A driver was pulled over for spending too long in the left lane, because her plate was flagged

Stops built on where your plate has been

Gizmodo reports that automated license plate readers scan plates, log time and location, and let police flag cars whose travel patterns look suspicious without suspicion of a specific crime. WIRED reported that Flock built an AI tool that can track vehicles by movement patterns alone.

The examples: Grace Knowles was stopped for allegedly spending too long in the left lane, with the officer saying her plate had been flagged. A California grandmother was questioned by Customs and Border Protection about how often she visited casinos. A driver repeatedly traveling between Wisconsin and Michigan was stopped, with police calling one state a "known source State for Marijuana". A federal jury in Texas sided with a man flagged after a Houston-to-Carrizo Springs round trip with an overnight stay.

An August YouGov poll found 46% opposed police use of the technology in their communities and 38% supported it. Some agencies, including the LAPD, have suspended or ended Flock contracts. Flock did not respond to Gizmodo.

"That's the opposite of a dragnet"

One commenter wrote that "we used to call those dragnets which are illegal under the 4th amendment". A skeptic pushed back on the term:

"A dragnet being an indiscriminate stop and search of suspects that sweeps up lots of people."

By that definition, he argued, plate readers are the exact opposite, since police could always watch a license plate. Another commenter said they should be banned federally.

Can a model's reasons be audited?

One commenter asked whether AI legalizes racial profiling:

"Nobody knows why the AI flagged suspicious behavior, it usually can't tell even if you are able to ask follow up questions."

A reply disagreed that it is opaque: you can edit the person in the image to be white and feed it back to see if the model's opinion changes. A developer of clinical software at a health company said his company's data scientists run bias checks before publishing models, and another commenter warned that a model trained on biased human decisions keeps the bias while being "a lot harder to prove/sue".

The counterfactual test only works when someone outside the vendor can run it, and Knowles' stop shows the officer's explanation was a map.