An English police force has taken an officer off frontline duties and opened a criminal investigation into the alleged use of AI to create evidential material in a number of cases . Derbyshire Constabulary calls the allegation perverting the course of justice. No arrests have been made. The Crown Prosecution Service, which decides what gets charged in England and Wales, says it is working with defense teams and the courts on the cases that might be affected. It is the first known case of its kind in UK criminal justice.

The case will turn on what one officer intended. The problem it exposes is bigger than one officer, but it has a simple fix.

The Original Recording Is The Real Evidence

The original transcription source, the body-worn camera audio, the recorded interview, the raw notes, is the evidence. When a model turns that source into a clean written statement, the statement is a derivative. A copy that has been through a process you cannot see. If the source is preserved and stays in the chain of custody, you can play it back and check the AI's version against what was actually said. If the source is gone, the derivative is all you have, and it proves only itself.

This is old digital forensic doctrine in new clothes. Keep the original. Work from copies. Go back to the original to settle any dispute. AI does not change the rule. It makes the rule load-bearing, because the copy now writes itself and can drift from the source without anyone touching it dishonestly.

AI Cannot Be Cross-Examined - A Human Must Answer

Preservation is half of it. The other half is a person. Evidence in a criminal case has to be sponsored by a human who can swear it is accurate and answer for it under cross-examination. The defense gets to ask the officer a plain question. Did the witness say this, in these words, or did the software write it this way? The officer has to answer, under oath, with the recording there to check. A model cannot be sworn, cannot be questioned, and cannot be held to account. Take the human out of that loop and you are left with a document no one can stand behind.

Without The Source, AI Error Looks Identical To AI Fabrication

Without the original to check against, an honest mistake and a deliberate fabrication look identical on the page. A model that smooths a witness’s hesitation into a flat assertion, and an officer who edits a quote to mean something it never did, produce the same clean paragraph. The sloppy and the corrupt become impossible to tell apart. The preserved source is what separates them, and it is the same fix for both.

The most widely used AI report-writing tool for U.S. police is Axon’s Draft One — a different workflow than the one alleged in Derbyshire, but with its own audit problem. When an officer exports a finished report, the system erases the initial draft When an officer exports a finished report, the system erases the initial draft , and with it the record of which words the AI wrote and which the officer added. The Electronic Frontier Foundation, or EFF, a civil liberties group that reviewed the tool, found agencies frequently cannot even tell which reports were AI-assisted. The product removes the audit trail as a feature. That is the opposite of preservation. It is designed deletion of the one artifact that could check the output.

These reports already misfire in the open. One AI-generated police report had an officer transforming into a frog , the kind of error obvious enough to catch on the way out the door. The errors that matter are the plausible ones, and those are exactly the ones the deleted draft would have let you check.

Police already have a live example of a useful tool turning into a liability because a minority abused it. Flock Safety runs a national network of automated license-plate-reader cameras that thousands of departments tap. The capability is real. So is the record of misuse. EFF documented searches that tracked protesters, hundreds of lookups run with racial slurs, and queries aimed at people seeking reproductive care. An audit in one California county found a vendor error re-exposed local data to out-of-state agencies hundreds of thousands of times, with some searches citing immigration enforcement. Individual officers have used the cameras to stalk romantic interests .

Most departments use Flock within policy. It did not matter. The abuse by a few produced lawsuits, canceled contracts, statewide restrictions, and a public that now treats the whole network as suspect. A handful of bad actors set the terms for everyone.

AI-written evidence is standing in the same spot. A few officers fabricating with AI, plus a larger number of honest errors no one can catch because the source was not kept, is enough to make judges and juries distrust every AI-assisted statement, including the accurate ones. The officer in Derbyshire, if the allegation holds, is more than a problem for his own cases. He is a preview of the bill the whole system pays when the check is missing.

AI Evidence Laws And Federal Rule 707 Are Arriving Late

U.S. prosecutors and lawmakers are moving, slowly. The King County prosecuting attorney in Washington state refused to accept police narratives produced with AI. California became the second state, after Utah, to require written disclosure when AI drafts a police report . The federal judiciary has a proposed Federal Rule of Evidence 707 working through approval, meant to hold machine-generated evidence to the same reliability standard as expert testimony before a jury sees it. Disclosure rules and reliability standards help. Neither one brings back a source that has already been deleted.

Every case that Derbyshire officer touched now has to be pulled apart, because no one can replay an original and hear which words were the witness's. Multiply that by every department running interviews through AI and discarding the source on the way out. The records keep looking authoritative. The proof of what they originally said keeps disappearing. Keep the original recording, put a human who can be cross-examined next to it, and most of this risk goes away. Skip that, and a few bad actors get to decide how much anyone can trust the rest.