Thread by @sifutweety

SO! This is incredibly unsurprising but it also makes me super ranty, in part because I have published papers on how you build a deep network that can estimate (people's first impressions of -- that's THE KEY BIT) personality traits from face images (ranty thread)

https://twitter.com/hatr/status/1361756449802768387

There are a billion problems with how people talk about/commercialize this. The first and maybe biggest problem is that accurately judging somebody's (inherent, stable) personality from an image of their face is LIKELY IMPOSSIBLE

How do I know it's impossible? Well, because people can't do it. In fact, if you show people images of other people's faces and ask them to judge if somebody is trustworthy or whatever, their answers robustly fail to correlate with that person's actual behavior in their life.

This is work by Alex Todorov at Princeton, by the way -- I'm feeling to ranty to cite, but look him up if you want somebody actually smart on this stuff.

So if people -- who are by any estimate orders of magnitude better at social inference than any computer system designed -- can't actually judge somebody's personality, what's going on?

Well, two things. One is that irrespective of how well it correlates with actual behavior people DO judge other people's personality from face images, all the time, and they do it reliably.

So there are certain faces where if you show them to 100 people every single one of those people will say "oh, I can totally trust her."

I said "her" because when I collected massive scale data on this the most trustworthy-appearing faces were all young white women with no makeup.

Which is another aspect of this -- obviously but obviously these kind of judgments embed every sort of bias. To a first approximation, this kind of social signal is ALL bias. Which, in practice, can be kind of ok? We use heuristics to understand all kinds of things

You just need to be aware of them and think of ways to work through them in cases where your personal bias is going to cause negative outcomes to other people/social groups. What happens in practice, though

is that the people training these algorithms lack the clarity of understanding to realize what they're doing, and end up building systems that launder that bias as algorithmic clarity.

This is made even WORSE because most of the people publishing/commercializing work DON'T UNDERSTAND that they aren't actually detecting stable personality traits.

They're approximating how PEOPLE would describe those traits. It is a useful thing to estimate, or can be, but it has absolutely nothing to do with somebody's actual personality.

This division, between personality and perceived personality, is reasonably clear -- could be a little better -- in the psychological literature, but is absolutely lost on most people working on this in ML.

An additional problem comes when you naively apply the techniques of computer vision to something like the Big 5; computer vision is designed to extract ground truth about the world -- that's an objective, knowable, empirical fact about the location or nature of things around you

the Big 5 is EXTREMELY not that. It's a medium-reliable psychological measure which takes a lot of care and, honestly, some crossing of fingers and hoping, to measure reliably.

And notably, the Big 5 was NEVER designed as a measure you could estimate from face images The gold standard tests use large batteries of questions as the input into their factor model.

The idea that you can do it in real-time with a deep network trained on face images paired with trait factor labels is so antiscientific it might as well be dowsing.

It's a shame, because there is important and useful work to be done on how we can expand the definition of computer vision to include estimation of characteristics that don't easily map to ground truth

But instead we get charlatans and incompetents making claims that don't hold up to even minimal scrutiny.