To spare you the search of his comment history, my memory is that it's roughly: the stolen car databases have tons of cars listed that weren't actually stolen, or have already been recovered, so you end up pulling over a ton of unsuspecting people just driving around.

One answer is that ALPRs have much too high an error rate (John Oliver claims it exceeds 30%) to rely on them for dragnet surveillance.

Maybe it makes sense to track specific vehicles, with safeguards such as warrants and also checking make/model/color, but it makes less sense to just give the government a 30-day rolling database of the public's movements in total, especially when 30% of those data points are inaccurate and thus not useful as evidence.