4 ms·

> In comparisons of black and white individuals, there was a 4% difference from demographic parity in favor of white individuals. It's hard to believe that the

by fwip 5y ago

> In comparisons of black and white individuals, there was a 4% difference from demographic parity in favor of white individuals. It's hard to believe that the bias was only 4% - there were a lot of people testing with images that they sourced themselves, and the preference for white people seemed much closer to 80-20. The paper authors mention that their training data is from Wikidata (pictures of celebrities). I wonder if the types of photos in that dataset are meaningfully representative of the kinds of photos that people usually post to Twitter.

gwern 5y ago
> It's hard to believe that the bias was only 4% - there were a lot of people testing with images that they sourced themselves, and the preference for white people seemed much closer to 80-20. It's very easy to believe the bias was near-zero given you are citing highly motivated people on Twitter cherrypicking from thousands of examples and a little baffling you find that to be more credible than controlled systematic experimentation; note, for example, the extremely striking fact that the fuss completely missed the other bias they found which was several times larger - showing how totally useless people on social media are for testing these things and how they can conjure up "80-20" biases which don't exist.
- eyelidlessness 5y ago
  > highly motivated people on Twitter cherrypicking One of the reference threads that identifies the issue happened on it by accident highlighting a surprising experience in another product (Zoom). Believe it or not, people who care about this stuff are not looking for stuff to complain about, we’re tired and overwhelmed. And I would hope that people who, upon discovering a vulnerability find and catalogue the ways it can be exploited, would be celebrated here.
- fwip 5y ago
  I admit that it wasn't especially rigorous testing, but I personally tested this along with other people I knew. I used real photos from my camera and my wife's (we are of different races), featuring photos of ourselves, friends, and family. I of course hope that the systems I use aren't racist against my loved ones. I am motivated to confirm whether or not they are, but I didn't go on to parlay my findings into an essay for clout. I gained nothing from doing this, except the knowledge that Twitter was suckier than I knew.

by fwip 5y ago

gwern 5y ago

> It's hard to believe that the bias was only 4% - there were a lot of people testing with images that they sourced themselves, and the preference for white people seemed much closer to 80-20. It's very easy to believe the bias was near-zero given you are citing highly motivated people on Twitter cherrypicking from thousands of examples and a little baffling you find that to be more credible than controlled systematic experimentation; note, for example, the extremely striking fact that the fuss completely missed the other bias they found which was several times larger - showing how totally useless people on social media are for testing these things and how they can conjure up "80-20" biases which don't exist.

eyelidlessness 5y ago
> highly motivated people on Twitter cherrypicking One of the reference threads that identifies the issue happened on it by accident highlighting a surprising experience in another product (Zoom). Believe it or not, people who care about this stuff are not looking for stuff to complain about, we’re tired and overwhelmed. And I would hope that people who, upon discovering a vulnerability find and catalogue the ways it can be exploited, would be celebrated here.
fwip 5y ago
I admit that it wasn't especially rigorous testing, but I personally tested this along with other people I knew. I used real photos from my camera and my wife's (we are of different races), featuring photos of ourselves, friends, and family. I of course hope that the systems I use aren't racist against my loved ones. I am motivated to confirm whether or not they are, but I didn't go on to parlay my findings into an essay for clout. I gained nothing from doing this, except the knowledge that Twitter was suckier than I knew.