Gaydar
September 7, 2017 7:47 PM   Subscribe

Stanford study: New AI can guess whether you're gay or straight from a photograph given a single facial image.

A classifier could correctly distinguish between gay and heterosexual men in 81% of cases, and in 74% of cases for women. Human judges achieved much lower accuracy: 61% for men and 54% for women. The accuracy of the algorithm increased to 91% and 83%, respectively, given five facial images per person.
posted by 445supermag (17 comments total)

This post was deleted for the following reason: There's the core of a discussible idea here, but the sort of glib, brief mainstream news article just sets up too many unwarranted assumptions for people to get any traction -- restless_nomad



 
How's it do with identifying bisexual people?

SEXUALITY ISN'T A BINARY, COMPUTER SCIENTIST BROS. STOP IT.
posted by SansPoint at 7:48 PM on September 7, 2017 [6 favorites]


No word on if it can recieve interference from a gay weather ballon.
posted by 445supermag at 7:48 PM on September 7, 2017 [3 favorites]




I mean...I guess a binary works if your categories are "marked for death" and "not marked for death."

I confess I haven't RTFA yet because HOLY SHIT WHY, but is there any conceivable use for something like this that isn't, you know, fucking evil?
posted by schadenfrau at 7:54 PM on September 7, 2017 [5 favorites]


Since this is going to go completely off the rails, and do way more harm than good anyway, I'm just looking forward to the day someone points it (very publicly) at a nice large corpus of high resolution portraits of the entire Republican congress.
posted by trackofalljades at 7:57 PM on September 7, 2017 [4 favorites]


Post flagged for all kinds of reasons.
posted by tzikeh at 7:58 PM on September 7, 2017 [2 favorites]


Another linked bit of info related to Kosinski.

I've directly collaborated with people who have directly collaborated with him. His name has been popping up in a lot of places over the past year. And it's all kind of... yikes.

In case the link doesn't resolve: it's a picture of Kosinski in a helicopter with the caption "Off to give a talk for Prime Minister Medvedev"
posted by supercres at 8:01 PM on September 7, 2017 [2 favorites]


Is a 81% hit rate even good? I guess it depends on the distribution of their test data. Like, I don't know what percentage of men are gay, but I have heard the number 10% thrown around. Always guessing "straight" on a random sample should then give you a 90% correct rate.
posted by RustyBrooks at 8:01 PM on September 7, 2017 [2 favorites]


Since this is going to go completely off the rails, and do way more harm than good anyway, I'm just looking forward to the day someone points it (very publicly) at a nice large corpus of high resolution portraits of the entire Republican congress.


Ugh, can we not do the "homophobes are secretly gay!" thing here.
posted by daybeforetheday at 8:02 PM on September 7, 2017 [3 favorites]


> I confess I haven't RTFA yet because HOLY SHIT WHY, but is there any conceivable use for something like this that isn't, you know, fucking evil?
I'm afraid not.

Basically, follow the money. Surveillance capitalism pays.
posted by runcifex at 8:04 PM on September 7, 2017 [1 favorite]


Is a 81% hit rate even good? I guess it depends on the distribution of their test data. Like, I don't know what percentage of men are gay, but I have heard the number 10% thrown around. Always guessing "straight" on a random sample should then give you a 90% correct rate.

I was talking to a friend who does natural language processing (he hates the paper, by the way), and he mentioned that their results are from equally balanced sets of gay and straight people, so this is beating random 1-to-9 guessing.
posted by codacorolla at 8:08 PM on September 7, 2017 [1 favorite]


Yeah, just this morning I was thinking that what the world really needs is a way to identify people's sexuality without any regard for how they feel about it. It'll be like my teen years all over again! Thanks, Science.
posted by shapes that haunt the dusk at 8:09 PM on September 7, 2017 [3 favorites]


Always guessing "straight" on a random sample should then give you a 90% correct rate.

This is an extremely apt point. They should be reporting AUC or F1 score for a classifier, which would be robust to unbalanced classes. Their samples are balanced, which is not ecologically valid. They're doing the much easier task of balanced samples.
posted by supercres at 8:10 PM on September 7, 2017 [1 favorite]


I mean...I guess a binary works if your categories are "marked for death" and "not marked for death." or cake.
posted by octobersurprise at 8:15 PM on September 7, 2017


Consistent with the prenatal hormone theory of sexual orientation, gay men and women tended to have gender-atypical facial morphology, expression, and grooming styles.

...yes, because we all made our grooming style decisions in the womb? That line confused me a lot. The fact that I don't pluck my eyebrows doesn't mean that I couldn't start doing so if I felt the need to.
posted by Sequence at 8:15 PM on September 7, 2017 [1 favorite]


Yeah, just this morning I was thinking that what the world really needs is a way to identify people's sexuality without any regard for how they feel about it

I'm having horrible, horrible Gattaca flashes.
posted by corb at 8:17 PM on September 7, 2017 [1 favorite]


If this were a thing I'd love to use it to plan my day. I move around in preferences a lot.
posted by poe at 8:20 PM on September 7, 2017


« Older Paging Dr. Dolittle. Bring a microscope.   |   taxonomic vandalism Newer »


This thread has been archived and is closed to new comments