The problem with predictive policing is in the name. Inference (ML) predicts the future from the past. If the past is racist, then inference will create a racist future. Since racism is systemic[1], especially when it comes to policing, predictive policing is actively working against an anti-racist future.
There may be statistical ways to factor out systemic racism. There are two reasons I don't think that works:
1. I don't see how one evaluates the correctness of the process that controls for racism. What is ground truth for anti-racist policing?
2. These systems are likely snake oil and the vendors of these systems are (possibly inadvertently) profiting off racist policing. If cops arrest more black people per capita, then send the cops to black neighborhoods and have them follow black parolees. The system works (according to an objective function which maximizes arrests.) Remove racism and send the cops to white neighborhoods. Now the cops don't arrest as many people. The system fails. So I think it's likely that if you remove racist policing from predictive policing, you get the null hypothesis.
I'd be happy to hear a counterargument from someone who has actual statistics on this though.
[1] If you don't believe this, you're in the minority now: https://www.vox.com/2020/6/11/21286642/george-floyd-protests...
There may be statistical ways to factor out systemic racism. There are two reasons I don't think that works:
1. I don't see how one evaluates the correctness of the process that controls for racism. What is ground truth for anti-racist policing?
2. These systems are likely snake oil and the vendors of these systems are (possibly inadvertently) profiting off racist policing. If cops arrest more black people per capita, then send the cops to black neighborhoods and have them follow black parolees. The system works (according to an objective function which maximizes arrests.) Remove racism and send the cops to white neighborhoods. Now the cops don't arrest as many people. The system fails. So I think it's likely that if you remove racist policing from predictive policing, you get the null hypothesis.
I'd be happy to hear a counterargument from someone who has actual statistics on this though.
[1] If you don't believe this, you're in the minority now: https://www.vox.com/2020/6/11/21286642/george-floyd-protests...