New AI can guess a person's sexual orientation from a photograph
theguardian.com
theguardian.com
Now in the general population, the ratio of gays to straights is about 16 to 984 (1.6%)†. So if their test gives someone a positive reading, that increases the odds to 162 to 984, or 14%. So you can't use this test to accurately guess someone's sexual orientation. Simply because gay people are so rare that even a few percent of straight people misclassified will overwhelm the number of actual gay people.
But nonetheless it is much more accurate than humans or the base rate. And it is scientifically interesting that there are physical features that strongly correlate with sexual orientation.
EDIT: another article claims this:
>However, when asked to pick out the ten faces it was most confident about, nine of the chosen were in fact gay. If the goal is to pick a small number of people who are very likely to be gay out of a large group, the system appears able to do so.
†These estimates vary by a lot and could be 3 times that, which would make the test much more accurate. Perhaps as high as 42% accuracy. And some people believe sexuality is more of a spectrum than a binary straight/gay. If so the test might be identifying straight individuals that lean more gay/bisexual than normal, and it's misclassifications wouldn't be so unreasonable.
The test may also give varying degrees of confidence, and it may give much higher confidence to some people than others. The edit I made confirms that, and there are some individuals that it can tell are definitely gay or straight. But for most it is more uncertain.
Another interesting thing is that both the straight faces appear to be older and heavier set than the gay faces. I can't even imagine a possible explanation for that. I'd like to see the true composite of faces, rather than this one reweighted by their algorithm.
The bot didn't just go for fixed features.
Besides the fact that this sounds deeply offensive for gay people, it also sounds terribly, awfully unlikely.
In fact, it reminds of the recent "phrenology" paper, where deep nets were used to, allegedly, "identify criminals from their faces":
https://www.technologyreview.com/s/602955/neural-network-lea...
Nowadays we know that sexual orientation is something that is neither set in stone, nor absolute (people are known to change throughout their lives, as well as have occasional partners of the "other" sex, whichever is opposite to their normal orientation). So the very idea of an accurate "gaydar" is ludicrous in and of itself.
Finally, I wonder why anyone with an ounce of brains would seriously consider training a machine learning algorithm to recognise peoples' orientation. What morally justifiable practical purpose could there possibly be?
Offensive or not, sexual orientation is DNA based and so does your facial structure. What do you make of the claim that based on 5 photos they achieved a 90% success rate? That's significant.
Sorry, but do you have a source for that claim?
It has nothing to do with gender. You're just trying to proliferate your opinion that gender = biological sex. Anyone that is transgender would disagree with your opinion.
"As of 1999, about 500 species, ranging from primates to gut worms, have been documented engaging in same-sex behaviors.[3][page needed][4] According to the organizers of the 2006 Against Nature? exhibit, it has been observed in 1,500 species.[5]"
So if not DNA, then what else?
https://en.wikipedia.org/wiki/Homosexual_behavior_in_animals
I'm reminded of Gerd Gigerenzer's book "Reckoning with risk". It's about medical testing, so forgive the language used:
> A condition exists. There is a test for that condition. The test is good, but not perfect. If someone has the condition there is a 90% chance they'll return positive. If someone does not have the condition there is a 1% chance they'll return positive. About 1% of the population have the condition. Bob has the test, and it comes back positive. What's the probability Bob has the condition?
Most people, even doctors who give these types of tests, get this wrong. The answer is about 50%, but most people put it much higher at 99% or 90%.
I think the Guardian reporting is irresponsible because the general public do not understand percentages, and any reporter using percentages is misleading (albeit inadvertantly) the public.
Here's a link to the book: https://www.amazon.co.uk/Reckoning-Risk-Learning-Live-Uncert...
I'm assuming their model is badly trained and overfitting or they're misinterpreting (though I hope not misrepresenting) their results.
>Nowadays we know that sexual orientation is something that is neither set in stone, nor absolute
I don't think that's really correct. Sure there are exceptions, but most people stick to one sexual orientation, or at least somewhere in between on a spectrum. Attempts at converting gay people to become straight were unsuccessful. In fact this was one of the biggest arguments for gay rights, that it isn't a choice and can't be changed.
>What morally justifiable practical purpose could there possibly be?
This proves a biological origin for sexuality. Which is both scientifically interesting, and a strong argument for gay rights in places where it is still seen as some kind of sin or mental illness.
Anyone hypothetically using this tool to discriminate against homosexuals would have to admit the tool works of course. And that would mean homosexuality is biological, which would contradict their own beliefs.
Lastly do you have any actual argument against the papers methodology? Or the phrenology paper you link for that matter. Your comment is basically "I don't like the conclusions so I won't accept them". That's not terribly scientific.
The whole "gender is a social construct" movement is based on rejecting actual scientific evidence, so don't be surprised...
Because it's based on the idea that gay men aren't real men. You have to be a little suspicious when scientific research comes out with findings that happen to exactly match the prejudices of the experimenters and/or the broader society. There used to be lots of "scientific" evidence for racist ideas too (see e.g. http://lrc.ohio.edu/lrcmedia/Streaming/lingCALL/ling270/myth...).
>Which is both scientifically interesting, and a strong argument for gay rights in places where it is still seen as some kind of sin or mental illness.
It's only an argument for gay rights if you presuppose that "fixing" gay people would be the right thing to do if it were possible. If you don't presuppose that, then it makes no difference whether or not people are "born gay".
If a study found that male, violent criminals had wider jaws than average men, would you think that the study was based on the idea that average men aren't real men, unlike criminals?
I think it's quite likely that both sexual orientation and jaw size are influenced by the same hormones (e.g. testosterone), since both features are involved in sexual dimorphism. Essentially, if a person is similar to average members of the opposite sex in one dimension, they are likely to be similar in other dimensions as well.
That doesn't make anyone inferior or less "real", it just means that the averages of two groups are slightly different on a certain axis that roughly corresponds to the feminine-masculine distinction. On that axis, you will have
(most feminine woman) (average straight woman) (average gay woman) (most feminine man) (most masculine woman) (average gay man) (average straight man) (most masculine man)
That axis is very different from the inferiority-superiority axis, where you have
(nobody) (everyone else)
Ok. I think it's also quite likely that they're not.
>(most feminine woman) (average straight woman) (average gay woman) (most feminine man) (most masculine woman) (average gay man) (average straight man) (most masculine man)
Are you serious about this? Gay men are just men who are sexually attracted to men. You'd think that the idea that gay men are sexually attracted to men because they are more feminine than other men would have died in the face of all the obvious counterexamples, now that fewer people are hiding their sexuality, but it seems surprisingly persistent.
The problem with this kind of thing is that despite representations to the contrary, no-one is really interested in the group means for their own sake. They are interested because they want to pedal the tired old theory that gayness is a defect resulting from a lack of masculinity. People are often too polite (or afraid?) to say that now, but gay people are capable of reading between the lines.
I'm interested in group means for their own sake. Or rather, I'm interested in them because they hint at underlying mechanisms, and I'm interested in mechanisms for their own sake.
FWIW, I'm asexual and do not regard homosexuality as some kind of defect. Otherwise, what would I have to think of myself?
Additionally, I don't think you can "lack" masculinity in the sense that you "need" some amount and are deficient otherwise. There are some people who are closer to one extreme than to the other, and if likelihood of attraction to a specific sex changes somewhat monotonically with position on this axis, that tells you very little about individuals, but a bit more about biological interactions, which, again, I find interesting for their own sake.
And this study and other similar ones would seem to prove you wrong?
>You'd think that the idea that gay men are sexually attracted to men because they are more feminine than other men would have died in the face of all the obvious counterexamples
A few exceptions don't disprove a statistical correlation.
Besides it's more complicated than just "gay people are more feminine". There are various hormones and biological interactions other than just testosterone. And the effects happen early on in development; Whether an adult has lots of testosterone and muscles is irrelevant.
As this study shows, humans are much worse than this algorithm at identifying gay people. It's not merely a matter of "that face looks feminine" or human scorers would have done much better. Presumably many of the correlations are very small or subtle, but in aggregate they can get surprising accuracy.
A team reported their results with a model they trained. The results are accompanied by some hopeful theory about things that the authors have no way to know. There's no proof of anything in there.
A large number of exceptions to a generalization absolutely is sufficient to undermine a causal hypothesis. If X causes Y, then then you shouldn't frequently find Y in the absence of X. You do frequently find gayness in the absence of above average femininity.
>Besides it's more complicated than just "gay people are more feminine".
Of course it is more complicated that that. That's what I've been saying. But in the media, and on HN, the idea of relating homosexuality directly to a masculine-feminine continuum is very popular. Many do not realize that this is not science, but just old-school homophobia dressed in pseudoscientific garb.
Only if it's a deterministic causal hypothesis. Otherwise, the proportion of counterexamples only serves to bound the effect size, which is obviously quite small in this case.
I don't think anyone in this thread seriously believes that "almost all gay men [are] unusually feminine". If they were, they wouldn't be called "gay men" but "women with a penis" or something like that.
> That's what I've been saying. But in the media, and on HN, the idea of relating homosexuality directly to a masculine-feminine continuum is very popular. Many do not realize that this is not science, but just old-school homophobia dressed in pseudoscientific garb.
A direct relation would certainly be surprising, but that doesn't mean you can't study correlations between sexual orientation and certain measurements of the body. That's science and might lead to insights how and when sexual orientation develops. I don't really care what people make of it who confuse correlation for causation or don't understand that statistical significance doesn't imply a large effect size.
What a bizarre thing to say. People used to really think that all gay men were unusually feminine, but they still didn't call them women with penises, because they aren't.
>Otherwise, the proportion of counterexamples only serves to bound the effect size, which is obviously quite small in this case.
But why the interest in an effect which is small if it exists at all, and for which there is little evidence? There are probably hundreds of other features which correlate equally strongly (i.e., not very strongly at all) with homosexuality. This one happens to tap into pre-existing stereotypes.
>A direct relation would certainly be surprising, but that doesn't mean you can't study correlations between sexual orientation and certain measurements of the body. That's science and might lead to insights how and when sexual orientation develops.
The topic of the article isn't a scientific study. It's a couple of people in a business school playing around with deep learning for marketing purposes. They are not doing it because they're genuinely interested in uncovering the mechanisms underlying the development of human sexuality.
>I don't really care what people make of it who confuse correlation for causation or don't understand that statistical significance doesn't imply a large effect size.
In other words, you don't care what the majority of the population make of this research. That's fine for you if you're not gay and you're not affected by it. But the question of how this research is marketed, and how it's perceived by the general public, is potentially very important to gay people. It matters to us what the general public thinks about us.
"Gay bears" would disagree with you: https://www.google.de/search?q=gay+bears
The latter paper's approach is harshly criticised here: http://callingbullshit.org/case_studies/case_study_criminal_...
It is offensive in the sense that it seems to assume that gay men are less masculine than straight men. Why would that be the case?
Not to mention that, in practice, the vast majority of gay men are not effeminate.
>> This proves a biological origin for sexuality.
A claim in a paper never "proves" anything. In this case we're talking about a paper on sociology that presents a machine learning model so it's even less likely to prove anything about biology.
>> Lastly do you have any actual argument against the papers methodology?
There should be a law about requesting "arguments" and name-dropping "methodology" on the internet. Something like Godwin's law. Along the lines of: "When an online discussion reaches the point where someone mentions the word 'argument' it's time for everyone to go home".
At best, the only thing we know is that this is true for some people.
You say it yourself: "people are known" to change / have sexual partners of both genders -- because such people are an exception!
These people are a minority, and you really can't generalise and affirm that this is the rule for the majority.