Yeah, the consensus "here" is that deep learning, self-driving cars, all that, is "half-there". Presently, deep learning works, has a fairly high accuracy rate, in some percentage of cases. Sure, you can implement it right now in all sorts of places. The question is what the costs of the small but significant failure rate is going to be and
who will pay that cost?. If some nobody pays the cost, well, a high failure rate may be fine with the institution that implements this.
With cars, the idea of a 0.1% chance of dying each time you step into a vehicle is terrifying - that's much higher than you get now, humans are actually amazingly good at driving. IE, fortunately, the system probably won't stomach the costs of bad self-driving cars.
But with the justice system spotting potential criminals, yeah, pulling in some innocents isn't going to bother a system with N degrees of repression. Most people in the justice system have no choice but to plea bargain now. Most people in the justice system are also guilty but the innocent-but-not-wealthy minority is generally SOL. And that is in our ostensible democracy, that's known by most sophisticated observers of the system, and only handful make an effort to change it. So a half-assed evidence gatherer of whatever sort will make lots of people happy, put lot of guilty and innocent people in jail and hardly needs great reliability. Indeed, with AI parole-granting systems, we're already there.