So the book touched on this. Their model was literally a model of the judge that it was compared against - as in they took a judges decision history and turned it into a model where it was the decision the judge made on average in a case with these circumstances, then compared it to the judge themselves. They did this with a bunch of judges, and found that while neither the judge nor the model did much better than 50/50 at guessing if someone would commit a crime/do a runner while on bail, there was a big improvement in the amount of time spent in custody with the computer decision - as in the computer model did much better at picking the ones that led to reduced time in custody for the same success rate vs the judge. So there's a good possibility of being more cautious at keeping more risky people in custody while overall not needing more space to keep them.

The issue is that people aren't comfortable with a computer making the decision, and also that the model will inevitably be biased against certain races because of a history of biased decisions that the model is based on (even of the model doesn't take race into account). There's a feedback loop, where if you remand someone in custody they can't keep their job and are more likely to commit crimes in future. So you don't want to feed into that if possible. But I'd think there's a good argument for violent crimes (like in this case) to be treated more cautiously and ease up on nonviolent crime to balance.

The book talks about US judges but I'd think many of the benefits and concerns also apply in NZ.