UK Police AI Evidence Case Exposes Chain-of-Custody Crisis

The first criminal investigation

Derbyshire Constabulary in England has removed an officer from frontline duty and opened a criminal investigation into alleged use of AI to create evidential material across multiple cases. The force is treating the matter as potential perverting of the course of justice, according to reporting by Forbes contributor Lars Daniel. No arrests have been made, but the Crown Prosecution Service is now working with defense teams to assess which cases may be compromised. It marks the first known criminal probe of its kind in UK law enforcement.

While the investigation centers on one officer's alleged conduct, the case illuminates a structural vulnerability in how police departments are adopting AI transcription and report-writing tools.

Why it matters

This case could establish legal precedent for how courts treat AI-generated police evidence across common-law jurisdictions. More immediately, it demonstrates that a handful of misuse incidents—whether through deliberate fabrication or uncaught errors—can undermine public confidence in an entire category of law enforcement technology. Departments that have already deployed AI writing tools without preserving original source recordings now face potential challenges to evidence in past cases.

The chain-of-custody problem

The original audio recording—from a body-worn camera, interview room, or field notes—is the actual evidence. When an AI model converts that source into a written statement, it produces a derivative. If the original recording remains in the evidence chain, courts can verify the AI's interpretation. If the source is deleted, the derivative becomes the only record, and no independent verification is possible.

This principle comes from established digital forensics practice: preserve the original, work from copies, return to the source to resolve disputes. AI transcription does not change the rule, but it does raise the stakes. The derivative now generates itself through an opaque process, and can diverge from the source without any human intervention.

The cross-examination gap

Criminal evidence requires a human sponsor who can testify under oath and face cross-examination. Defense counsel must be able to ask an officer directly whether a witness used specific words or whether the software paraphrased. An AI model cannot be sworn in, questioned, or held accountable. Without a preserved recording and an officer who can vouch for its accuracy, the evidence loses its foundation.

When the original source is missing, honest transcription errors become indistinguishable from deliberate alterations. A model that converts a witness's hesitant statement into a confident assertion looks identical on paper to an officer who manually edits testimony to support a preferred narrative.

Built-in deletion

Axon's Draft One, the most widely deployed AI report-writing tool for U.S. police departments, deletes the initial AI-generated draft when an officer exports the final report. The Electronic Frontier Foundation reviewed the system and found that many agencies cannot even identify which reports received AI assistance. The tool erases the audit trail by design, removing the one artifact that could verify which words came from the AI and which from the officer.

Public examples of AI report failures include an instance where generated text described an officer transforming into a frog—an error obvious enough to catch. The concerning errors are the plausible ones that slip through, precisely the cases where a preserved draft would enable verification.

The Flock Safety precedent

Law enforcement has recent experience with how individual misuse can poison an entire technology's reputation. Flock Safety operates a national automated license-plate-reader network used by thousands of departments. The EFF documented searches targeting protesters, hundreds of queries containing racial slurs, and lookups tracking people seeking reproductive healthcare. One California county audit revealed vendor errors that re-exposed local data to out-of-state agencies hundreds of thousands of times. Officers have used the cameras to stalk romantic partners.

Most departments use Flock within policy, but the minority abuse generated lawsuits, contract cancellations, and statewide restrictions. A few bad actors defined public perception of the entire system.

AI-generated evidence now occupies the same position. A small number of fabrication cases, combined with undetectable honest errors, will lead courts and juries to distrust all AI-assisted statements—including accurate ones.

Regulatory response

U.S. jurisdictions are beginning to respond. King County, Washington's prosecuting attorney has refused to accept AI-produced police narratives. California and Utah now require written disclosure when AI drafts a report. The federal judiciary is considering proposed Federal Rule of Evidence 707, which would subject machine-generated evidence to the same reliability standards as expert testimony before reaching a jury.

These measures help, but disclosure rules cannot recover a source recording that was deleted before anyone knew to preserve it. Every case the Derbyshire officer touched now requires review because no original recording exists to verify the written statements.

The fix remains straightforward: retain the original recording, ensure a human officer can testify to its accuracy under cross-examination, and maintain the audit trail showing how the AI transformed source to output. Without those safeguards, a minority of bad actors will determine whether anyone can trust the majority of legitimate AI-assisted evidence.

Details of the Derbyshire investigation were first reported by Lars Daniel in Forbes.