An officer turning into a frog got corrected because frogs are impossible [12][13], but the case offers little comfort beyond that one obvious catch. A model drafting narrative from body camera audio produces errors spread across a range of plausibility, and the ones written in ordinary policing prose are the ones an officer skimming a shift's worth of drafts will wave through. The Utah report failed loudly. A fabricated sentence placing a hand near a waistband would fail quietly. It would sit in a file that gets read at trial months or years after the incident [11].
To weigh the frog case you need a denominator, and the source supplies none. It is a single-authored argument published by phys.org rather than an audit: no count of how many drafts the RCMP pilot produced, no measured rate at which officers altered them, no error rate of any kind [20][10]. Without those figures, the frog is an anecdote from one tail of a distribution nobody has characterised.
Which is why the exposure is evidentiary before it is about accuracy. Defence counsel does not have to demonstrate a specific fabrication. Asking whether an AI tool touched a report, and what review it then received, is enough, and no accuracy figure repairs an agency that cannot answer. Answering is a records problem. According to the phys.org piece, there is no uniform model for AI adoption across Canadian police services [17], and some services have declined to discuss their AI policies with media organisations [18]. Police in the United States and Canada have also officially published AI-generated images of drug seizures [6], which places synthetic material in the public record under a police letterhead. It is also unknown whether body camera material collected this way will train later systems [16], and the people recorded can neither opt out nor consent [15].
The misidentification cases run on a different mechanism and are worth keeping separate. A facial recognition hit is not a drafted narrative but a ranked list a human treats as a lead. The North Dakota match, per the same account, kept a Tennessee grandmother in custody for more than five months, upward of 150 days [8][21]; a Maryland school surveillance system flagged a bag of chips and an officer drew his gun [9]. The author's claims about demographic bias arrive without figures attached [7][4]. I would not lean on them, because the evidentiary argument does not require them: a tool with no demographic skew that fabricates at a low rate creates the same disclosure duty.
The piece argues that selective police use of AI is inherently biased and must cease immediately [19]. That is a policy position, and it reaches past the evidence in front of it. The narrower requirement I would defend is dull, and cheap next to one retrial: per-report provenance, a logged difference between machine draft and officer-signed final, and a published correction rate from the pilots that have already run [10][14]. If those logs showed officers materially rewriting drafts, the courtroom exposure would shrink to something arguable, and the agencies holding the logs are the only parties who can settle it.
