September 8, 2017

"The study from Stanford University – which found that a computer algorithm could correctly distinguish between gay and straight men 81% of the time, and 74% for women..."

"... has raised questions about the biological origins of sexual orientation, the ethics of facial-detection technology, and the potential for this kind of software to violate people’s privacy or be abused for anti-LGBT purposes."
“It’s certainly unsettling. Like any new tool, if it gets into the wrong hands, it can be used for ill purposes,” said Nick Rule, an associate professor of psychology at the University of Toronto, who has published research on the science of gaydar. “If you can start profiling people based on their appearance, then identifying them and doing horrible things to them, that’s really bad.”...

Rule speculated about AI being used to actively discriminate against people based on a machine’s interpretation of their faces: “We should all be collectively concerned.”
"We should all be collectively concerned" — what a phrase! Mind if I be individually concerned or if I let other individuals or collectives engage in concern without my participation, at least for the nonce?

I'm not reading the details here, but since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time? Seems to me the machine it terrible at its work. 

And here's a simple solution: Let's not actively discriminate against people based on sexual orientation — except in the formation of sexual relationships — and then the machine won't have any relevant function. Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

Oh, but maybe in our formation of sexual relationships, we'd like a machine to double check whether our partners are actually sexually oriented to us. Maybe you'd like it built into Tinder or Match.com or whatever, some kind of notice that there's a 19 or 26% chance that this person is advertising the wrong sexual orientation. Again, the best solution is to stop making life difficult for gay people: They'll pursue their own desires and not people they don't really want. It's a win-win.

121 comments:

n.n said...

Transgender spectrum.

J. Farmer said...

Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

Exactly. This has been my position on the "born gay" argument since I was in high school. To quote Hillary, "what difference, at this point, does it make?"

Lance said...

since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time? Seems to me the machine it terrible at its work.

You could enjoy a second career as a software quality assurance engineer. Well, maybe not "enjoy", but at least you'd be good at it.

The Godfather said...

The article says "gay men appeared more feminine and vice versa." What is "vice versa" supposed to mean, that feminine men appeared more gay?

Henry said...

What does biology have to do with it?

A haircut is probably sufficient.

Henry said...

For 75% accuracy.

Rabel said...

The eye liner is the tell.

richlb said...

It's not the same percentage for men and women? That machine is sexist!

Nonapod said...

Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

There's an entire grievance industry whose very existence depends upon people caring about stuff like this. Since we don't have to worry too much about a lot of traditional day-to-day survival problems, making a big deal about things that have no effect on your own life personally is what this modern era is all about. We have to constantly invent new things to be concerned about or we might not ever feel like we're part of anything. We have to constantly search for new causes and form new tribes of concern. Being human who doesn't have to worry is a worry in and of itself.

n.n said...

The transgender spectrum only matters if it represents a progressive condition (e.g. dodo dynasty). As there is no evidence that transgenderism is a progressive condition, and since expression of the orientation can be moderated and does not pose an explicit threat to people, it seems reasonable to conclude that it can be tolerated, but since it has no redeeming value to society or humanity, it seems odd to construct a political congruence (e.g. "="), which effectively establishes active discrimination against politically incorrect (e.g. inconvenient) orientations. So selective.

Gahrie said...

If you can start profiling people based on their appearance, then identifying them and doing horrible things to them, that’s really bad.”...

People have been doing this ever since there have been people.

Charles said...

I'm not reading the details here, but since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time? Seems to me the machine it terrible at its work.

That's not correct. The machine is presented a single male picture and asked to identify whether the male is straight or gay strictly from the evidence of the picture. It has no capacity to judge what the respective populations are or whether the picture is randomly selected. It can only assess the probability on a single picture, therefore its chances of being right are 50% (like the boolean problem of flipping coins; what are the chances that the next coin toss will come up heads if you have already flipped ten heads? 50%)

For a single picture, the AI system is correctly identifying the individual's orientation 81% of the time (91% of the time when shown five pictures of the same individual) instead of the expected 50% correct.

Henry said...

Ah, here is the key phrase from the article. The algorithm looked at "features, expressions and “grooming styles”

Only one of these is hereditary. Yet that is what the article focuses on. That's the sexy.

Biological determinism (including the male / female kind that Google engineers ponder) is utterly co-opted by the fact that gene expression depends on environment.

Gahrie said...

The article says "gay men appeared more feminine and vice versa." What is "vice versa" supposed to mean, that feminine men appeared more gay?

Probably that gay women looked more masculine.

Gabriel said...

since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time? Seems to me the machine it terrible at its work.

An excellent way of benchmarking its performance.

A seemingly high accuracy rate, when applied to a small fraction of population, has notoriously counterintuitive rates of false positives.

Let's pretend that 10% of men are gay. If the machine is correct 81% of the time, then 19% of straight men are erroneously identified as gay, and 81% of gay men are correctly identified as gay. In a population of 100 men, 17 straight men will be wrongly identified as gay, and 8 gay men will be correctly identified as gay.

So if the machine says "gay", it's wrong 68% of the time (17 / (17 + 8)).

Widdershins likewise, if the machine says "straight", it's wrong 2.7% of the time (2 / (2 + 73)).

In both cases, the machine is accurate 81% of the time.

Now Ann's hypothetical--machine always says "straight", would be wrong 10% of the time, which is actually not better than the machine (wrong 2.7% of the time).

Gahrie said...

Again, the best solution is to stop making life difficult for gay people:

Is that really the problem today?

They'll pursue their own desires and not people they don't really want.

Yeah...tell that to Christian bakers.

The Vault Dweller said...

I'm eagerly awaiting fallout from this bit of technology. I predict a minor celebrity, being declared gay by this AI, the celebrity turning around and suing the creator of the AI as well as the person who published the claim for defamation, and in that case the Lambda Legal foundation filing and amicus brief with the court advising the court to declare that being called gay can not be defamatory per se.

tcrosse said...

There is no substitute for well-developed Gaydar.

rhhardin said...

Hips and breasts are a positive sign for female, at least if you're estimating at a distance.

That's cultural, since you're not born that way.

n.n said...

Male transgender/homosexuals do tend to be more feminine (i.e. female-oriented gender), not limited to the transposition of roles in a couplet's sexual relationship.

Gabriel said...

@Charles: Ann's machine doesn't guess, and neither does the AI. Yes, you are right, normally you would compare the AI to chance. But do the same with Ann's machine and it is right 90% of the ime which is ALSO way better than guessing, so why build a machine at all?

Her hypothetical was well worth exploring, because "right 81% of the time" simply doesn't mean what people think it does, and Ann's machine actually does worse than the AI.

Ray - SoCal said...

Are the pictures of the usual test sets, members of college campuses?

The challenge is most psychological tests, is they use as their test subjects, college students. Which are a bit different than the general population.

My guess is grooming. It would be interesting if the same tests were done with pictures from other cultures. Would the results be similar?

Gabriel said...

@Ray:pictures from other cultures. Would the results be similar?

Many other cultures reject the concept of gay identity, regardless of how much same-sex copulation is going on. Many subcultures in the US do as well.

Left Bank of the Charles said...

Althouse's sex robots are going to need this technology.

Henry said...

That AI gaydar machine needs to compete against actual gay people. IBM didn't build Deep Blue to beat a bunch of cub scouts.

I search the Althouse articles for "gaydar" and found this excellent hit, from 2008:

John McCain on Saturday Night Live

"$160 million to the Department of Defense for developing a device that can jam gaydar. Now, I don't know if this is anti-gay or pro-gay or if such a device would even work, but I do know this. Jamming gaydar is not a federal responsibility. That's something best left to state and local government."

Nobody tell the unsettled Mr. Rule.

johns said...

this machine is going to be wildly popular at parties. way better than beer pong.

Bob Ellison said...

I'm not reading the details here, but since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time?

Charles said, "That's not correct.... It can only assess the probability on a single picture, therefore its chances of being right are 50%..."

That's so wrong it's difficult to assess. It's not 50%. Althouse is correct that an automaton saying "straight" would get 90% (probably 97%) accuracy.

But Althouse is probably incorrect in assuming that the Stanford folks were so stupid as to not take this into account. Basic statistics analysis would handle the fact that most people are straight.

Stanford folks can be dense, but not that dense.

lgv said...

"this kind of software to violate people’s privacy or be abused for......"

This is the key issue with many kinds of software, in far worse ways that straight vs. gay. The concept of technology or AI used for "evil" purposes has always been the case. The moment I saw Gattica I was alarmed because I could see something similar happening in the future. When will gene mapping become mandatory and people with genetic pre-dispositions be treated differently?

As for this gaydar software, what's the big deal? At least here in the states. It would definitely be a great device for women who unwittingly marry gay men. A vast majority would have liked to have known beforehand. I'm also curious as to the error type of the 19%. Straight but labeled gay or gay, but identified as straight. Also, as mentioned by poster(s), this software does not support the genetic gayness theory as the variables are mixed.

Gabriel said...

@Bob Ellison:Althouse is correct that an automaton saying "straight" would get 90% (probably 97%) accuracy.

If you apply the analysis of false positives that I did above, but straight is 97% of the population, you get Ann's machine with 3% false positive rate, and the AI would be wrong 0.07% of the time, when it said "straight", and it would be wrong 88.4% of the time when it said "gay".

But the machine has not changed, it still has 81% accuracy. The AI is still better than Ann's machine at predicting straight or gay.

The Vault Dweller said...

If this AI ever gets implemented in a downloadable App for your cellphone this will instantly become part of a College drinking game. Probably various rounds of whoever can make the gayest face as determined by the App gets to make another person take a drink.

Anonymous said...

Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

Pedophilia, ianm, is highly correlated with having been sexually abused as a child. It becomes pretty damned "ingrained" as a sexual orientation. I'm told that there are some unenlightened busybodies, in light of that, "who care" about "how sexual orientation arises".

Again, the best solution is to stop making life difficult for gay people...

Yes, stop making life difficult for gay people. Just bake the damned cake. (And don't say anything less than worshipful on social media about dirtbags if they're gay.)

Bob Ellison said...

Gabriel, I bow to your superior math. Makes sense.

Bruce Hayden said...

"It's not the same percentage for men and women? That machine is sexist!"

Maybe because the divide between heterosexual and homosexual appears to be notably sharper for males. With females, there seems to be a lot more bisexuality. I do remember one night when I was flirting a bit with a woman in a club at another table. Winks, smiles, etc. she responded a bit, so I moved in, only to be possessively told by the other woman at the table that she was her date that night. Both were nicely dressed in cocktail dresses, looked very feminine, etc.

That said, I would suggest that the gay guys here in MT would probably easily pass as straight in much of the rest of the country, and, thus, maybe this study. Maybe a little bit in the mouth, and a little better grooming. But they still like their guns, hunting, pickup trucks, ATVs, snowmobiles, etc. Indeed, I got into a discussion with one of the right after the Orlando mass shooting at the gay nightclub, where we talked about guns, self-defense, etc, and he was adamant that concealed carry was THE solution to homophobia, and esp Muslim homophobia. And, yes, he had a Trump son gen out in front of his shop.

Fernandinande said...

"We obtained facial images from public profiles posted on a U.S. dating website. We recorded 130,741 images of 36,630 men and 170,360 images of 38,593 women between the ages of 18 and 40, who reported their location as the U.S. Gay and heterosexual people were represented in equal numbers. Their sexual orientation was established based on the gender of the partners that they were looking for (according to their profiles)."

Unknown said...

"Let's not actively discriminate against people based on sexual orientation — except in the formation of sexual relationships —"

Careful, you're dangerously close to being transphobic!

Birches said...

I would assume gay men would be posting pictures that appeal to gay men. Lesbians and heterosexuals do too. This isn't really ground breaking research. Smart people are so dumb sometimes.

gspencer said...

"Oh, but maybe in our formation of sexual relationships, we'd like a machine to double check whether our partners are actually sexually oriented to us."

The father character in Tootsie (played by Charles Durning) sure coulda used one of those.

CJinPA said...

Rise of the Hate Robots

Henry said...

But Althouse is probably incorrect in assuming that the Stanford folks were so stupid as to not take this into account. Basic statistics analysis would handle the fact that most people are straight.

The AI competed against human subjects:

Human judges performed much worse than the algorithm, accurately identifying orientation only 61% of the time for men and 54% for women.

If the test images actually matched population statistics, simply guessing "straight" all the time would be more accurate than the AI. But the researchers didn't do that. They culled the test images to have equal numbers of gay and straight:

The researchers’ program, details of which are soon to be published in the Journal of Personality and Social Psychology, relied on 130,741 images of 36,630 men and 170,360 images of 38,593 women downloaded from a popular American dating website, which makes its profiles public. Basic facial-detection technology was used to select all images which showed a single face of sufficient size and clarity to subject to analysis. This left 35,326 pictures of 14,776 people, with gay and straight, male and female, all represented evenly.

However, even if the AI is much better than a human being at identifying sexual orientation, there's still going to be a huge number of false positives, no matter what.

Ignorance is Bliss said...

Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

I imagine a lot of parents-to-be would care if it means that sexual orientation could be decided by, say, a slightly different mix of prenatal vitamins. Imagine the expecting mothers who are going to take prenatal vitamins, and will have to choose which formulation to buy- gay, straight, or "surprise me".

Of course there will be some activists who would specifically choose gay children, and some who would choose "surprise me" to show how accepting they are. I suspect the vast majority, even among those who are perfectly accepting of homosexuality, would choose straight. What parent would prefer their child to face extra harassment and discrimination? What parent wouldn't want their child to have the opportunity to raise children with their spouse where the children were the biological children of both parents?

Ann Althouse said...

"That's not correct. The machine is presented a single male picture and asked to identify whether the male is straight or gay strictly from the evidence of the picture. It has no capacity to judge what the respective populations are or whether the picture is randomly selected. It can only assess the probability on a single picture, therefore its chances of being right are 50% (like the boolean problem of flipping coins; what are the chances that the next coin toss will come up heads if you have already flipped ten heads? 50%)"

Seems to me it's like flipping a loaded coin.

Bob Ellison said...

This presents a classic problem in innumeracy. If you say "this magic machine can tell whether you're gay or straight with 81% accuracy, and normal humans are only 61% accurate", most people will assume that those are straight-up numbers, as in, those dumb humans get it right only a little more than half the time, and "Ann's Machine" would achieve 97% accuracy. That's not what's being measured here.

It doesn't worry me a hell of a lot. This will all be outlawed pretty soon.

Saint Croix said...

Althouse is correct that an automaton saying "straight" would get 90% (probably 97%) accuracy.

The 10% of the population is gay meme came from Kinsey, I think. Who did most of his research on people in prison. The idea that maybe some people go gay in prison didn't seem to have occurred to him.

Gabriel said...

@Ferdinande:. Gay and heterosexual people were represented in equal numbers.

Ho ho. Then on the population Stanford tested, it's 81% correct for both straights and gays and 19% wrong for both.

But apply the AI to the real population, and it will usually be wrong when it identifies gays. Still the same machine though. Conditional probability is so hard to develop a feeling for.

81% accurate means, "here is a man whose orientation is known, and the machine gets it right 81% of the time".

But using the machine on NEW people, is "here is a man who the machine has said is straight or gay". The probability that the machine gets is right, will be very different from 81%, unless men are close to 50% gay.

Henry said...

That last quote is from The Economist

Gabriel said...

@Ann:Seems to me it's like flipping a loaded coin.

That's a very good way to think of it. But your example is a COMPLETELY loaded coin, and it does worse than the AI on identifying a persons unknown orientation correctly. You can do the analysis using a loaded coin, and you will still find that the AI outperforms it.

Annie C. said...

"at least for the nonce?"

Reading Althouse is delightful. I got to that phrase and I got such a grin on my face.

Ray - SoCal said...

It not actually saying if your gay or not. What it's doing is if you have a picture, on a site where you are looking for a Gay date, then it can identify you most of the time. So it's a subset of the Gay population. This could include grooming and how you dress, and the background. Unless it's just head shot with a white background.

Bob Ellison said...

The real question is: can the blog comment filter tell that you're a dog?

Unattorney said...

There is no limit on how dense Stanford folks are. By the way,try assuming a gay rate of 3% which is probably closer to the truth. In that situation, the false positives far outnumber the true positives, since 19% of 97 straight guys would be identified as gay while 19 % of the three gay guys would be identified as straight, for every hundred guys tested. (Since only really stupid people don't know everyone is at least a little gay and a little straight, the whole study is useless.)

MadisonMan said...

How do the authors of this study know if the test subjects are straight or gay? Because they said so? Even if they are using pictures of straight/gay men/women -- how are the study authors certain about the test subject's sexuality?

It's like they assume some kind of binary either/or relationship.

Owen said...

John Deere post: what characterizes AI as exemplified in a weed-killing algorithm?
Stanford "gaydar" post: how can we use AI to sort our sexual selves?

Creepy. Speaking of which, I am now reading Nicholas Carr's "Utopia Is Creepy." A miscellany of cranky brilliant telegrams from the cultural collision with technology, especially social media.

buwaya said...

If the test was done correctly they would have had a very large proportion of "gay" subjects in order to properly test the "skill" of the system, to avoid a spurious result exactly as mentioned due to the small minority it is intended to distinguish.
This is not strange, all sorts of tests have been developed to test for conditions that are present in much smaller minorities - genetic tests for rare diseases for instance, where the negative result can be expected from 99.99% of the population.

n.n said...

The gendar will aid progressives seeking to increase [class] diversity.

Michael P said...

I should write papers like this. Imagine the implications on institutional racism if a computer can look at someone's picture and at least 74% of the time, identify whether the person is black!

walter said...

...just waiting for Titus to comment...

Roughcoat said...

Stanford has become a locus of bullshit.

tcrosse said...

I wonder how the AI Gaydar works on Stealth Gays, those who are deeply closeted, or those who are unaware or in deep denial that they're Gay ?

exiledonmainstreet, green-eyed devil said...

" Again, the best solution is to stop making life difficult for gay people"

Life is difficult for gays in 2017?

It is difficult for them in Iraq and Saudi Arabia and undoubtedly in certain neighborhoods in Dearborn where nobody will bake a wedding cake for a gay couple. Wander into the wrong street in Brussels or Paris holding hands with your partner and life might become very difficult for you very quickly.

In the civilized parts of the Western World, they generally have it pretty good.

exiledonmainstreet, green-eyed devil said...

tcrosse said...
I wonder how the AI Gaydar works on Stealth Gays, those who are deeply closeted, or those who are unaware or in deep denial that they're Gay ?"

I've known a few of those. They might have denied they were gay, but everyone else knew it, and you didn't need particularly strong Gaydar to figure it out.

Abdul Abulbul Amir said...

Poor quality app. Its percentage correct would go up if it guessed straight all of the time.

Bandit said...

My dog can identify sexual preference 100% of the time with just a simple snoot exam

Gabriel said...

@Abdul Abulbul Amir:Poor quality app. Its percentage correct would go up if it guessed straight all of the time.

No, it wouldn't, multiple comments have addressed this and worked it out in detail.

Phil 314 said...

As alluded to above I'm trying to get the discrimination angle in western society. I see it being used to identify gay men to hire (I.e. "Diversity!")

Henry said...

I wonder what AI can tell from handwriting analysis.

buwaya said...

"Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is? "

This gets down to some very fundamental matters, of the "why are we here?" and "what is my purpose in life?" sort.

If we assume that we are each autonomous, unburdened entities with no obligations to others however defined, parents, family, clan, nation, religion, ancestors, humanity in general, then this position is defensible. This is however a very modern idea and historically speaking, in almost any human culture, it is heretical.

If however one considers each human individual as existing as a unit in a theological-social-biological system, then it matters a great deal indeed. Reproduction is a fundamental human obligation, and it matters a great deal if a default in this great duty comes from an unavoidable biological defect, infertility as an act of God, or if it is caused by an injury of some sort, whether biological or psychological. If caused, there is someone at fault, even if it is the victim himself.

Henry said...

In Japan, blood type is supposed to indicate personality. From Wikipedia:

discussion of blood types is widely popular in women's magazines as a way of gauging relationship compatibility with a potential or current partner. Morning television shows feature blood type horoscopes, and similar horoscopes are published daily in newspapers. The blood types of celebrities are listed in their infoboxes on Japanese Wikipedia.

tcrosse said...

I wonder what AI can tell from handwriting analysis.

It looks for i's dotted with smiley faces or little hearts.

buwaya said...

To be clear, statistically speaking the default in the duty of reproduction in the modern world is, in the main, not due to being homosexual.

Feminism, modern social conditioning and offical/semi-official brainwashing, the collapse of the modern family, and the prevalence of certain forms of entertainment such as video games and pornography, have much more to do with it. Homosexuality, to the degree that it is NOT biological, is just one case among those others.

But at root it comes down to this question, whether human beings are social creatures or not, whether we are born "free" or burdened.

And this question of burdens is not a "liberal" POV. Traditional conservatism of every flavor assumed "burdens", much more so than any modern liberal or progressive ideology.

Modernity assumes "free", biology seems to say "burdened", because "free" means, ultimately, extinction.

Ralph L said...

the gay guys here in MT would probably easily pass as straight in much of the rest of the country.... Maybe a little bit in the mouth,
All sorts of dirty comments possible here.

sparrow said...

Ordinarily an AI false positive rate is based on a test sample held out of the the training data that was used to develop the machine. If they included the test data in the training set they've biased the answers. They may have used a more sophisticated method like 10 fold cross validation but ultimately AI methods are best tested by entirely new data sets. I wonder what effect drawing from a larger age range or others factors would have on accuracy.

An aside though, every device ever created that could be adapted was applied to positive and negative ends. AI is no different.

buwaya said...

An aside, if the accuracy of this system is as stated it would be a rather good and useful manufacturing quality control test for many processes.

Hammond X. Gritzkofe said...

"...or be abused for anti-LGBT purposes."

But abused for pro-LGBT purposes? Impossible!

Todd said...

lgv said...

The moment I saw Gattica I was alarmed because I could see something similar happening in the future. When will gene mapping become mandatory and people with genetic pre-dispositions be treated differently?

9/8/17, 11:08 AM


I have said half jokingly that all of the recent marketing for personal gene testing for the general public (ancestry, 23andme, etc.) is really a devious plot by the government/industrial/insurance complex to gather gene information on the population. Either because they are looking for something specific or they want to use the information to make decisions about individual people (thinking insurance). If they just came out and said "we want everyone to get tested" the ACLU, etc. would be all up in arms BUT instead they figured out a way to get many people to not only sign up to have it done but to pay for it themselves too! What a great, evil genius plan!

Todd said...

The Vault Dweller said...
If this AI ever gets implemented in a downloadable App for your cellphone this will instantly become part of a College drinking game. Probably various rounds of whoever can make the gayest face as determined by the App gets to make another person take a drink.

9/8/17, 11:10 AM


Or take something else?

I know, I know, I should NOT have gone there...

Bay Area Guy said...

You don't need algorithms, Boolean logic or artificial tntelligence to determine if a man is gay. Just ask him if likes to suck dick. If the answer is in the affirmative, well, odds are he's gay.


Martin said...

Interesting that the article fears such tech could be used for anti-LGBT discrimination when everything happening in the culture today suggests that if it were used to discriminate it would be used for PRO-LGBT purposes.

Tim Cook told me so.

But in either case, it better be 6-sigma accurate to be used for anything at all. This isn't like determining whether X would respond favorably to a pop-up ad from Chevrolet, and how much to charge Chevy for that ad.

Owen said...

buwaya: you always raise the tone in these discussions. Thanks.
sparrow: great points about biasing the training/test data. We are under a strange spell of credulity when we are told a "computer model" is involved. "GIGO" was never more true than now, and a nose for BS was never more prized.

Owen said...

Todd: your comment about people paying for/volunteering information of an ABSOLUTELY personal nature --what could be more individuate and specific than your own genome?-- seems like a home run. Depressing.

Sebastian said...

Huh, 81%. They call that AI?

Next, they'll tell us that some devious software can correctly figure out people's actual intelligence, based on a single test.

David Docetad said...

Few are the subjects that really smart people can get totally wrong more than probability. Even mathematicians.

David Docetad said...

For those doubting Charles, consider the following scenarios:

The machine is placed in a gay bar and used to analyse the patrons. Next to it is Ann's automaton. Which one is more accurate?

Todd said...

David Docetad said...
Few are the subjects that really smart people can get totally wrong more than probability. Even mathematicians.

9/8/17, 2:06 PM


I read once that 83.95% of all statistics are made up and that the probability of any specific statistical analysis coming up with the same answer under reanalysis is 27.15%!

Big Mike said...

In this sort of AI a lot depends on the training set -- the examples used to train the algorithm. So if all the pictures in the training set of men with blonde hair but dark roots are gay, then it will create a rule that says dark roots only if gay. If every bald man in the training set is straight and no bald men are gay, it will create another rule that says bald only if straight. This a straight man with dark roots or a bald gay man will be mis-assigned.

The trouble is that this is Stanford, and they should know about training set artifacts

Ralph L said...

no bald men are gay
That's true. They're banished from Gaydom.

HoodlumDoodlum said...

Ann Althouse said...Let's not actively discriminate against people based on sexual orientation — except in the formation of sexual relationships

That's homophobic and transphobic. Anyone who believes it's ok to discriminate on the basis of biological sex or sexual orientation is an ugly bigot. I'm saddened to hear you're one of them, Professor Althouse.

Think I'm kidding? That's the new line--if you're a guy and you don't want to date someone who self-identifies as a woman but who has a penis, then you're a bigot, a transphone, and a bad person.

Ann Althouse said...

"tcrosse said...
I wonder how the AI Gaydar works on Stealth Gays, those who are deeply closeted, or those who are unaware or in deep denial that they're Gay ? "

That's a question I was going to raise.

exiledonmainstreet, green-eyed devil said...

Ralph L said...
no bald men are gay
That's true. They're banished from Gaydom.

9/8/17, 2:41 PM

Is it OK if they shave their heads completely rather than have the Friar Tuck fringe or the awful comb-over?

The shaved head look looks better on blacks than on whites, although a man I know who does it says it makes hats and caps an absolute necessity, especially during Wisconsin winters. I would guess a gay guy would have to be careful about the sort of hats and caps he wears. My guess is that those north woods plaid things with the ear flaps wouldn't work.

Hammond X. Gritzkofe said...

OMG! I just got the results back, and I'm queer. I mean gay. Shouldn't use that word now that I'm one of them. I mean one of us.

How could I have kept this secret from myself all these years?

Pinandpuller said...

81% of the time it works every time.

Matt Sablan said...

"That's not correct. The machine is presented a single male picture and asked to identify whether the male is straight or gay strictly from the evidence of the picture. It has no capacity to judge what the respective populations are or whether the picture is randomly selected. It can only assess the probability on a single picture, therefore its chances of being right are 50% (like the boolean problem of flipping coins; what are the chances that the next coin toss will come up heads if you have already flipped ten heads? 50%)

For a single picture, the AI system is correctly identifying the individual's orientation 81% of the time (91% of the time when shown five pictures of the same individual) instead of the expected 50% correct."

-- Someone's sexuality isn't the same as flipping a fair coin though. For example, with the fair coin, if the Machine guessed randomly, it would be about as right as simply always guessing heads.

With guessing if Bob is gay, if the machine picks it at random, it is significantly more likely to be wrong than right if it randomly picks Gay (assuming that the 5%-10% population figures are correct.) If the machine changes behavior to always pick straight, it becomes significantly more correct than if the machine above chose to always pick heads.

Bob Ellison said...

Why is AI Gore so concerned about sexual orientation?

Gahrie said...

What if I am not gay, but simply fabulous?

n.n said...

Transgender/homosexuals are characterized primarily by mental cross-gender features, which would make distinguishing between normal and transgendered individuals using static visible artifacts a hard and possibly wicked problem. Still, for the sake of [class] diversity... and selective-child doctrine to recycle unwanted, inconvenient, or profitable babies.

HoodlumDoodlum said...

Ann Althouse said...Who cares if sexual orientation arises "biologically" or through some other process that nevertheless becomes ingrained in the biological entity that each of us is?

Who cares? Certainly the Baby I Was Born This Way side cares. Certainly the Pray the Gay Away side cares. Everyone cares. People agitating for cultural acceptance of homosexuality for years based their appeal around the idea that homosexuality was not something that is subject to an individual's choice--that it was an identity they had no control over.

It's widely agreed upon that it's wrong to discriminate against people on the basis of things they cannot change. It's only slightly less widely agreed upon that it's ok to discriminate against people on the basis of things they can change. Whether some trait is immutable or not matters a great deal! If a trait "becomes ingrained" then it might "become un-ingrained." That matters.

White nationalism is wrong and bad because being a member of that identity group involves having or not having a trait over which an individual has no control. The Black Power movement makes a similar appeal to racial identity but it's not wrong because you're not required to have an immutable trait to be a member of that identity group. That's the line, right? Some forms of identity politics are unacceptable because they require possession or non-possession of an immutable trait--something one cannot choose. Other forms of identity politics are acceptable because they do not have that requirement. I think that's the accepted PC line.

Aside:
There's a small group of people now following the transexual activists' lead and pushing for acceptance of "transableism." The transabled are people who don't have a given physical handicap but who IDENTIFY as someone who does--or who wants to sometimes so identify. That, so far, is not widely accepted. Why not? The logic of "I'm a woman if I self identify as a woman" holds for "I need a wheelchair and deserve the accommodations society gives to people in wheelchairs if I think I do." But there's some non-logical emotional difference; people seem to feel that one case is acceptable and one case is not.

HoodlumDoodlum said...

Back on topic:
The argument used to be "it's wrong to discriminate against gay people because being gay isn't a choice. You can't make yourself gay and you can't make yourself straight--it's not a choice!" The argument was that homosexuals deserved more consideration than other interest groups--greater legal protection as a class, etc--precisely because they had no say in that portion of their identity.

If that's not true then why should we have more compunction about discriminating against homosexuals than we do against any other self-chosen group? Lets say I don't much care for Yankee fans and I made it a policy not to hire people who were Yankee fans. The law would have no way to stop me. Being a Yankee fan is a matter of choice--maybe you live in NY and your whole family were Yankee fans and you didn't really choose being a fan for yourself (it was "biologically ingrained" on you through repeated trips to the ball park, hours in front of the TV cheering w/your family, etc) and so on, but we all agree that you remain a Yankee fan simply by choice. You could choose not to be a Yankee fan! Most people would say that my strong preference for not hiring Yankee fans is a little weird, but not illegal. You can't say the same if my preference was against a person of a given skin color.

But wait! Religion! Religious belief and creed are given the same kind of protection as sexual orientation in the law. Since one can choose one's religion that must mean choice doesn't matter. I'll give 2 responses to that 1.) many of the same people arguing that homosexuality must be a protected class argue against religious freedom/that religious belief and creed should be treated traits and argue that what they define as harmful beliefs (whether religious in nature or not) should never be protected (or even allowed/tolerated) 2.) we purposefully treat religious belief and profession of creed as something greater and more profound than an individual's choice in may other ways so it's not unusual to treat it differently here. The latter response is admittedly weak (almost to the point of question begging) but I don't think it's unusual for someone to assert that their faith isn't something they consciously chose.

At any rate it seems a bit unsporting to argue for a few decades that something must be treated a given way primarily because that thing is wholly beyond the individual's control only to turn around (having gained the sought-after widespread acceptance) and say "it doesn't really matter if the trait is within the individuals control or not!"

Curious George said...

Apparently 74% of lesbians wear flannel shirts, work boots, and have buzz cuts.

Curious George said...

Aaron Rodgers call your agent.

HoodlumDoodlum said...

Ann Althouse said...Seems to me it's like flipping a loaded coin.

Feels to you like it's biased, you mean.

HoodlumDoodlum said...

I dunno, I guess the fear is that it's possible to tell a difference when it's not polite to admit there may be any visible/detectable difference at all--is that it?
We're supposed to believe that everyone's the same and there's no way to tell anyone apart, but if some machine can do a decent job of telling us apart then we have to drop the polite fiction that everyone's the same? Is that the deeper problem?

If the machine were just as accurate at predicting a person's political leanings based on a picture would anyone have the same objection to it and it's possible nefarious uses? Would the response then be "Let's not actively discriminate against people based on political beliefs and then the machine won't have any relevant function?"

Gabriel said...
This comment has been removed by the author.
Gabriel said...

@David Docetad:For those doubting Charles, consider the following scenarios:

The machine is placed in a gay bar and used to analyse the patrons. Next to it is Ann's automaton. Which one is more accurate?


Already worked out above. The AI is more accurate for both gays and straights, than Ann's automaton is.

Assuming a 97% of the population is straight, Ann's machine is wrong 3% of the time for straights and 100% of the time for gays.

The 81% accurate AI is wrong 0.7% of the time for straights and 88.4% of the time for gays.

3% > 0.7%. 100% > 88.4%. AI beats Ann's machine.

Bob Ellison said...

"The machine is placed in a gay bar..."

Guy at bar: "What's that for?"

Bartender: "It tells you who's gay."

Guy: "EVERYONE here is gay!"

Bartender:

Hari said...

It would be far more valuable to know what percentage of straight people the algorithm identifies correctly and what percentage of gays it identifies correctly.

For example, if it was given 90 straight pictures and identified 81 correctly and was given 10 gay pictures and identified zero correctly, that would be significantly different is it identified 71 of 90 straights correctly and identified 10 out of 10 gays correctly.

I'm sure this information is in the actual study.

Freeman Hunt said...
This comment has been removed by the author.
Freeman Hunt said...
This comment has been removed by the author.
Michael K said...

Just think of how valuable it will be in interior decorating.

JHapp said...

I don't care if someone is gay and I really don't understand why anyone else would. What a waste of time and money.

Jupiter said...

Ignorance is Bliss said...

"I imagine a lot of parents-to-be would care if it means that sexual orientation could be decided by, say, a slightly different mix of prenatal vitamins."

Yeah, right, vitamins. Imagine if we discovered that Downs syndrome was genetic, and there was an early-pregnancy test for it.

Jupiter said...

If you could detect homosexuality in the first trimester, only Christians would have homosexual children.

tim in vermont said...

Pretty girls can tell who's gay pretty easily. If a guy doesn't check her out, no matter how subtly, he's probably gay. I doubt it would work for hot guys and lesbians though, because lesbians can often be tempted.

Anonymous said...

Ann writes:

since the great majority of people are straight, wouldn't a machine that just guessed everyone is straight be right more than 90% of the time? Seems to me the machine it terrible at its work.

1: It's more like 3% are gay
2: From the paper:

Facial images. We obtained facial images from public profiles posted on a U.S. dating website. We recorded 130,741 images of 36,630 men and 170,360 images of 38,593 women between the ages of 18 and 40, who reported their location as the U.S. Gay and heterosexual people were represented in equal numbers. Their sexual orientation was established based on the gender of the partners that they were looking for (according to their profiles)

(See Table 1 for numbers)

Here's the part where they demonstrated basic competence:

To prevent overfitting, we used a 20-fold cross-validation when estimating the predictions. The sample was split into 20 subsamples; one of the subsamples (test set) was put aside, while the remaining 19 subsamples (training sets) were used to train the prediction model.

So their training data, and their test data, was both 50% gay and 50% straight. So if you just picking "straight" every time, you'd be 50% correct.

Final bit:
Given a single facial image, a classifier could correctly distinguish between gay and heterosexual men in 81% of cases, and in 71% of cases for women. Human judges achieved much lower accuracy: 61% for men and 54% for women

So human gaydar sucks for women (essentially chance), and mostly sucks for men, the computer actually got decent results.

FIDO said...

This seems to be sloppy thinking on the part of Ms. Althouse.

Why care if it is biological or can be taught? Well, with this whole tranny movement there are significant risks to the psychology of our children if some militant Lefty teacher starts gender indoctrination.

And if we can prove it is biological, it cuts the feet out from under many religious people in their bias.

However, unfortunately I don't think life is that easy. I think it is both.

madAsHell said...

My PC has gaydar.

walter said...

Jupiter said...If you could detect homosexuality in the first trimester, only Christians would have homosexual children.
--
Ka-boom...that's pretty potent at a level of tolerance and inclusion. But then, what would (a) God be up to doin' that?

walter said...

Blogger FIDO said...This seems to be sloppy thinking on the part of Ms. Althouse. Why care if it is biological or can be taught?
--
Interesting..have we transcended "nature vs nurture"?

TomHynes said...

If you click through to the Economist link, the study used "35,326 pictures of 14,776 people, with gay and straight, male and female, all represented evenly." Therefore, if you always guessed "straight" you would be right only 50% of the time.

Snark said...

As Tom Hynes notes, chance is 50% in the population sample they used. This is of course different from the general population, where guessing straight would have a statistical advantage. The interesting thing to me in the data was the degree to which the machine outperformed the humans.

iowan2 said...

I actual enjoy my work. Its sales and service, in agriculture, so lots of transfer of information, helping others attain their goals. I Retired, when I found the same gig as a consultant, so same 'job' less stress. My FIL was in building trades. He was good at it, could do anything. Spent his free time, building stuff. Most peoples sense of worth, comes from being of service to others. Robots my take some of the monotony out of that, but the 'work' still has to be done.

stlcdr said...

Doesn't this validate the statement 'so-and-so looks gay'? Or is ok for a computer to say it but not a human?

If a robot decides to not serve someone because they are gay, is that ok? What about if a robot bouncer at a gay bar does not let in someone because they look straight?

(Ironically, I checked the 'I'm not a robot' checkbox...I think robots would be able to check the box if it said 'I am not a robot').

Gabriel said...

Not sure anyone else still cares, but Ann's question inspired to me work out if there was ever an AI that performs so badly that a guessing strategy would outperform it. And here is what I learned:

1) The difference between guessing and a real test is that the real test is correlated with what the person's orientation is, however weakly, but guessing is statistically independent of the person's orientation. This makes the probability equations take fundamentally different forms.

2) There are only three guessing strategies that make any sense to pursue:

a) Always guess positive, if the population is known to be more than 50% positive. The accuracy will be whatever the prevalence of positive in the population is.

b) Always guess negative, if the population is known to be less than 50% positive. The accuracy will be 100% - whatever the prevalence of positive in the population is.

c) Guess positive 50% of the time, if you have no idea what the prevalence of positive in the population is. This is the most likely real-world scenario; you're not testing everybody, or even necessarily a random sample of the general population. (Imagine that you are trying to determine if a bar is a gay bar by testing their patrons, for example.) In this case your accuracy is 50% regardless of the prevalence of positive in the population.

3) False positive and false negative rates are identical for all guessing strategies; the false negative rate is the prevalence of positive in the population, and the false positive rate is 100% minus that.

4) The only test that does worse than guessing, for a population where the prevalence is totally unknown, is a test that is less than 50% accurate. If the test is greater than 50% accurate, by however little, its false positive rate and its false negative rate will be lower than in any guessing strategy whatever. Its accuracy might not be higher than choosing the "always positive" or "always negative" strategy, in a population with a known prevalence, but it will still have lower rates of false positives and false negatives.