This story is just a lot of hype for nothing. Gay men are less than 5% of the population. Even if this software were 91% accurate in identifying gay men as gay, and moreover, 91% accurate identifying straight men as straight (this was not stated in the article), the odds that the software correctly identifies a random man who could be straight or gay would still be about 50%. You are better off just flipping a coin.
3
Given that gay men utilize views of other men's faces when cruising, there might very well be physical features that a machine system can pick up. Straight people can also notice some of the same features that gay men pick up on. Gaydar also depends on responses responses from other faces. A program that captured interaction between faces might be more accurate--a far more complicated problem, however. "How" we recognize each other isn't a tribal secret.
A state or a corporation with lots of data to mine would find many clues about many features of individual behavior, besides the sexual. A subscriber to only The Nation is probably more liberal than a subscriber to only TIME. A subscriber to the Advocate is probably gayer than a subscriber to the National Review.
2
The article's paragraph that contrasts false negatives with false positives begins:
The software would also mistake 9 percent of gay people as straight people.
Instead of repeating the false positive statistic, it should give the number of false negatives:
The result: Of 50 gay people in the population of 1000, the facial scan would identify about 5 as straight.
2
There are two sides to this (at least). On the one hand, it can help prove that sexual orientation is innate, and not chosen as still believed by most of the world. On the other, it can be used in the 76 countries where being gay is illegal, to detect gay people, and either kill them, or otherwise deprive them of civil rights.
9
This is another instance of "political correctness" getting in the way of evidence-based academic research. NYU's Janathan Haidt has been calling attention to the lopsided irrational ideology on campuses that precludes honest intellectual battle of ideas. Remember how the research of STEM cells was held hostage by zealous religious fanatics? Extremes exist on both left and right side of the cultural divide.
5
Another reason to be concerned about the ubiquity of Facebook.
4
Michal Kosinski and Yilun Wang look like a nice couple. Hopefully their next potential employer or landlord won't hold that against them if he or she thinks they're gay. Especially in Trump's America, where already sparse legal protections against sexual discrimination and sexual minorities are being chipped away - including "legal" discrimination by religious employers and institutions. This unethical misuse of science (or more accurately pseudo-science) could literally be a matter of life or death in certain countries or cultures, an excuse for exclusion, harassment, punishment or even murder. We must insist that we - our images, our DNA, our likes and dislikes and habits, our relationships, our internet history, our medical history, every bit of our being - are not playthings to be dissected, bartered, violated or exploited without our informed consent. These attributes belong to us, they are us - and we have a right to privacy, the right to be left alone, if we so choose. Witness the recent Equifax debacle, which highlights how little control we have in such matters,
and the fact that legal restraints and remedies to protect our privacy are sorely needed. Past & present tribes and cultures have tried to combat the theft of their images, selfhood and dignity not by arguing invasion of privacy but by maintaining that it's a theft of their spirit and essence - a stealing of their souls and persona. Same concept, different terminology. And needed now - more than ever.
8
Beautifully said.
1
The invasion of privacy is repulsive and it is hard to see how it serves a legitimate purpose. In the meantime, like anything, it can be used for good or evil, including for early noninvasive disease detection. Can you imagine early ovarian cancer being detected this way? In the meantime, can you trust those who have this capability to only go after what is good? The real question to me is: how can you insure that this technique is not abused?
5
I wonder how many of these people commenting on this study actually read the paper. This is the last line of the abstract,
"Additionally, given that companies and governments are increasingly using computer vision algorithms to detect people's intimate traits, our findings expose a thread to the privacy and safety of gay men and women."
It takes a dozen lines of Python code or less to put together a CNN model that you can then train to do this, it isn't hard to do.
In response to the American vs. Russian tank parable, further down in the paper, the researchers state that "facial features are extracted using a DNN that was specifically developed to focus on non-transient facial features, disregarding the head's orientation and the background."
As far as "Gay people... tend to post higher-quality photos", one normally scales all images to a specific size before feeding to a neural network, for training and for testing. So even if they took a selfie with the Canon 5DS, it must be scaled down before being fed to the network.
If Yemen officials want this, the technology is literally available on GitHub for anyone to grab and perfect. These people simply carried out an experiment with the technology and see if it would yield any note-worthy results.
I have to say that I don't agree with their conclusion though. To say that we can detect sexual orientation from faces is to undermine the value, complexities, and experiences of human beings.
14
When Abraham Lincoln was chastised by a cabinet member for saying he di not like a particular General's face, Lincoln retorted "everyone over 40 is responsible for his face'
6
But by then it's too late.
2
This raises the crucial question of whether it is possible for government investigators, by using facial recognition techniques, to tell when a person is very likely to be a political dissident, or even likely to become one. For example, what if the photo of the face of a person is taken at a moment when he or she is reading a news story by the MSM which supports an establishment lie, as they so often do, shows a deep disgust? This might be interpreted as a sign the reader is a right wing partisan or a left wing politically aware activist interested in the truth? Conversely, if the person seems to be pleased by the story this could be interpreted to indicate the person is politically uninformed or misinformed such as your typical Hilbot?
As it is now we have government intelligence agencies trying to intimidate the American people into censoring themselves with stories of their great illegal surveillance capabilities. What would it be like if they were able to focus their efforts to intimidate on much fewer people?
3
You point out a great danger.
The only challenge is capturing the image at the right moment, though with a compromised PC and some software that knows what's on the screen and detects eye movement (so it verifies what's actually being looked at) it's easily doable.
1
It is possible people consciously on unconsciously exaggerate their sexuality on dating sites. I would be more convinced if they could identify sexuality from DMV photos.
Also if the observed morphological differences proves true this suggest Homosexuality is genetic and not a life style choice as the Religious Fundamentalist claim. I can't think of any genetic discrimination secular or religious that is legal under our laws.
6
I can't even identify myself from my DMV photo ;-)
But although your point is well taken, it's pretty clear discrimination persists irrespective of legality. And in the hands of a less-than-ethical government legality isn't very relevant. You'll merely vanish.
5
Problem one: Sample is not random.
Problem two: Sample is taken from profiles where men are trying to look attractive to other men.
Problem three: Sample includes bisexual men and straight identifying men. Yes, there are straight men on Grindr etc. Unaccepting of themselves, perhaps. Looking for a good time, definitely. Meeting the sexual needs of non-reciprocating guys has a long history.
Problem four: not all profiles have pictures.
Problem five: the telling is more in what the face looking at.
5
Why is it that the NYT always froths at the mouth and is the first to publicize any questionable study that pigeonholes and targets the LGBTQ community? A few years ago they apologized for publicizing an erroneous study about male bisexuality ("Straight, Gay, or Lying?"). But now they're happy to share a study that the authors gleefully say uses "off the shelf" technology, but yet can be used by any homophobic bully (or worse) who doesn't like the shape of your jawline.
2
I don't know how reliable this technology really is, but if it does as well as the article claims, it is pretty good. If it really works, it is an interesting insight into human sexuality. But one thing I am sure of, if a bully wants to attack gays, he doesn't need this software. Bullies were attacking gays before Alan Turing was born.
If there is any cultural relevance to this research it is this message: Your sexuality is as much a personal preference as your jaw line.
2
Matt - "Why is it that..."
Agenda!
1
The various researchers' justifications for their research is questionable. Basically the beg off on the the ethics of it and so it is just science. Yes so was Phrenology in the 1880s which started as science and was quickly used for all kinds of deeply racist practices.
Even more questionable is that there is an algorithm that can be so complex as to handle every nuance of a person's face. And lighting and shadow in a photo, etc. Bogus.
3
Oh, don't fool yourself. Facial recognition is here already, and it's going to get much better in the near future. This technology won't even be necessary when systems will know your many identities, mostly offered up for free via your online activity.
3
If it's possible to make, someone will make it (e.g. thermonuclear weapons.)
If someone makes it, someone will use it (again, thermonuclear weapons.)
If it is thought effective, even if it's not, someone will make it and someone will use it (e.g. the polygraph, which does not distinguish fear of being found out from fear of being falsely accused).
A detector that is partially effective will produce false positives. One is one too many; 15 million is 15 million too many.
To use any device to discriminate unethically is unethical, no matter what its efficacy.
To warn about the unethical use of technology is to encourage those who want to want to use it unethically (as people say, it gives them ideas).
Not to warn against these dangers makes them harder to defend against once the scheme is underway.
What are good people to do?
7
Suppose this machine is developed and marketed and an order comes in from a country that stones homosexuals to death such as Yemen. Will the manufacturer sell the machine to Yemen knowing it will be used to determine who will be buried in the ground up to their neck and have stones thrown at their head till they die? How will these two scientists feel if this happens? How will anyone feel who approved of this or published this paper?
3
You miss their point: this kind of software exists today, and is trivially simple to harness.
They're trying to warn society. They're the good guys.
13
Not entirely convinced by explanations that it was done as a warning of the potential over-sell of this type of machine "learning." Have never heard of Amazon Mechanical Turk before. Need to hear more about this reference.
Google it...
I worked in tech until a couple of years ago.
Machine learning is both real and highly effective. It's used today in numerous applications.
2
A poor distinction is being made between what is personal and what is private. I have high melanin skin and kinky hair and I was born a baby girl and am now a woman. None of intrinsic qualities are "private". If sexual attraction is determined in the womb, it is not private, it is just personal.
Black people get discriminated against in the U.S. . But if I were to say "My race is private, don't ask me about it." people would, justifiably, look at me as if I was talking nonsense because I am so obviously Black. But even low melanin AA people "passing" in the past were not saying "My race is private." They were straight out lying and pretending to be White people.
People get discriminated against based on age. That does not make age "private". With the technology available today it is easy enough to tell, within a 2 decade span, how old someone is. If someone age 75 was asked "Would you like to show your ID and take advantage of our senior citizen discount?" by a store clerk and that someone responded "Whether or not I am a senior citizen is private and it is a privacy violation for you to ask." that person would be, justifiably, looked at as if he/she was talking nonsense.
If LGBT people are going to claim their equality struggle is a civil rights issue and their sexual orientation is not a choice but an intrinsic quality, that's fine. But they can't also argue that it is "private" any more than I can argue that my Blackness or my being a woman is "private" information.
23
What? The core question is not "private versus public information." That is a tertiary issue at best. The problem is that the "science" behind this work is not accurate enough, or perhaps even true. It would be like the USA govt telling someone who is of trackable african descent that she/he is not black at all because one scientist from Stanford said that person is unquestionably black. Sometimes science is not correct.
"If LGBT people are going to claim their equality struggle is a civil rights issue"
And sadly THAT is the sentence where you reveal your real sentiments and politics.
8
Please avoid statements like this one:
"A facial scan that is 91 percent accurate would misidentify 9 percent of straight people as gay; in the example above, that’s 85 people."
A screening device has TWO accuracy numbers. There's the probability of correctly classifying a straight person as straight. There's also the probability of correctly classifying a gay person as gay. These two accuracy numbers (called also sensitivity and specificity) do not have to be the same; in fact, they might not even be close.
Daniel Kahneman and Amos Tversky devoted their careers to cleaning up confusions like this one.
22
When I first read about this study I was deeply distressed. I wonder what a computer would have said about my face then and I'm still distressed by this endeavor, which seems a form of targeting, of bully-empowerment.
Useful was the statement that gay people tend to post better-quality photos. And yet, as I wander around the world as a late middle age homosexual male, what my brain seems to isolate in trying to identify other gay men is -- that gay men's faces often seem to have a tiny bit more detail to them, slightly more defined cheekbones, eye sockets, chins -- as if the Creator, a heavenly Michelangelo, had gone back to work on our faces one last time with a velvet hammer and satin chisel. Does this make sense to anyone else?
1
I have two theories as to why liking curly fries could correlate with intelligence:
First, they may be more expensive or sold in more expensive restaurants. There's probably some correlation between income and intelligence.
Second, randomness? If you test enough factors there will be correlation just by chance. Sort of like if one sports league wins the championship the stock market goes up, etc.
Shaun Eli Breidbart
www.BrainChampagne.com
4
I've lived in Boston, New York, L.A. & San Francisco almost my entire life. I'm an "elite," literary college professor. I've never seen a curly fry on any menu outside of street fairs and carnivals in that same life. The curly fry thing is the great mystery here.
3
You've never seen one? I've never even heard of one (or them). What, if anything, are they?
2
There is a fundamental problem with this entire study. It does not actually detect who is gay. It only documents that people post different pictures depending on whether they are looking for same-sex or opposite-sex dates. Is anyone surprised by that? Also, they study was a head to head comparison, "which is these two photos is 'more gay'". That is a huge bias and advantage for the program. Until they do this with photos that are not intentionally selected to convey sexual orientation (i.e. not dating pictures) there is no news here beyond the hype created.
9
This is true. It does not detect who is gay it looks a probabilities of who is more likely gay in pairs of one straight and one gay dating profile image. Something I found surprising and interesting that I could not fit in the story was that Nicholas Rule, who does a lot of studies of "gaydar" has found that gay people actually tend to look less gay and straight men tend to look less straight in dating profile photos than other photos. This is what he told me: "When i first started doing this work, journal editors and reviewers were saying of course gay men look extra gay or straight and when we started coding, we found that gay men want to seem masculine, straight men want to seem extra caring." That is not based on same sample and this is referring to more than faces, but interesting point.
2
"It does not actually detect who is gay."
If people think it detects who is gay, some of them will use it to "detect" who is gay. Bad news for everyone, except for those who would make money off the algorithm.
That, I believe, is the purpose of these computer scientists - to send a warning about electronic snooping.
Think of the polygraph. Not good at detecting lies, and bad for everyone except those who profit from its use as a lie detector.
In the process of warning about data mining, these electronic whizzes have sent another warning to those willing to hear - concerning the dangers of knee-jerk judgment. That too can do harm, and when it's combined with the desire to snoop, tyrannize, discriminate, or even influence policy - look out!
3
Please spend your youthful minds & tech energy on curing cancer already. We have enough ways to discriminate w 100% accuracy as it is.
Jeez!
25
I think you're missing the point and motivation of this research as described at the beginning of the article. The researchers are concerned about the possible uses of facial recognition algorithms and deliberately chose the most controversial and arguably alarming application they could think of in order to publicize these concerns.
17
I agree with Harvey Wachtel.
But in any case, CJ, what does it matter if the accuracy is 100%? In many situations (being black, being gay) there is always someone who can "pass." The evil is that the discrimination goes on.
BTW, I am miriam from Astoria, Queens. I don't know how my location was changed to (br/), but I hope it can be changed back.
1
No, the question is not "can you live with yourself if you knew it’s possible and you didn’t let anyone know?” What a bogus and cowardly statement. The question is why in the world is this necessary; what good does this do for society? It was only as recently as 30 years ago that homosexuality was removed from the DSM as a disorder. Decades of sham academic "research", like Dr. Kosinki's, hurt generations of people. Here we are again, but now we're fighting against AI stereotyping.
Shame on Stanford
8
What if a sonogram could determine LGBT characteristics leading to abortion spike? No more/less ethical than Amniocentesis?
2
In neither the original article nor this one, does anyone address how or if the researchers determined if the subjects used their own photos. Notoriously, online daters use fake ID in their profiles.
Noting this statistical anomaly in their research would further the point that trusting this type of pseudo-science in judging people is less than intellectually sound.
Turn this to more useful prospects. Take photos of Limbaugh, Ailes, Weinstein, O'Reilly, Cosby 45 and contrast with non-sexual predators.
45
Isn't this a little too binary for the complicated, fluid world we live in? When is one's sexuality (or gender) set exactly? Age five? fifteen? fifty-five? This also sounds very race-based. I bet the algorithm works really well on white men in Silicon Valley.
4
It's amusing beyond laughter to see "a threat to the privacy of LGBTQ and non-LGBTQ people alike" dismissed as "junk science".
Junk science claims to tell you things that are factually not true.
This study shows what anyone familiar with the digital ecology and the power of "trained" computer algorithms will tell you up front, without apology or qualification: computers can find out a lot about you, whether you like it or not.
I understand the parochial interests of narrow advocacy groups to champion their base. But the larger question affecting us all is: are we all really ready to become digital products -- targets for digital advertising, case files in digital surveillance, chits in the digital economy, profit units in the system of corporate profit generation?
Y'all better work out your answer, and stand by it, fast. Because the algorithms are coming for you -- gay and straight, atheist and believer, left and right ... and once they have you fully in their thrall, the battle will be over.
44
When all the composite images are white you know you have a biased(bad) model!
4
No, you don't, Sylvia. The researchers said:
"All were white ... because they could not find enough dating profiles of gay minorities to generate a statistically valid result."
You've just shown that you don't know what is involved in scientific model making - one thing that is essential for good model making sample size. Say that I wanted to learn something about, say, eye color and Klopstokians. So I say, OK, here are 5 blue-eyed Klopstokians, here are 5 brown-eyed Klopstokians, let's see how many in each group have been diagnosed with X Fever, and bingo, we'll know which eye color puts Klopstokians at risk for X Fever.
Not so fast. Why not so fast? The sample is too small. You don't know how else these 10 Klopstokians differ from each other and from other Klopstokians, and what they have in common. You don't know what variables other than eye color come into play. But make the sample big enough, and bring in whatever else you know about Klopstokians, control for independent variables so you can reasonably tell you're not measuring something else (e.g. all men, all women), and you have a much better chance of learning something.
Then, when other scientists do other controlled studies, scientists will learn even more.
If you believe in science, then use it as a means to find truth, not as a weapon.
2
And someone " thought" this was a good use of time and resources????
Seriously STUPID.
6
The combination of facial recognition software, deep neural network software, and massive databases is a dangerous mix of technologies. Imagine the following scenario:
A government has access to all driver licences (and so has photographs and residential locations), all voter registration data (which includes party affiliation), data on the elections in which the person voted, all passports (and so have photographs and their travels outside the nation), their tax returns (and so their employment locations and contributions to political and religious charities) and data on all arrest, convictions and sentences. Imagine if the government collected all posts on blogs and comment sections in newspapers. Imagine if photo radars at intersections captured all drivers. Imagine if data on all cell phone calls and all emails were captured - not the content of the calls or emails, but simply whom communicated talked to whom and their locations at the time of the communication. The government could develop algorithms that predict political party affiliation, the likelihood of voting in elections, the likelihood of committing specific types of crime, patterns of employment by political party and other scenarios that can't be imagined now.
This is a serious issue that is unlikely to be addressed in the next few years.
29
Not just governments: employers, medical providers, schools, advertisers, political parties and their candidates, realtors, credit companies & banks, etc. "Imagine" that all the data a government has goes on sale to the highest bidders (who already have plenty of data on their own.)
7
no need to imagine, this is the state of affairs as of today.
3
Sounds like the basis for the 2011-2016 TV series "Person of Interest" or even 2002's "Minority Report". Eyes are everywhere and technology is being developed to gather our abundant data. Algorithms are developed to mine our private information. As long as there is the belief that the haves ought to be in charge of the have-nots then our personal beliefs and public records will be for sale to the highest bidder. Perhaps Kosinski & Wang are the canaries in the coal mine saying our private information is no longer as private or as safe as you think. We're already giving away so much of our private data and buying preferences because we are lazy; we want to trust it won't be misused or mislaid.
Given what recently happened with the Equifax breach, isn't it time for there to be true consequences to mismanagement of our personal data (or use of faux news)? Oh, disclaimers signed by users of certain websites or software programs protect businesses from going bankrupt over poorly executed client protections. "Oops! I didn't think of that!" Perhaps Kosinski & Wang are just telling us Politicians can be bought and Corporate entities will not be held accountable for taking advantage of advances in technology. So, how can we secure our privacy from invasive algorithms developed not only by do-gooder Americans but also by other countries who want to incite infighting amongst our own citizens? We are all connected but the misuse of technology is being used to divide and destroy us...
1
There was no accounting of age in these photos. What about older people over the age of 45 or 50 and up. I am becoming more and more disenchanted with my computer ... The dangers of hackers, terrorists, distortion of facts, more and more ads to ignore. I now wish it had never been invented. I rather enjoyed using encyclopedias and dictionaries . Even reading well-written mysteries by good writers can help learn about other societies from many ages.
22
That's an interesting point -- targeting faces by age, I'm sure many would say is quite doable. This particular sample was limited to a certain age group.
1
It is my understanding that subjects in a scientific experiment must be given the facts relating to the study and must give their signed consent. The men posting on Grindr, who were used as "subjects" in this research did not give their approval for their faces to be used.
17
There is some serious flaws in their study.
Because they worked from dating profiles the photos represent the subjects trying to look sexually attractive rather than neutral faces. They claim in the paper that their algorithm is designed to filter out facial expressions but if you look at the facial landmarks it is clear that the big difference is that gay men and straight women are smiling more with more raised eyebrows - a choice of presentation rather than inherent characteristics.
Secondly, the choice to compare pairs of photos is odd - since they selected a 50/50 population they could have simply guessed on individual photos. The pairing means that if some of the dating photos are distinctly "gay" - pursed lips or make-up on men or tight-lipped butch seriousness on women - each time those photos appear the computer will get the pair right - increasing the success rate.
7
Seconding bcw's second point. Reporting on this work is not noting that they construct a substantially easier problem that does not have practical applications. Who would look for a pair of people where they already know one is gay and try to assign the gay one? As bcw describes, in the regular individual classification problem, you get errors in the region where the scores of gay and straight people overlap. This problem is substantially easier because individuals in this ambiguous region have a better chance of being classified correctly in a pair. If there is some shift in the distributions of gay and straight people, an ambiguous point is now marked correctly with probability equal to the whole mass of the other distribution that is farther toward it's own end. The authors have been asked to release classification rates on individuals, within classes, but they have refused to do so. It appears that they are obfuscating low performance on the intuitive problem that a malicious agent might actually use.
1
I note your amusing stereotype that gay men will pose with "pursed lips" and makeup to point out something you apparently missed: this is a proof in concept of what computers can do when there is a specific type of data and a specific categorization task.
It does not matter if the algorithm "spuriously" identified gay men "because they posted higher quality photos." This merely restates the fact that gay men are more cunning in their photograph manipulation, which becomes a diagnostic sign.
It's also silly to object, as the article reports, that this is some form of "physiognomy". Physiognomy was a junk science that claimed to identify mental traits from facial features. This report only shows that people present facial features to correspond to a self chosen social category. Physiognomy claims to peer inward; algorithms detect what is projected outward.
Finally, the article gets mired in the debate about the statistical *power* of the algorithm, which technically is its ability to avoid false positive (incorrect) identifications. The discussion here indicates that the author of the article does not have a serious grounding in statistics.
Imagine, as a corrective, what would happen if the algorithm also had your likes, your web visits, your clickthroughs and your demographics to include in the data.
All of these are now available to aggregating sites such as Facebook and Google.
Be very afraid.
11
If this study has two co-authors, both at Stanford, then why is the article focused entirely on one, whereas the other only receives perfunctory, minimally necessary co-authorship mentions?
15
That's a good question. One of the authors, Dr. Kosinski was the sole writer the paper (though both did research) and all media inquiries were routed to him.
1
"Knowledge is a deadly friend
If no one sets the rules
The fate of all mankind I see
Is in the hands of fools"
- King Crimson
15
Only one question- WHY?
11
It all the controversy, it seems that the whole point of Dr. Kosiniski's project has been lost, which is usually what happens when knee-jerk reactions take hold. Dr. Kosinski's concern, and it should be a concern for us all, is "the potential for facial analysis to be misused and for findings about its effectiveness to be distorted."
Why should it acceptable to use facial recognition software to detect potential criminals, but not sexual orientation? The point is that any program that attempts to categorize, criminalize or penalize people, or claims to predict behavior, based on nothing but facial features should be abhorrent to all of us. Dr. Kosinski's project has nothing to do with being gay.
80
Hobby Lobby and Chik-a-Fil-A will be lining up to purchase this for their HR departments.
1
I agree with Ms Pea. The point of the study was to make people realize the danger of facial recognition - "Few seemed concerned. So to call attention to the privacy risks, he decided to show that it was possible to use facial recognition analysis to detect something intimate, something “people should have full rights to keep private.” This point even seems to gets lost in this article
49
I strongly agree that the controversy misses the point of the study, which is to warn us of the threat to privacy from computer algorithms. But it's a mistake to think that the study has nothing to do with being gay.
As Dr. Kosinski himself stated, there were a number of other characteristics -- e.g., atheism -- that he could have chosen to focus on, but he chose homosexuality. Why? The three characteristics mentioned in the article (atheism, homosexuality, & potential for criminality) have this is common: they are all widely considered socially negative and even immoral -- potential grounds for discrimination & punishment. Of the three, homosexuality was guaranteed to generate the most controversy. Homosexuality is one of the major battlefields in the Culture War. Dr. Kosinski wanted attention, and he got it.
1