With the crime you actually really need to catch (e.g. murders, terrorism, armed robberies) the problem is that they are rare, compared to them not happening. Most people on here have probably never even witnessed a murder for example, and terrorism is even rarer.

That means your prediction that eats data with mostly legal stuff all day must be extremely accurate to be useful. Let's say you wanna predict who a terrorist ist by the way they are walking at an airport. Given the high volume of people the answer will be: nearly nobody, except when they are. So your mechanism just produces false positives all day long and might encode stereotypes on the way.

The most important aspect for me would be that we actually demand scientific evidence of such a thing working before our executive is allowed to use it. Everything else is a huge risk for civil rights and doesn't help to prevent actual crime either.