The famo AI gaydar study was repeated – and, no, n't tell if you're straight or not jt om your face • The Register

Contents:

• A STANFORD SCIENTIST SAYS HE BUILT A GAYDAR G “THE LAMT” AI TO PROVE A POT
• THE FAMO AI GAYDAR STUDY WAS REPEATED – AND, NO, N'T TELL IF YOU'RE STRAIGHT OR NOT JT OM YOUR FACE
• GAY SIMULATOR
• ROW OVER AI THAT 'INTIFI GAY FAC'
•  GAY ROBOT
• BIG GAY AI
• SOUTH PARK: THE FRACTURED BUT WHOLE BIG GAY AL'S BIG GAY CATS SVENGER HUNT LOTN GUI

A STANFORD SCIENTIST SAYS HE BUILT A GAYDAR G “THE LAMT” AI TO PROVE A POT

Last week, The Enomist published a story around Stanford Graduate School of Bs rearchers Michal Kosski and Yilun Wang’s claims that they had built artificial telligence that uld tell if we are gay or straight based on a few imag of our fac. It me om those who follow AI rearch, as well as om LGBTQ groups such as Gay and Lbian Advot & Defenrs (GLAAD).

What their technology n regnize is a pattern that found a small subset of out, whe gay and lbian people on datg s who look siar. Those two fdgs should not be nflated, ” Jim Halloran, GLAAD’s chief digal officer, wrote a statement claimg the paper uld e harm exposg methods to target gay the other hand, LGBTQ Natn, a publitn foced on issu the lbian, gay, bisexual, transgenr, queer muny, disagreed wh GLAAD, sayg the rearch intified a potential threat.

Kosski asserted an terview wh Quartz that regardls of the methods of his paper, his rearch was service of gay and lbian people that he se unr siege morn society. He says his work stands on the shoulrs of rearch happeng for s—he’s not reventg anythg, jt translatg known differenc about gay and straight people through new technology. You take some data— this se was 15, 000 pictur of gay and straight people om a popular datg webse—and show to a ep-learng algorhm.

THE FAMO AI GAYDAR STUDY WAS REPEATED – AND, NO, N'T TELL IF YOU'RE STRAIGHT OR NOT JT OM YOUR FACE

Rather, should foc on permanent featur like the length of the nose or width of the face, bee those stay nsistent when regnizg a thors say this is necsary bee, sce the pictur are om a datg se, gay and straight people might e different facial exprsns, mera angl, or pos to seem attractive based on the sex they’re tryg to attract.

GAY SIMULATOR

It’s an ia backed by soclogists, and if your claim is that facial stcture is different for gay people and straight people, then you fely don’t want somethg as fleetg as a facial exprsn to nfe that.

ROW OVER AI THAT 'INTIFI GAY FAC'

A heat map of where the algorhm looks to tect signs of homosexualyImage: Kosski and WangBut some AI rearchers doubt that VGG-Face is actually ignorg exprsn and pose, bee the mol isn’t beg ed for s tend e, to simply intify people. For stance, they found that gay and straight men wore different facial hair, and gay women were more prone to wear baseball is jt one crique: we don’t know for sure that this is te, and ’s impossibly difficult to fd out given the rmatn provid the paper or even if you had the algorhm. But this potential explanatn based on the tt of another AI rearcher throws doubt to the ia that VGG-Face n be ed as a perfect oracle to tect somethg about a person’s facial featur while ignorg nfoundg algorhm generated the optimally gay and straight Kosski and WangThe send aspect of this rearch, apart om the algorhm, are the data ed to tra the facial-regnn system.

 GAY ROBOT

The pany uld have disclosed technologil or cultural bias herent the data for rearchers to way, ’s unclear how imag of people taken om datg webs and sorted only to gay and straight tegori accurately reprent their sexualy. The thors also assume that men lookg for male partners and femal lookg for female partners are gay, but that’s a stunted, bary distillatn of the sexual spectm soclogists today are tryg to unrstand. “We don’t know who’s gay.

Is an inty where you stand up and say ‘I am gay, ’ is an unrlyg attractn, or is a behavr?

To only measure terms of gay or straight don’t accurately reflect the world, but stead forc a human nstct onto —a hallmark of bad that, Kosski says the study was nducted wh the nf of what ers reported themselv to be lookg for on the datg s—and back to the pot that someone g this malicly wouldn’t spl hairs over whether someone was bisexual or about the numbers? Dpe the 91% accuracy reported other news outlets, that number only om a very specific algorhm is shown five pictur each of two different people who were lookg for the same or oppose sex on the datg se, and told that one of them is gay. Another experiment tailed the paper lls the difference analyzed between gay and straight fac to further qutn.

BIG GAY AI

Estimat om the paper put roughly 7% of the populatn as gay (Gallup says 4% intify as LGBT as of this year, but 7% of lennials), so om a random draw of 100 people, seven would be gay.

Then they tell the algorhm to pull the top 100 people who are most likely to be gay om the full 1, algorhm do ; but only 43 people are actually gay, pared to the entire 70 expected to be the sample of 1000.

The remag 57 are straight, but somehow exhib what the algorhm thks are signs of gayns. At s most nfint, asked to intify the top 1% of perceived gayns, only 9 of 10 people are rrectly labeled.

SOUTH PARK: THE FRACTURED BUT WHOLE BIG GAY AL'S BIG GAY CATS SVENGER HUNT LOTN GUI

While accuracy is a measure of succs, Kosski said he didn’t know if was ethilly sound to create the bt algorhmic approach, for fear someone uld replite , stead optg to e off-the-shelf realy, this isn’t an algorhm that tells gay people om straight people.

After readg Kosski and Wang’s paper, three soclogists and data scientists who spoke wh Quartz qutned whether the thor’s assertn that gay and straight people have different fac is supported by the experiments the paper.