[yuzi]

There's a huge flaw in there. When I rent a unit, I will do some client research and to the extent I can I will block the rental based upon my findings. For example if I read their blog and they come across as a risky customer (a profile photo with machine guns, money, and drugs or postings showing a they have a lack of character). Now it could be they also happen to be black, but when they create a fake profile and then get accepted it does not mean I blocked them due to race.

At a minimum you would need to make sure you fake both profiles to ensure everything but skin color is the same.

[abalone]

> ensure everything but skin color is the same

Which would be utterly obvious and ineffective. Instead, they could look for discriminatory rejection patterns.

[yuzi]

> discriminatory rejection patterns.

Maybe I'm lacking imagination, but how do you account for the renter and Airbnb having vastly different sets of information to work from? Also your sample size would have to be huge to rule out coincidental outcomes, which I dare say would rarely be found for the average property.

For example my property gets about 100 customers per year. I'm guessing in the last 4 years I've had 3 black customers, all accepted by the way, but if I rejected all 3 can an algorithm reasonably rule out coincidence having such a low sample size? It just seems like it would be the kind of algo doomed to fuck it up.

[abalone]

It's like spam filtering. It may be hard to imagine but it can work. Algorithms look at multiple signals and produce a confidence score.

Consistently rejecting the one black customer you get every year would probably raise your score enough to prompt an investigation, and the outcome of that (like a fake white test booking) would give more confidence that it's not just coincidence.