Judge Warns DOJ Over AI Hallucination in ICE Detention Filing

A quick correction comes first: the search phrase “Lorenzo Thompson ICE detention asylum case update” is mixing two different immigration stories. Thompson’s reported July 14 ICE detention and pending asylum matter is not the AI-citation incident. The court warning over a nonexistent case arose two days later in Daghra v. Hinkley, an ICE detention habeas proceeding in federal court.

The verified record for this Risk Digest entry is narrower and more useful: in Daghra v. Hinkley, Chief Judge Hala Y. Jarbou found that the government’s filing cited “Taylor v. Hott, 724 F. App’x 387, 392 (6th Cir. 2018),” a case the court concluded “does not exist.” The order language available for review comes through a verbatim reproduction published by the Volokh Conspiracy at Reason, rather than a machine-readable PACER order reviewed directly here.[1]

The Short Record

Bloomberg Law reported that the government filing was made by Assistant U.S. Attorney Carolyn Almassian, and identified Alissa Heynen of the Michigan Immigrant Rights Center as counsel for the petitioner.[2] Those names matter for recordkeeping, but they are not the center of the incident. The operative fact is what the judge did with the citation once it appeared in a filing asking the court to reject habeas bond relief.

What The Citation Was Supposed To Do

The government cited Taylor v. Hott for a jurisdictional proposition in its response to a habeas petition brought by an ICE detainee. That is the place where citation checking stops being a housekeeping chore. A detained person was asking a federal court to intervene over bond-related custody. The government was opposing relief. If authority is used to narrow the court’s role, delay release, or justify continued custody, it has to be traceable.

Judge Jarbou’s order did not treat the problem as a misspelled party name or a bad pincite. The court wrote that “the case does not exist,” and added that the citation was “likely produced by generative artificial intelligence.” The order then connected that failure to the duty of candor: “A lawyer may use generative AI, but the duty of candor towards this tribunal demands no less than scrupulous review of the output before incorporating it into a court filing.”[1]

That language is doing two things at once. It does not ban AI-assisted drafting. It does not announce that every AI error will draw discipline. But it does remove the easy excuse: if generative AI is used, the lawyer still owns the citation that reaches the docket.

Why This Was Not A Harmless Reporter Mistake

The citation looked plausible in the way hallucinated legal citations often look plausible: a familiar party format, a Federal Appendix volume, a page number, a circuit, a year. That is precisely the danger. A busy reader may register the citation as ordinary Sixth Circuit support unless someone checks whether the case exists.

Law360 reported the more concrete problem: the real case at 724 F. App’x 387 is Atkins v. CGI Techs. & Sols., not Taylor v. Hott. And Atkins was a commercial arbitration matter, not an immigration bond case.[3] That means the citation did not merely point to the wrong page of a relevant immigration decision. It opened onto a different legal universe.

This is the part of AI-hallucination risk that gets flattened when the shorthand becomes “fake cite.” In litigation, a citation is a route. It tells the judge and opposing counsel where the asserted rule can be found, what court said it, and how closely that rule maps onto the dispute. When the route leads to a commercial arbitration case instead of the immigration authority represented in the brief, the filing has not just made a clerical error. It has required the court and the other side to spend attention clearing debris from the record.

The Court Warned, But Did Not Sanction

The docket outcome is easy to overstate if the politics of the case are allowed to outrun the order. Judge Jarbou did not impose sanctions. Reuters framed the ruling accurately as a warning to the Justice Department about AI use in an immigration case, not as a penalty order.[4]

The reason matters. According to the reproduced order, the court declined to sanction the government because the respondent had not moved for sanctions. That procedural posture kept the remedy at warning level. The warning is therefore the operative court response, not a softened version of some hidden punishment.[1]

The habeas petition itself was dismissed as moot for reasons separate from the AI citation. The automatic stay had expired, and the immigration judge had already set bond at $35,000.[1] That is not a finding that the hallucinated citation was harmless in any broader professional sense. It only explains why the habeas petition did not proceed as a live request for relief.

So the precise answer to the central risk question is this: when no sanctions motion was before the court, an AI-hallucinated government citation in this ICE detention habeas dispute triggered an express judicial warning tied to the duty of candor, not monetary sanctions, suspension, referral, or other discipline.

What Public Framing Adds, And What It Does Not

Newsweek’s headline placed the episode in the “Trump DOJ” frame, and Reuters also noted the administration’s immigration-policy context.[4][5] That context explains why the story traveled quickly. It does not change the narrower docket event: a federal judge identified a nonexistent citation in a government habeas filing, described it as likely AI-generated, declined sanctions, and issued a warning.

The distinction is not pedantic. Immigration detention filings move under pressure, and the person in custody bears the cost of delay and confusion. But a useful record has to keep separate the detention policy fight, the habeas mootness ruling, the bond amount, and the professional-responsibility warning. Once those are blurred, the next lawyer trying to cite the incident will have inherited the same problem in a new form.

Risk Context: Warning Here, Sanctions Elsewhere

Daghra sits inside a larger AI-litigation-error pattern, but it should not be inflated into the worst example in that pattern. Norton Rose Fulbright’s June 2026 update counted more than 1,148 U.S. AI hallucination cases, a figure useful as background risk context rather than proof of anything about this particular AUSA’s drafting process beyond what the order found.[6]

A sharper comparison is Lnu v. Blanche. Bloomberg Law reported that the Ninth Circuit imposed sanctions of $2,500 and a six-month suspension on immigration attorneys in an order warning about AI-hallucinated briefs.[7] That contrast matters because it shows the Daghra outcome was not inevitable leniency across immigration-related AI citation errors. Different posture, different conduct record, different remedy.

The Litigation Signal

Daghra is not a sanctions case. It is also not a nothing case.

The court put a specific warning on the record: a lawyer may use generative AI, but the lawyer must scrupulously review the output before filing it. In a detention dispute, that is a modest sentence with a hard edge. The government asked the court to accept legal authority against a person seeking custody-related relief; the authority did not exist.

The Lorenzo Thompson story remains separate from this record. Daghra v. Hinkley is the AI-hallucinated citation incident, and its risk signal is confined but clear: the absence of sanctions this time depended on the posture before the judge, not on any judicial tolerance for fake legal authority in government filings.

References

1. Judge Faults Federal Government Lawyer for Apparent AI Hallucination, Volokh Conspiracy / Reason.com, July 17, 2026
2. DOJ Called Out by Judge for Alleged AI Use, Fake Case Citation, Bloomberg Law, July 16, 2026
3. Judge Warns Gov't Over Apparent AI-Hallucinated Citation, Law360, July 16, 2026
4. US judge warns Justice Department about AI use in immigration case, Reuters, July 16, 2026
5. Judge Calls Out Trump DOJ Over 'Hallucinated' Citation, Newsweek, July 2026
6. AI in litigation: Update on Gen AI sanctions in 2026, Norton Rose Fulbright, June 2026
7. Ninth Circuit Warns of AI Hallucinated Briefs in Sanctions Order, Bloomberg Law, June 3, 2026