Airbnb reduced partying by 55% in 2 years

They also called it an "AI system", but from the article it looks more like a statistical model that assign a risk score to every attribute (age, location, when, duration, ecc) and then do sum/weighted average. If the sum/weighted average are above a threshold, the reservation is flagged.

This testing would be an interesting article of its own. How does one do this while keeping the model within the bounds of the law?

For example: Say you're in a jurisdiction that prohibits discrimination based on a renter's youth. You might be able to argue that denying a college-aged person a 2 day rental over their birthday for fear of a party is not illegally discriminatory. It's much less defensible if it's a 2 month rental over their birthday. But the local law probably doesn't specify a threshold like that. So creating a threshold-based test -- which is constructively a policy -- could itself be legally thorny. (But IANAL)