Adongo case shows AI risk asymmetry for immigration lawyers

Daniel Adongo’s deportation is not, on the available record, an AI hallucination case. That boundary matters. The former Indianapolis Colts linebacker’s removal has become a useful legal-analysis entry point for immigration lawyers because it moved through a system now governed by EOIR’s attorney-facing AI policy, while the government’s own AI use is expanding with much less case-level visibility.

The verified timeline is narrow. ICE arrested Adongo in September 2025; an immigration judge ordered him removed on March 23, 2026; ICE deported him on June 20, 2026. Reporting has also tied the case to mandatory detention under the Laken Riley Act, and separate reporting has described mental-health concerns, including a schizophrenia diagnosis and suspected CTE reflected in family accounts and legal filings, not in a published medical determination.[1][2][3][4]

That is enough context to explain why immigration counsel are paying attention. It is not enough to claim AI tainted the Adongo proceeding. The AI implication comes from the surrounding legal architecture: private lawyers are now expressly warned that AI-assisted filings remain sanctionable if they contain false authority, while agency-side automation is described in strategy plans and inventories rather than exposed in each respondent’s record.

The EOIR Policy Does Not Ban AI, But It Leaves Counsel Holding the Risk

EOIR Policy Memorandum 25-40, issued August 8, 2025, is measured in one respect and unforgiving in another. It does not impose a blanket prohibition on generative AI use. It also does not create a mandatory disclosure rule for AI-assisted filings. But it makes the traditional duty unmistakable: the practitioner remains responsible for the filing, including fabricated citations, false factual assertions, and material offered knowingly or with reckless disregard for its falsity.[5]

For removal-defense lawyers, that is not an abstract ethics reminder. A detained client may have days, not months, to preserve a claim. A lawyer may be working across incomplete records, translation problems, prior counsel issues, and rapidly changing custody facts. If a generative tool invents a case citation in that environment, the court does not sanction the tool. It sanctions the lawyer.

PM 25-40 therefore creates a practical floor: AI may be used, but no lawyer can outsource verification. The absence of a mandatory disclosure rule should not be mistaken for leniency. It means the court may not know AI was used unless something goes wrong, and when something goes wrong the professional-responsibility analysis begins with the lawyer’s signature.

Agency AI Is Visible in Plans, Not Necessarily in the Case File

The harder problem is not that EOIR warned lawyers about hallucinations. It should have. Fabricated authority burns court time and client time, and in immigration court both are scarce. The harder problem is that the warning lands in a system where government-side AI use is being developed at the institutional level without an equivalent per-case disclosure burden.

Analyses of EOIR’s 2024 Strategic Workforce Plan describe contemplated AI use in transcription, translation, and drafting-related work, including opinions and orders.[6][7] Those functions are not peripheral in removal proceedings. Transcription affects what the reviewing body can see. Translation affects what the record appears to say. Drafting assistance affects the language through which credibility, burden, relief, and discretion are memorialized.

DHS materials point in the same direction on the enforcement side. An American Immigration Council analysis identified 105 active AI use cases across immigration agencies, including 27 classified as rights-impacting. A June 2026 DHS inventory update listed 235 total AI use cases, 200 active use cases, and 109 deployed use cases.[8][9] The counts vary by reporting date and classification method, so they should not be collapsed into one fixed number. They do show scale.

That is the asymmetry the Adongo matter exposes, even without an AI allegation. A respondent’s lawyer files in a sanctionable, reviewable posture. The agency’s AI use may be documented somewhere else: in an inventory, a strategic plan, a use-case library, or an implementation document that the respondent may never see in the individual adjudicatory record.

This distinction is procedural, not ideological. A court can discipline a lawyer for a hallucinated case because the lawyer’s name is on the brief and the false citation is in the record. A party trying to understand whether AI affected a government transcript, translation, draft order, case triage decision, or enforcement-support system may have to work from public inventories that were not designed to litigate the reliability of a particular case event.

Why Lnu v. Blanche Is the Risk Ceiling Immigration Lawyers Can No Longer Treat as Theoretical

The Ninth Circuit’s June 3, 2026 decision in Lnu v. Blanche gives immigration lawyers the concrete endpoint that PM 25-40 only foreshadows. Counsel in an asylum matter filed briefs containing fabricated case citations generated by AI. The panel did not treat that as a harmless technology mistake. It imposed monetary sanctions, suspended counsel from Ninth Circuit practice for six months, referred the matter to the California State Bar, and required Sethi Law Group to disclose AI use and certify citation verification in future filings.[10]

The opinion’s language is blunt because the failure was blunt. The panel wrote that “fabrications are the most notorious hallucinations,” a sentence that will now sit behind every AI-assisted asylum brief filed in the circuit.[10] It is also the phrase that risk managers should take seriously: the court was not disciplining mere use of a tool; it was disciplining the submission of legal authority that did not exist.

The certification remedy matters as much as the suspension. Monetary sanctions punish the event that already happened. A bar referral escalates professional consequences. A six-month practice suspension directly affects livelihood and client service. But a forward-looking AI-use disclosure and citation-verification certification changes the firm’s filing posture going forward. It turns a technology mistake into a continuing compliance condition.

The Ninth Circuit also judicially noticed research from Stanford RegLab and the Stanford Institute for Human-Centered Artificial Intelligence, including reported hallucination rates of 17% for Westlaw AI and 33% for Lexis AI in the cited study context.[10] Those figures should not be stretched into a universal product ranking. Their function in Lnu was narrower and more important for practitioners: the panel treated hallucination risk as known enough that lawyers could not plausibly claim surprise when AI-generated authorities proved false.

A law-firm analysis of the decision reached the same practical point: the sanction package makes AI verification a litigation-control issue, not a style preference.[11] In immigration work, where a brief may carry the only preserved argument for withholding, CAT protection, reopening, or review of an adverse credibility finding, the professional exposure and client prejudice move together.

The Scale Evidence Is Real, With Uneven Source Strength

The broader sanctions landscape reinforces the point, but the sources are not all equal. The most useful current tracker is the Charlotin AI Hallucination Cases Database, which reported 1,782 total documented hallucination cases globally and more than 1,228 U.S. lawyer hallucination cases as of its July 18, 2026 update.[12] That is a database count, not an official judiciary census, but it is more systematic than anecdote.

The often-repeated figure that more than 40 AI-related attorney sanctions occurred in the first six weeks of 2026 is best treated as a reported signal from Thomson Reuters, not as an official national tally.[13] It still matters because it captures the professional mood: courts are no longer encountering AI hallucinations as rare curiosities.

Other reported sanction amounts show that the financial side can be substantial. Norton Rose Fulbright’s AI-in-litigation update identified a $60,000 sanction in an Illinois circuit court in December 2025 and a $29,877 sanction in a Texas bankruptcy matter in July 2026.[14] Those are not immigration cases and should not be treated as a removal-practice average. They are useful outer markers for what courts are willing to do when AI-assisted litigation failures consume adversary and judicial resources.

What the Adongo Case Actually Signals

A search for AI in the Adongo case itself leads to the wrong question. The better question is what his fast-moving removal timeline reveals about counsel exposure in the current system. Mandatory detention compresses communication. Removal defense already operates inside a heavy backlog. EOIR’s policy tells attorneys they remain fully answerable for AI-assisted filings. Lnu shows that the answer can include money, suspension, bar referral, and future certification. Meanwhile, agency AI use is disclosed mostly through public program-level materials rather than a matching obligation in each case record.

That combination is the legal significance of the former NFL player’s deportation by ICE. Not because Adongo’s case proves AI error. It does not. The significance is that the case arrived at the same time immigration lawyers were being placed under an explicit AI-accountability regime, while the institutional actors around them continued to expand AI uses through less individualized forms of transparency.

For counsel, the lesson is not to avoid every AI tool. The record does not support that kind of blanket conclusion, and EOIR has not adopted it. The lesson is narrower and harsher: if AI touches a filing, the lawyer must be able to verify the authorities, facts, quotations, procedural history, and record citations at a level that would survive the Lnu certification standard. If that sounds like ordinary lawyering, it is. The difference is that generative tools can produce false legal materials with enough fluency to make ordinary lawyering easier to fake and harder to audit after midnight.

The unevenness remains. Immigration lawyers in high-volume removal practice now operate under a visible discipline regime for AI-assisted mistakes. The agency may expand AI use in transcription, translation, drafting-related functions, and enforcement-support systems without the same per-case disclosure burden. Lnu supplies the sanction ceiling that makes the lawyer-side risk concrete. Adongo supplies the procedural setting that makes the asymmetry hard to ignore.

References

1. Exclusive: Former NFL player deported by ICE, Fox News
2. Former NFL player Daniel Adongo deported, NewsNation
3. Daniel Adongo mental-health investigative report, Indianapolis Star
4. Daniel Adongo deportation coverage, Newsweek
5. Policy Memorandum 25-40: Use of Generative Artificial Intelligence in EOIR Proceedings, U.S. Department of Justice Executive Office for Immigration Review, August 8, 2025
6. EOIR Strategic Workforce Plan analysis, Hoppock Law Firm
7. EOIR Strategic Workforce Plan analysis, Austin Kocher Substack
8. AI Use in Immigration Enforcement, American Immigration Council
9. AI Use Case Inventory Library, U.S. Department of Homeland Security, June 2026
10. Lnu v. Blanche, No. 24-4790, U.S. Court of Appeals for the Ninth Circuit, June 3, 2026
11. Ninth Circuit sanctions attorneys for AI-generated hallucinated citations, Sabrina Damast Law Firm
12. AI Hallucination Cases Database, Damien Charlotin, July 18, 2026
13. AI-related attorney sanctions social post, Thomson Reuters, 2026
14. AI-in-litigation update, Norton Rose Fulbright, 2026