It's a small leap to combine panoptic surveillance with predictive policing (the latter has been used in large cities already) - along with automated classification labeling "suspect" behaviour.
Note for example that it is rather likely, if you go from an historic "know good" dataset containing more convictions of poor people of colour, that you would automate racial profiling... Quite possibly mixing cause and effect in the training period. Consider further the implications of a company running prisons (paid per inmate) going into the business of selling city/country-wide surveillance systems to "assist police".