Stanford Invents AI Gaydar, Flubs Write-Up

Yilun Wang and Michal Kosinsksi, researchers at Stanford’s Graduate School of Business, have developed a neural-net classifier that purportedly detects sexual orientation (in caucasians).

The authors report an avalanche of experimental results, and claim the classifier can “correctly distinguish between gay and straight men 81% of the time, and 74% for women.”  OK, that’s the sensitivity of the gadget.  What about specificity, i.e. how well does it correctly distinguish folks who are not-so-gay?  Without that second number (as well as an estimate of prevalance), it’s not possible to estimate the false positive and false negative rates for this thing.  Very important, if some of the more Orwellian applications mentioned by the authors come to pass.

I give the authors a “C,” for incomplete work.

Good layman’s analysis in the Guardian.

Update: Dan Simmons, writing at the Andrew Gelman blog, writes a rambling, fascinating takedown of this “research,” from both the scientific and MSM points of view.  Based on just the statistical problems, I’m changing the grade to a “D-.”