A UK Police Officer Is Under Criminal Investigation for AI-Fabricated Evidence. The Failure Pattern Isn't Unique to Policing.

The line between "AI-assisted" and "AI-fabricated" is something most professionals believe they're on the right side of. Few have a concrete method for confirming it.

That distinction just produced the UK's first criminal investigation for AI use in evidence. On June 14, a Derbyshire police officer was removed from duty and placed under criminal investigation for allegedly using AI to fabricate evidential material in multiple cases. The Crown Prosecution Service is reviewing the affected cases and notifying defence teams. The HN thread discussing it reached 364 points. It is believed to be the first criminal case of its kind in UK criminal justice.

The officer's situation involves alleged criminal intent. The failure pattern it exposes doesn't require intent at all.

What AI fabrication actually looks like

When an AI generates a quote, a summary, or a specific fact, it produces text that reads like it came from somewhere. The formatting is appropriate. The phrasing fits the context. The claim is plausible. What's often missing is the source — because the AI didn't retrieve one. It produced text that looked retrieved.

That's AI fabrication: not random nonsense, but coherent, contextually appropriate content that doesn't trace back to a real prior source. It's designed — not deliberately, but structurally — to pass a reading-for-sense check. The only check it reliably fails is a source check.

The same pattern appears in professional work. AI produces a statistic that sounds right but doesn't trace to a real study. AI summarizes a customer complaint and attributes a detail that wasn't in the original. AI formats a quote in a way that's close but inaccurate. The professional reads it, finds it reasonable, and includes it in the deliverable. The difference between the police case and the professional case isn't the failure. It's what happens when the failure is found.

What to do differently Monday morning

Any specific claim that appears in a formal document — a quote, a statistic, an attributed fact, a cited incident — needs to be traceable to a source that existed before your AI prompt. Not plausible. Not approximately right. Traceable.

If you can't find the source in 90 seconds, you have a fabrication risk, not a fact. That's the rule, regardless of how confident the AI sounds when it produces the claim.

The check: find the original document, not a secondary reference. Pull the statistic from the study that produced it, not from a summary. Verify the quote exists in the source before you attribute it. If you can't close that loop, the claim doesn't go into the document.

Sarah: the strategy consultant who almost published it

Sarah works at a 90-person management consultancy. Her team uses AI to pull supporting evidence for client recommendations — industry statistics, competitor examples, market sizing figures to include in board presentations.

She was building a section for a financial services client when the AI returned what looked like a perfect data point: a specific percentage figure attributed to a named industry association, with a plausible year. She drafted it into the slide deck. When she went to add the citation link, she couldn't find the source. The association's publications page had no matching report. The figure didn't appear to exist.

The AI had produced the number in the correct format for that kind of claim — specific, attributed, credible-sounding. It passed a read-aloud test. It failed a source check.

Her process now: any figure that appears specific gets a 30-second source check before it's included. Find the original document, not just a secondary reference to it. If the source can't be found, the figure goes back to the AI with a request for an alternative that includes a real citation. Claims that can't be sourced don't make the final version.

Marcus: the HR director who caught the contradiction

Marcus is HR director at a 220-person logistics company. His team uses AI to draft formal HR documentation — first-draft summaries of disciplinary incidents, meeting notes converted into written records, template letters completed with case specifics.

After the police investigation story came across his feed, he pulled three finalized documents from the previous month. In one — a formal written warning — the AI had produced a sentence summarizing the employee's stated explanation that slightly but materially misrepresented what was in the meeting notes. The draft was internally consistent. The sentence didn't match the source.

That's a legal exposure. The document had been signed off because it read correctly, not because anyone had checked the attributed statement against the original record.

The process change: any document that attributes a specific statement to an employee gets that statement verified against the original meeting record before the document is finalized. Not read for plausibility. Cross-checked against the source.

The practical rule

When AI generates content that looks sourced — specific figures, direct quotes, attributed statements — the question is not "does this sound right?" It's "can I find where this came from?"

If the answer is no, you have two options: find the source and include it, or cut the claim. Submitting AI-generated content as verified fact when you can't verify it is how the line gets crossed — whether the crossing is deliberate or not.