They have lots of non-actionable data on thousands of individuals ("individual X is a potentially violent extremist but hasn't committed any actual crimes yet").
One could make an argument that what they are after here is to get actually actionable data ("individual X is plotting an attack next week together with individuals Y and Z").
I don't think this is devil's advocate arguing.. There's a slippery slope from what you describe to the concept of "thought crime", which sounds hokey and conspiracy-ish but given the level of surveillance we have, we're not that far off.
Thought crime already exists for a long time already. Planning to commit murder (thinking about it) is punished nearly as harshly as succeeding. Is it really so controversial?
I suppose machine learning could study communication patterns to see what patterns are there before an attack.. e.g. if suddenly they started talking about birds when they never did before, did they just establish some code words? The dark part is the computers will need training data...
There's a genuine reason why machine learning seems to be the Holy Grail - man has dominated this planet, solved his problems and overcome challenges by his intelligence (ability to learn and adapt). Bequeathing these same abilities to machines is a big deal and might be the solution to everything.
That's not actionable. The dark part here is the serious threat of applying oppressive prior restraint policies at will. Employing machine learning in the process would only offer a false sense of legitimacy.
They have lots of non-actionable data on thousands of individuals ("individual X is a potentially violent extremist but hasn't committed any actual crimes yet").
One could make an argument that what they are after here is to get actually actionable data ("individual X is plotting an attack next week together with individuals Y and Z").