Only as good as the inputs, yes. But if it's a halfway decent filter, it will include machine learning, e.g. a Bayesian filter, and if "is Muslim" turns out to have low correspondence with actual criminal activity that input will quickly be deweighted. Or perhaps paired with other aspects- e.g., perhaps "is Muslim" is of no consequence and "Googles Jihad" is also of no consequence, but "is Muslim" && "Googles Jihad" gives you a point. Just as one example of the patterns a good filter could recognize.

Machines can't be racist, so the arrest score going after lots of poor, Black men must mean that there's something to it.

If a learning Bayesian filter targets a certain demographic, there probably IS something to it.

That really would be amusing/pleasing, if all this work we've spent developing spam filters became the lead-up to an accurate, learning crime filter. Perhaps the fork to spamassassin will be known as crimeassassin?