Hacker News Reader

Comment by sliverstorm | Hacker News Reader

Are you saying you fear they will develop algorithms that are non-discriminatory but produce results that essentially mirror discriminatory results?

View on HN

"Fear" wouldn't be the word I would use, and I wouldn't say they're mirroring discriminatory results so much as approximating reality with some level of error. Reality is not egalitarian. Good approximations of reality will probably not be egalitarian, either. A nakedly discriminatory heuristic might be a better approximation of reality than "everyone is equal." A non-discriminatory heuristic might be a better approximation than that and also have the side effect of falling very unequally across arbitrary groupings of people.

For example, according to actual data I've seen, poor people really are disproportionately likely to a) have missed payments before, b) not repay your loan, and c) belong to certain ethnic groups. Some people may fit into one of these categories and not the others, but the correlation exists, so you can't clamp down on one without also measurably reducing the others unless you go out of your way to do so.

(Don't know who's downvoting, but I'm just stating pretty well-known facts. Don't take this to indicate approval — I'm just saying that, statistically, some groups do get the short end of the stick. What you do with this information is a different question entirely.)

Here is some actual data comparing ethnic groups. It does indeed conclude that blacks are more likely to exhibit risky financial profiles than whites.

http://www.bos.frb.org/economic/wp/wp1992/wp92_7.pdf

However, the study also concludes that for a sample of black people statistically identical to whites, the approval rate is only 82% (compared to 90% for whites).

Interesting. That study was only shortly after the FICO scoring system was introduced, so I'm curious if anything has changed.

Reply on news.ycombinator.com