces on a dating website, and one used by mostly straight people. And also by ‘gaydar’ researchers who may or may not have been breaching the dating site’s confidentiality rules.

It does seem to be the case however that the AI here successfully picked up what the human eye didn’t. Though exactly what the AI was doing seems to be a bit of a mystery, as apparently is the case with AI and algorithms generally. What’s evident though is that the data processed was not just ‘fixed’ or inborn facial features, as much as these can anyway be reliably measured from self-posted photos on a dating website, but also what the report calls ‘expressions’ and ‘grooming styles’. (Interestingly, Cox’s studies debunking gaydar – strangely unmentioned in this paper – found that ‘gaydar’ was largely an artefact of the happenstance that gay and lesbian self-selected photos tended to be of better quality than ones used by heteros: once this was eliminated there was no ‘gaydar’ effect.)

The authors cover themselves here however by claiming that they weighted facial features more heavily (something currently impossible for anyone else to verify), and postulating that anyway the ‘gender atypical’ expressions and grooming of their ‘gay’ subjects was probably also down to the fact they weren’t exposed to enough/too much testosterone in utero. Gay grooming is, in other words, also a product of faulty prenatal grooming. PHT theory is almost womb-like in its all-enveloping-ness.

In a statement that for me draws perhaps the biggest question mark over their whole study they also claim that they found that gay men were less likely to have facial hair than straight men. And again wonder whether this might be down to, you guessed it, insufficient pre-natal man-horms.

I don’t know about you, but now I really want to know now what this dating site was where they found all these clean-shaven horny gay men. For the last decade or so – and long before straight men started growing beards en masse, following in fact gay men’s example – it has been practically impossible for a gay man to get laid in metropolitan areas without face-fur.

They also offer this ‘revealing’ nugget about baseball caps:

‘consistent with the association between baseball caps and masculinity in American culture…, heterosexual men and lesbians tended to wear baseball caps (see the shadow on their foreheads)’

So now we know their AI was definitely measuring things that are not ‘inborn’. No one is delivered wearing a New York Yankees cap – not even a lesbian positively marinated in amniotic androgens.

It may be true that out gay men generally don’t seem to be terribly fond of wearing baseball caps – especially when trying to get laid on a straight dating site. But perhaps this has less to do with any underexposure to testosterone than an overexposure to fashion sense.