Automated license plate readers do not arrest people.
People arrest people after information moves through a decision chain.
That sounds obvious until a machine result acquires the psychological weight of a fact.
The San Diego case is a contradiction test.
Times of San Diego reported that Hugo Parra was arrested in a violent-crime investigation after police had a general description of an Alfa Romeo, a witness identification, and a result from the city's Flock system that investigators treated as corroborating evidence.
Parra's attorney later argued that the same Flock data showed the vehicle about five miles away from the crime scene at the relevant time and that the image was captured before the police pursuit began.
If that account is correct, the important failure is not that “the camera hallucinated a plate.” The problem is that an investigative result was interpreted as support while its timestamp and geography contained information pointing the other way.
Machine evidence needs contradiction handling.
camera read → database match → investigator receives result → timestamp/location check → comparison with witness evidence → stop/arrest decision
Every step matters.
A correct plate read can still be used incorrectly. A correct timestamp can still be ignored. A vehicle similarity search can return the right make and color but the wrong vehicle. A witness identification can reinforce a machine result and vice versa until two weak signals feel like one strong conclusion.
That is evidence fusion, and evidence fusion can amplify error as easily as confidence.
Human verification must test the result, not salute it.
Flock transparency portals commonly state that hotlist hits are required to be human verified prior to action. That is good design language, but the phrase “human verified” can mean several different things.
Did the human confirm the plate characters? Confirm the vehicle make and color? Check the timestamp? Check whether the location is physically compatible with the alleged event? Confirm the hotlist record is current? Compare contradictory evidence?
A useful verification protocol names the checks rather than relying on the existence of a person.
The system should surface contradictions aggressively.
If a vehicle appears at two geographically incompatible places inside an impossible travel window, software can flag that. If a search result predates an alleged pursuit, the interface can highlight the chronology. If a match is based on vehicle characteristics rather than an exact plate, the result can make that distinction visually unavoidable.
The best safety feature is not merely a confidence score. It is a system that makes contradictory facts harder to overlook.
Do not overclaim the case.
The public reporting describes multiple pieces of evidence in Parra's arrest. It would be inaccurate to say Flock alone caused the arrest. It would also be inaccurate to dismiss the ALPR result as irrelevant when police used it as corroboration.
The value of the case is narrower and more important: it shows how a machine-generated investigative result can acquire authority inside a human evidence chain even when the record deserves a more skeptical read.
Cyberdelia is not claiming the ALPR result was the sole cause of Parra's arrest or that every Flock hit is erroneous. The article examines how machine evidence can be misinterpreted inside a multi-source police workflow.