AI-written documents keep showing up in places that carry legal weight. In November 2025, a federal judge raised concerns after reviewing evidence that immigration enforcement agents used ChatGPT to help write use-of-force reports tied to enforcement actions in the Chicago area. The story made headlines because of who was involved. But the underlying issue — AI generating a confident, polished narrative from limited input — is the same issue that's been showing up in law firms for two years.
What Happened
U.S. District Judge Sara Ellis included a footnote in a 223-page opinion describing what she saw while reviewing body camera footage from immigration enforcement operations. According to her opinion and multiple news reports, at least one agent used ChatGPT to help compile the narrative portion of a use-of-force report. Judge Ellis wrote that this practice may help explain why some of the reports she reviewed didn't match what the body camera footage actually showed.
The Example From the Ruling
Based on the body camera footage described in the opinion, the process looked like this: an agent gave ChatGPT a short description of an incident, along with a few images, and asked it to compile that into a narrative for the report. ChatGPT produced a written account based on what it was given — but a written account is only as accurate as the description and images fed into it, and it can't independently verify what actually happened.
Judge Ellis found discrepancies between the official narrative and the body camera footage itself. The concern wasn't just that AI was used — it was that AI was used to write a first-person account of an event the AI never observed, based on a secondhand summary.
Two Separate Problems: Accuracy and Privacy
The ruling points to two issues that are worth separating.
- →Accuracy. AI tools are good at producing fluent, confident-sounding text from whatever input they're given. They are not good at catching the difference between 'this is what happened' and 'this is what I was told happened.' If the input description leaves something out or gets a detail wrong, the AI output will too — and it will read just as authoritative either way.
- →Privacy. If the agent used a public, consumer version of ChatGPT, any images or sensitive details from the incident may have been uploaded to a third-party server. Reporting on the story noted that it was unclear whether the relevant agency had any policy on AI use by agents at all.
This Is an AI Accuracy Story, Not Just an Immigration Story
Strip away the subject matter, and this is the same problem that's been showing up in courtrooms since 2023: someone treats AI output as a finished product instead of a draft, and nobody catches the error before it matters. Stanford's RegLab has found that general AI models hallucinate in roughly one out of three legal queries. Multiple attorneys have been sanctioned for filing briefs with AI-generated citations to cases that don't exist.
A use-of-force report is a different kind of document than a legal brief, but the failure mode is identical: AI generates something that sounds right, and the person responsible for it doesn't verify it against the underlying facts before it becomes part of an official record.
What This Means If You're a Lawyer
Use-of-force reports are routinely used as evidence — in civil rights litigation, immigration detention challenges, and criminal proceedings. This ruling adds a new question to the list of things worth asking about any official report: was this document drafted with AI, and if so, what was the AI given to work with?
- →In discovery, it may now be reasonable to ask whether a report was AI-assisted, and what input (text, images, prompts) was used to generate it.
- →In cross-examination, a discrepancy between a report and other evidence (like body camera footage) can be explored through how the report was actually written.
- →For your own work, the same standard applies. If you use AI to draft a summary, declaration, or report based on a client's account of events, the AI's output reflects what it was told — not what happened. Read our guide on free AI for legal research for more on where AI tools are reliable and where they aren't.
Frequently Asked Questions
What did the judge actually say?
In a footnote to a 223-page opinion, U.S. District Judge Sara Ellis noted that at least one immigration enforcement agent appeared to use ChatGPT to help write the narrative portion of a use-of-force report, and that this practice may help explain inaccuracies she found when comparing reports to body camera footage.
Is it illegal for government agents to use AI to write reports?
That depends on the agency's policies, which weren't clear in this case — reporting noted that it was unclear whether the agency involved had any guidelines on AI use by agents. The legal issue raised by the judge wasn't that AI was used, but that the resulting reports didn't match the available video evidence.
Does this affect the underlying case?
The footnote was part of a larger opinion addressing the immigration enforcement operations themselves. The AI use was one detail the judge highlighted as a possible explanation for report inaccuracies she'd already identified by comparing reports to footage.
What should attorneys take away from this?
Any AI-generated document — a report, a summary, a declaration, a brief — is a draft based on whatever input it received. Before it becomes part of an official record, someone needs to check it against the actual facts, not just check that it reads well.
What free AI tools get right — and where you still need to verify the output
Editorial note: AI For Legal Research publishes independent content. We do not accept payment for editorial coverage or review scores. Nothing on this site constitutes legal advice. Always consult a qualified attorney for legal matters.