Such systems have already been used and are already resulting in perpetuation of racism in an automated manner. Moreover, these systems:
1. Are Proprietary; defendants don't get to question how it comes to its conclusion, as the inner workings are trade secrets.
2. Are Opaque; see prior. Without extensive access to the tools and a degree in statistics, no one will ever find out about its biases except through years of abuse after the fact.
3. These systems only give back results similar to what you were getting before *at best.* Some minorities are heavily over-represented in prisons in the US, and the trained goal, if everything works exactly as planned, is to match those results. Even assuming the company making it does its absolute best at this and trains the system perfectly, the best you can expect from automating a biased policing system is an automated biased policing system.
This all assumes a system that works as designed and *merely recreates the existing systemic biases*. But with the secrecy from 1 and 2, a third party can't even determine if the system is just as biased as the old way of doing things, let alone worse.
Well, if you take the current data, including racial profiling biases then there would be some component purely correlated with race, and other components correlated with secondary characteristics that are themselves correlated with race.
If you then eliminate race from your decision process, but allow the secondary characteristics, that's going to change your future data set such that the amount purely associated with race is going to tend towards zero. So yes even with a biased sample base, the process will reduce the amount of racial profiling. There will still be biases in there, but they're not going to be very strong when you eliminate the main one that caused the initial bias. e.g. hypothetically "has dreadlocks" could be a good predictor of being arrested, and that might remain as a correlation if you disallow race data from being entered, but other concrete factors are going to in fact chip away at how well "has dreadlocks" explains variable arrest rates.
As for this bit, it's just not true. The patterns you're training on have bias baked into them -- any well trained system will pick up on them, and represent that in whatever way it most effectively can so long as it can find any proxy for that information. If there is literally any way to represent a factor that results in a 10x likelihood of arrest (as being black in the US does
https://www.usatoday.com/story/news/nation/2014/11/18/ferguson-black-arrest-rates/19043207/ ), it will find those links and represent them through any means necessary. And herein lies the problem -- there are tons of very subtle and interrelated systems at work that determine whether a person released from prison will be re-arrested; most of which are outside the realm of anything being captured by the system. Reduce this to a tractable statistical problem, and the end result will be a load of uninterpreted noise plus a strong signal of easily captured bias. It won't be a recidivism predictor, it will be a black male predictor, used to lock people up with no more explanation than "the computer said so, and computers can't be biased." Or as some machine learning folks on twitter put it: "Bias laundering"