Psychologist Michal Kosinski
says artificial intelligence can detect your sexuality and politics just by
looking at your face. What if he’s right?
Vladimir Putin was not in
attendance, but his loyal lieutenants were. On 14 July last year, the Russian
prime minister, Dmitry
Medvedev, and several members of his cabinet convened in an office building
on the outskirts of Moscow. On to the stage stepped a boyish-looking
psychologist, Michal Kosinski,
who had been flown from the city centre by helicopter to share his research.
“There was Lavrov, in the first row,” he recalls several months later,
referring to Russia’s foreign minister. “You know, a guy who starts wars and
takes over countries.” Kosinski, a 36-year-old assistant professor
of organisational behaviour at Stanford University, was flattered that the
Russian cabinet would gather to listen to him talk. “Those guys strike me as
one of the most competent and well-informed groups,” he tells me. “They did
their homework. They read my stuff.”
Kosinski’s “stuff” includes
groundbreaking research into technology, mass persuasion and artificial
intelligence (AI) – research that inspired the creation of the political
consultancy Cambridge
Analytica. Five years ago, while a graduate student at Cambridge
University, he showed how even benign activity on Facebook could reveal
personality traits – a discovery that was later exploited by the data-analytics
firm that helped put Donald Trump in the White House.
That would be enough to make
Kosinski interesting to the Russian cabinet. But his audience would also have
been intrigued by his work on the use of AI to detect psychological traits.
Weeks after his trip to Moscow, Kosinski published a controversial
paper in which he showed how face-analysing algorithms could
distinguish between photographs of gay and straight people. As well as
sexuality, he believes this technology could be used to detect emotions, IQ and
even a predisposition to commit certain crimes. Kosinski has also used algorithms
to distinguish between the faces of Republicans and Democrats, in an
unpublished experiment he says was successful – although he admits the results
can change “depending on whether I include beards or not”.
How did this 36-year-old
academic, who has yet to write a book, attract the attention of the Russian
cabinet? Over our several meetings in California and London, Kosinski styles
himself as a taboo-busting thinker, someone who is prepared to delve into
difficult territory concerning artificial intelligence and surveillance that
other academics won’t. “I can be upset about us losing privacy,” he says. “But
it won’t change the fact that we already lost our privacy, and there’s no going
back without destroying this civilisation.”
The aim of his research,
Kosinski says, is to highlight the dangers. Yet he is strikingly enthusiastic
about some of the technologies he claims to be warning us about, talking
excitedly about cameras that could detect people who are “lost, anxious,
trafficked or potentially dangerous. You could imagine having those diagnostic
tools monitoring public spaces for potential threats to themselves or to
others,” he tells me. “There are different privacy issues with each of those
approaches, but it can literally save lives.”
“Progress always makes people
uncomfortable,” Kosinski adds. “Always has. Probably, when the first monkeys
stopped hanging from the trees and started walking on the savannah, the monkeys
in the trees were like, ‘This is outrageous! It makes us uncomfortable.’ It’s
the same with any new technology.”
***
Kosinski has analysed
thousands of people’s faces, but never run his own image through his
personality-detecting models, so we cannot know what traits are indicated by
his pale-grey eyes or the dimple in his chin. I ask him to describe his own
personality. He says he’s a conscientious, extroverted and probably emotional
person with an IQ that is “perhaps slightly above average.” He adds: “And I’m
disagreeable.” What made him that way? “If you trust personality science, it
seems that, to a large extent, you’re born this way.”
His friends, on the other
hand, describe Kosinski as a brilliant, provocative and irrepressible data
scientist who has an insatiable (some say naive) desire to push the boundaries
of his research. “Michal is like a small boy with a hammer,” one of his
academic friends tells me. “Suddenly everything looks like a nail.”
Born in 1982 in Warsaw,
Kosinski inherited his aptitude for coding from his parents, both of whom
trained as software engineers. Kosinski and his brother and sister had “a computer
at home, potentially much earlier than western people of the same age”. By the
late 1990s, as Poland’s post-Soviet economy was opening up, Kosinski was hiring
his schoolmates to work for his own IT company. This business helped fund him
through university, and in 2008 he enrolled in a PhD programme at Cambridge,
where he was affiliated with the Psychometrics Centre, a
facility specialising in measuring psychological traits.
It was around that time that
he met David
Stillwell, another graduate student, who had built a personality quiz and
shared it with friends on Facebook. The app quickly went viral, as hundreds and
then thousands of people took the survey to discover their scores according to
the “Big Five” metrics: openness, conscientiousness, extraversion,
agreeableness and neuroticism. When users completed the myPersonality tests,
some of which also measured IQ and wellbeing, they were given an option to
donate their results to academic research.
Kosinski came on board, using
his digital skills to clean, anonymise and sort the data, and then make it
available to other academics. By 2012, more than 6 million people had taken the
tests – with about 40% donating their data, creating the largest dataset of its
kind.
In May, New Scientist magazine
revealed that the dataset’s
username and password had been accidentally left on GitHub, a commonly used
code-sharing website. For four years, anyone – not just authorised researchers
– could have accessed the data. Before the magazine’s investigation, Kosinski
had admitted to me that there were risks to their liberal approach. “We
anonymised the data, and we made scientists sign a guarantee that they will not
use it for any commercial reasons,” he had said. “But you just can’t really
guarantee that this will not happen.” Much of the Facebook data, he added, was
“de-anonymisable”. In the wake of the New Scientist story, Stillwell closed
down the myPersonality project. Kosinski sent me a link to the announcement,
complaining: “Twitter warriors and sensation-seeking writers made David shut
down the myPersonality project.”
During the time the
myPersonality data was accessible, about 280 researchers used it to publish
more than 100 academic papers. The most talked-about was a 2013 study co-authored
by Kosinski, Stillwell and another researcher, that explored the relationship
between Facebook “Likes” and the psychological and demographic traits of 58,000
people. Some of the results were intuitive: the best predictors of
introversion, for example, were Likes for pages such as “Video Games” and
“Voltaire”. Other findings were more perplexing: among the best predictors of
high IQ were Likes on the Facebook pages for “Thunderstorms” and “Morgan
Freeman’s Voice”. People who Liked pages for “iPod” and “Gorillaz” were likely
to be dissatisfied with life.
If an algorithm was fed with
sufficient data about Facebook Likes, Kosinski and his colleagues
found, it could make more accurate personality-based predictions than
assessments made by real-life friends. In other
research, Kosinski and others showed how Facebook data could be turned into
what they described as “an effective approach to digital mass persuasion”.
Their research came to the
attention of the SCL
Group, the parent company of Cambridge Analytica. In 2014, SCL tried to
enlist Stillwell and Kosinski, offering to buy the myPersonality data and their
predictive models. When negotiations broke down, they relied on the help of
another academic in Cambridge’s psychology department – Aleksandr
Kogan, an assistant professor. Using his own Facebook personality quiz, and
paying users (with SCL money) to take the tests, Kogan collected data on
320,000 Americans. Exploiting a loophole that allowed developers to harvest
data belonging to the friends of Facebook app users (without their knowledge or
consent), Kogan was able to hoover up additional data on as many as 87 million
people.
Christopher
Wylie, the whistleblower who lifted the lid on Cambridge Analytica’s
operations earlier this year, has described how the company set out to
“replicate” the work done by Kosinski and his colleagues, and to turn it into
an instrument of “psychological warfare”. “This is not my fault,” Kosinski
told reporters from the Swiss publication Das Magazin, which was the
first to make the connection between his work and Cambridge Analytica. “I did
not build the bomb. I only showed that it exists.”
Cambridge Analytica always
denied using Facebook-based psychographic targeting during the Trump campaign,
but the scandal over its data harvesting forced
the company to close. The saga also proved highly damaging to Facebook,
whose headquarters are less than four miles from Kosinski’s base at Stanford’s
Business School in Silicon Valley. The first time I enter his office, I ask him
about a painting beside his computer, depicting a protester armed with a
Facebook logo in a holster instead of a gun. “People think I’m anti-Facebook,”
Kosinski says. “But I think that, generally, it is just a wonderful
technology”.
Still, he is disappointed in
the Facebook CEO, Mark Zuckerberg, who, when he testified
before US Congress in April, said he was trying to find out “whether there
was something bad going on at Cambridge University”. Facebook, Kosinski says,
was well aware of his research. He shows me emails he had with employees in
2011, in which they disclosed they were “using analysis of linguistic data to
infer personality traits”. In 2012, the same employees filed a patent, showing
how personality characteristics could be gleaned from Facebook messages and
status updates.
Kosinski seems unperturbed by
the furore over Cambridge Analytica, which he feels has unfairly maligned
psychometric micro-targeting in politics. “There are negative aspects to it,
but overall this is a great technology and great for democracy,” he says. “If
you can target political messages to fit people’s interests, dreams,
personality, you make those messages more relevant, which makes voters more
engaged – and more engaged voters are great for democracy.” But you can also, I
say, use those same techniques to discourage your opponent’s voters from
turning out, which is bad for democracy. “Then every politician in the US is
doing this,” Kosinski replies, with a shrug. “Whenever you target the voters of
your opponent, this is a voter-suppression activity.”
Kosinski’s wider complaint
about the Cambridge Analytica fallout, he says, is that it has created “an
illusion” that governments can protect data and shore up their citizens’
privacy. “It is a lost war,” he says. “We should focus on organising our
society in such a way as to make sure that the post-privacy era is a habitable
and nice place to live.”
***
Kosinski says he never set out
to prove that AI could predict a person’s sexuality. He describes it as a
chance discovery, something he “stumbled upon”. The lightbulb moment came as he
was sifting through Facebookprofiles for
another project and started to notice what he thought were patterns in people’s
faces. “It suddenly struck me,” he says, “introverts and extroverts have
completely different faces. I was like, ‘Wow, maybe there’s something there.’”
Physiognomy, the practice of
determining a person’s character from their face, has a history that stretches
back to ancient Greece. But its heyday came in the 19th century, when the
Italian anthropologist Cesare Lombroso published his famous taxonomy, which
declared that “nearly all criminals” have “jug ears, thick hair, thin beards,
pronounced sinuses, protruding chins, and broad cheekbones”. The analysis was
rooted in a deeply racist school of thought that held that criminals resembled
“savages and apes”, although Lombroso presented his findings with the precision
of a forensic scientist. Thieves were notable for their “small wandering eyes”,
rapists their “swollen lips and eyelids”, while murderers had a nose that was
“often hawklike and always large”.
Lombroso’s remains are still
on display in a
museum in Turin, besides the skulls of the hundreds of criminals he spent
decades examining. Where Lombroso used calipers and craniographs, Kosinski has
been using neural networks to find patterns in photos scraped from the
internet.
Kosinski’s research dismisses
physiognomy as “a mix of superstition and racism disguised as science” – but
then argues it created a taboo around “studying or even discussing the links
between facial features and character”. There is growing evidence, he insists,
that links between faces and psychology exist, even if they are invisible to
the human eye; now, with advances in machine learning, such links can be
perceived. “We didn’t have algorithms 50 years ago that could spot patterns,”
he says. “We only had human judges.”
In a paper published last year, Kosinski and a
Stanford computer scientist, Yilun Wang, reported that a machine-learning
system was able to distinguish between photos of gay and straight people with a
high degree of accuracy. They used 35,326 photographs from dating websites and
what Kosinski describes as “off-the-shelf” facial-recognition software.
Presented with two pictures –
one of a gay person, the other straight – the algorithm was trained to distinguish
the two in 81% of cases involving images of men and 74% of photographs of
women. Human judges, by contrast, were able to identify the straight and gay
people in 61% and 54% of cases, respectively. When the algorithm was shown five
facial images per person in the pair, its accuracy increased to 91% for men,
83% for women. “I was just shocked to discover that it is so easy for an
algorithm to distinguish between gay and straight people,” Kosinski tells me.
“I didn’t see why that would be possible.”
Neither did many other people,
and there was an immediate backlash when the research – dubbed “AI gaydar” –
was previewed in the Economist magazine. Two of America’s most prominent LGBTQ
organisations demanded that Stanford distance itself from what they called its
professor’s “dangerous and flawed research”. Kosinski received a deluge of
emails, many from people who told him they were confused about their sexuality
and hoped he would run their photo through his algorithm. (He declined.) There
was also anger that Kosinski had conducted research on a technology that could
be used to persecute gay people in countries such as Iran and Saudi Arabia,
where homosexuality is punishable by death.
Kosinski says his critics
missed the point. “This is the inherent paradox of warning people against
potentially dangerous technology,” he says. “I stumbled upon those results, and
I was actually close to putting them in a drawer and not publishing – because I
had a very good life without this paper being out. But then a colleague asked
me if I would be able to look myself in the mirror if, one day, a company or a
government deployed a similar technique to hurt people.” It would, he says,
have been “morally wrong” to bury his findings.
One vocal critic of that defence
is the Princeton professor Alexander Todorov,
who has conducted some of the most widely cited research into faces and
psychology. He argues that Kosinski’s methods are deeply flawed: the patterns
picked up by algorithms comparing thousands of photographs may have little to
do with facial characteristics. In a
mocking critique posted online, Todorov and two AI researchers at
Google argued that Kosinski’s algorithm could have been responding to patterns
in people’s makeup, beards or glasses, even the angle they held the camera at.
Self-posted photos on dating websites, Todorov points out, project a number of
non-facial clues.
Kosinski acknowledges that his
machine learning system detects unrelated signals, but is adamant the software
also distinguishes between facial structures. His findings are consistent with the
prenatal hormone theory of sexual orientation, he says, which argues that the
levels of androgens foetuses are exposed to in the womb help determine whether
people are straight or gay. The same androgens, Kosinski argues, could also
result in “gender-atypical facial morphology”. “Thus,” he writes in his paper,
“gay men are predicted to have smaller jaws and chins, slimmer eyebrows, longer
noses and larger foreheads... The opposite should be true for lesbians.”
This is where Kosinski’s work
strays into biological determinism. While he does not deny the influence of
social and environmental factors on our personalities, he plays them down. At
times, what he says seems eerily reminiscent of Lombroso, who was critical of
the idea that criminals had “free will”: they should be pitied rather than
punished, the Italian argued, because – like monkeys, cats and cuckoos – they
were “programmed to do harm”.
“I don’t believe in guilt,
because I don’t believe in free will,” Kosinski tells me, explaining that a
person’s thoughts and behaviour “are fully biological, because they originate
in the biological computer that you have in your head”. On another occasion he
tells me, “If you basically accept that we’re just computers, then computers
are not guilty of crime. Computers can malfunction. But then you shouldn’t
blame them for it.” The professor adds: “Very much like: you don’t, generally,
blame dogs for misbehaving.”
Todorov believes Kosinski’s
research is “incredibly ethically questionable”, as it could lend a veneer of
credibility to governments that might want to use such technologies. He points
to a paper that appeared
online two years ago, in which Chinese AI researchers claimed they had
trained a face-recognition algorithm to predict – with 90% accuracy – whether
someone was a convicted criminal. The research, which used Chinese government
identity photographs of hundreds of male criminals, was not peer-reviewed, and
was torn
to shreds by Todorov, who warned that “developments in artificial
intelligence and machine learning have enabled scientific racism to enter a new
era”.
Kosinski has a different take.
“The fact that the results were completely invalid and unfounded, doesn’t mean
that what they propose is also wrong,” he says. “I can’t see why you would not
be able to predict the propensity to commit a crime from someone’s face. We
know, for instance, that testosterone levels are linked to the propensity to
commit crime, and they’re also linked with facial features – and this is just
one link. There are thousands or millions of others that we are unaware of,
that computers could very easily detect.”
Would he ever undertake
similar research? Kosinski hesitates, saying that “crime” is an overly blunt
label. It would be more sensible, he says, to “look at whether we can detect
traits or predispositions that are potentially dangerous to an individual or
society – like aggressive behaviour”. He adds: “I think someone has to do it…
Because if this is a risky technology, then governments and corporations are
clearly already using it.”
***
But when I press Kosinski for
examples of how psychology-detecting AI is being used by governments, he
repeatedly falls back on an obscure Israeli startup, Faception. The company provides software
that scans passports, visas and social-media profiles, before spitting out
scores that categorise people according to several personality types. On its
website, Faception lists eight such classifiers, including
“White-Collar Offender”, “High IQ”, “Paedophile” and “Terrorist”. Kosinski
describes the company as “dodgy” – a case study in why researchers who care
about privacy should alert the public to the risks of AI. “Check what Faception
are doing and what clients they have,” he tells me during an animated debate
over the ethics of his research.
I call Faception’s chief
executive, Shai Gilboa, who used to work in Israeli military intelligence. He
tells me the company has contracts working on “homeland security and public
safety” in Asia, the Middle East and Europe. To my surprise, he then tells me
about a research collaboration he conducted two years ago. “When you look in
the academia market you’re looking for the best researchers, who have very good
databases and vast experience,” he says. “So this is the reason we approached
Professor Kosinski.”
But when I put this connection
to Kosinski, he plays it down: he claims to have met Faception to discuss the
ethics of facial-recognition technologies. “They came [to Stanford] because they
realised what they are doing has potentially huge negative implications, and
huge risks.” Later, he concedes there was more to it. He met them “maybe three
times” in Silicon Valley, and was offered equity in the company in exchange for
becoming an adviser (he says he declined).
Kosinski denies having
collaborated on research, but admits Faception gave him access to its
facial-recognition software. He experimented with Facebook photos in the
myPersonality dataset, he says, to determine how effective the Faception
software was at detecting personality traits. He then suggested Gilboa talk to
Stillwell about purchasing the myPersonality data. (Stillwell, Kosinski says,
declined.)
He bristles at my suggestion
that these conversations seem ethically dubious. “I will do a lot of this,” he
says. “A lot of startup people come here and they don’t offer you any money,
but they say, ‘Look, we have this project, can you advise us?’” Turning down
such a request would have made him “an arrogant prick”.
He gives a similar explanation
for his trip to Moscow, which he says was arranged by Sberbank Corporate University as
an “educational day” for Russian government officials. The university is a
subsidiary of Sberbank, a state-owned bank sanctioned by the EU; its chief
executive, Russia’s former minister for economic development, is close to
Putin. What was the purpose of the trip? “I didn’t really understand the
context,” says Kosinski. “They put me on a helicopter, flew me to a place, I
came on the stage. On the helicopter I was given a briefing about who was going
to be in the room. Then I gave a talk, and we talked about how AI is changing
society. And then they sent me off.”
The last time I see Kosinski,
we meet in London. He becomes prickly when I press him on Russia, pointing to
its dire record on gay rights. Did he talk about using facial-recognition
technology to detect sexuality? Yes, he says – but this talk was no different
from other presentations in which he discussed the same research. (A couple of
days later, Kosinski tells me he has checked his slides; in fact, he says, he
didn’t tell the Russians about his “AI gaydar”.)
Who else was in the audience,
aside from Medvedev and Lavrov? Kosinski doesn’t know. Is it possible he was
talking to a room full of Russian intelligence operatives? “That’s correct,” he
says. “But I think that people who work for the surveillance state, more than
anyone, deserve to know that what they are doing is creating real risk.” He
tells me he is no fan of Russia, and stresses there was no discussion of spying
or influencing elections. “As an academic, you have a duty to try to counter
bad ideas and spread good ideas,” he says, adding that he would talk to “the
most despicable dictator out there”.
I ask Kosinski if anyone has
tried to recruit him as an intelligence asset. He hedges. “Do you think that if
an intelligence agency approaches you they say: ‘Hi, I’m the CIA’?” he replies.
“No, they say, ‘Hi, I’m a startup, and I’m interested in your work – would you
be an adviser?’ That definitely happened in the UK. When I was at Cambridge, I
had a minder.” He tells me about a British defence expert he suspected worked
for the intelligence services who took a keen interest in his research,
inviting him to seminars attended by officials in military uniforms.
In one of our final
conversations, Kosinski tells me he shouldn’t have talked about his visit to
Moscow, because his hosts asked him not to. It would not be “elegant” to
mention it in the Guardian, he says, and besides, “it is an irrelevant fact”. I
point out that he already left a fairly big clue on Facebook, where he posted
an image of himself onboard a helicopter with the caption: “Taking off to give
a talk for Prime Minister Medvedev.” He later changed his privacy settings: the
photo was no longer public, but for “friends only”.
No comments:
Post a Comment