AI in Immigration Practice: What Attorneys Are Actually Required to Do

AI adoption is accelerating across the legal profession, including immigration practice, but governance has not kept pace. Informal use is widespread, but formal, written AI-use policies remain the exception rather than the rule at most firms — which means the overwhelming majority of the profession is using these tools with no documented answer to the basic questions: what can go in, what needs a client's sign-off, who's liable when the output is wrong.

Those questions now have real answers. The ABA, several state bars, and EOIR have now begun translating those existing duties into AI-specific guidance. This piece pulls together what the actual rules say, what the reliability research actually found, and what that adds up to in practice — reading the primary documents directly rather than working from summaries, since on a subject this consequential the gap between "someone paraphrased it" and "here's the sentence" tends to matter.

What the ethics rules actually say

The ABA's first opinion on generative AI, Formal Opinion 512, sets the national baseline by tying five existing duties to specific AI conduct.¹ On competence, it doesn't ask lawyers to become engineers — just to have "a reasonable understanding of the capabilities and limitations of the specific GAI technology" — but it's blunt about what happens without that understanding: "uncritical reliance on content created by a GAI tool can result in inaccurate legal advice," and the lawyer remains fully responsible regardless. On confidentiality, the opinion requires informed client consent before inputting case information into a self-learning tool, and it specifically closes a loophole a lot of firms are currently relying on: "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient." If your engagement letter has a generic AI clause and nothing more, it likely doesn't clear the ABA's own bar. On candor, the opinion is equally direct — AI output "must be carefully reviewed to ensure that the assertions made to the court are not false," and even an unintentional misstatement can trigger a misrepresentation finding under Rule 8.4(c). On fees, it gives a concrete example: fifteen minutes spent prompting a tool, plus the time spent reviewing the draft, is billable; time "saved" is not, and neither is the time a lawyer spends learning to use the tool in the first place.

EOIR wrote its own memo for immigration proceedings specifically, and it's thinner than most practitioners assume.² There's no blanket ban on attorneys using AI, and — more surprising — no mandatory national requirement to disclose that you used it. What EOIR did keep is a disciplinary hook: submitting hallucinated or fabricated AI content exposes an attorney to sanctions under 8 C.F.R. § 1003.102(c) for "knowingly or with reckless disregard offering false evidence," with the memo pointing to ABA 512 as persuasive authority on the underlying standard. Individual immigration judges remain free to impose their own disclosure requirements through standing orders, so "EOIR doesn't require it nationally" isn't the same as "no judge in front of me will ask." The memo is explicitly a placeholder — EOIR said it expected the Department of Justice to issue department-wide generative AI guidance by roughly the end of 2025, and as of this writing there's no public confirmation that guidance has actually been issued, which is worth checking before treating PM 25-40 as the final word.

State bars have moved faster and further than the ABA in a few notable ways. Florida's opinion contains the strictest billing language of any jurisdiction reviewed, barring lawyers from billing in any way that would "falsely inflate the lawyer's billable hours" and requiring AI output to be reviewed the same way a lawyer would review a paralegal's work.³ Texas's opinion — finalized in February 2025 as Opinion No. 705, after circulating as a draft the previous year — flags a specific and easy-to-miss risk: "many generative AI tools invite a 'conversation' in which the lawyer... will explain relevant facts," which can leak more confidential detail than a lawyer intends if the tool retains or trains on what's typed in.⁴ New York's task force coined a term worth knowing — "techno-solutionism," the assumption that AI automatically solves a profession's access-to-justice problems — and used it as a caution rather than a dismissal, going so far as to suggest that refusing to adopt genuinely useful, well-vetted tools could itself become a competence problem over time.⁵ California is considering amendments to its Rules of Professional Conduct that would expressly address AI use — going further than any bar in the country by proposing changes to actual enforceable rule text (Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3), not just non-binding guidance.⁶ The proposed amendments would require lawyers to independently review and verify AI output and would clarify duties concerning fabricated, misstated, or out-of-context authorities. The proposal remains subject to the State Bar's rulemaking process and should not be described as binding law; confirm its current status directly with the State Bar before relying on it.

What the reliability research actually found

Two Stanford studies get cited constantly and conflated just as often, because they tested different things. The first, run by a team at Stanford's RegLab, put general-purpose models through more than 800,000 queries across fourteen legal-reasoning tasks, checking their answers against a validated dataset of real cases.⁷ The hallucination rates were not close: 58% for GPT-4, 69% for GPT-3.5, 72% for PaLM 2, 88% for Llama 2 — and the errors got worse, not better, on the exact kind of material an immigration practice generates most: circuit and district-level cases performed far worse than Supreme Court questions, and when researchers asked about a fabricated legal premise, the models invented a confident answer instead of catching the fiction more than half the time.

The second study tested the tools lawyers are actually paying for — Lexis+ AI, Westlaw's AI-Assisted Research, and Ask Practical Law AI — against 202 pre-registered legal queries, scoring each response for both accuracy and whether its citations were properly grounded.⁸ Lexis+ AI came out ahead of the other two, but "ahead" meant 65% accurate, 17% outright hallucinated, and 18% incomplete or refused — meaning even the best-performing commercial legal-AI product on the market was wrong or ungrounded on roughly a third of the queries it actually answered. Westlaw's tool hallucinated a third of the time; Practical Law AI hallucinated a fifth of the time but declined to answer nearly two-thirds of questions outright. Purpose-built legal AI is real progress over a general chatbot — but hallucination rates of 17-33% in a controlled evaluation are still nowhere near reliable enough to justify unsupervised use in a filing, which is exactly the gap the verification duties in ABA 512 and California's proposed rule are built to close.

The tools built specifically for this practice area

AILA partnered with the legal-tech company Visalaw.ai in 2023 to build a tool called Gen — GPT-4-based, trained on AILA's own Practice and Procedures Manual, offering research, drafting, summarization, and translation.⁹ AILA states the tool keeps its immigration-law library architecturally separate from the underlying model "to ensure the information exchanged remains secure and private." That's worth taking as a claim to verify, not a settled fact — no independent, third-party benchmark of Gen's accuracy exists in any source reviewed for this piece, and the same is true of the broader immigration case-management market: Docketwise, INSZoom, Imagility, eimmigration, and LawLogix all show up regularly in comparison content, almost all of it written by the vendors themselves reviewing each other. None of it is independently benchmarked, so "market leader" claims from any single vendor blog should be read as marketing, not evaluation.

What this actually means for a practice

The rules converge on a handful of things that aren't optional, whichever state you're in. Every citation and factual claim needs independent verification before it goes in a filing — not a skim, given that even the strongest commercial tool is wrong close to a third of the time on non-refused answers. Generic AI-use language in an engagement letter almost certainly doesn't satisfy the ABA's confidentiality standard, so it's worth rewriting into something specific rather than assuming the boilerplate covers you. Lawyers should bill only for time actually spent on the work and should not inflate or duplicate charges merely because AI made the work faster. And a written AI policy — even a short one, even at a solo practice — closes the gap between how widely these tools are already used informally and how rarely that use is actually documented. If you supervise anyone, from an associate to a paralegal, that policy stops being optional under both ABA 512 and Florida's opinion specifically.

Two things in this piece couldn't be nailed down from public sources and are worth a direct check before anyone relies on them: whether the Department of Justice's department-wide AI guidance — which EOIR said it expected by the end of 2025 — actually shipped, and whether California's proposed rule amendments have moved past public comment toward final adoption. Both are the kind of fact that ages within months, not years.

This is general information about published ethics rules and research, not legal advice. Cooperative is not a law firm. Consult your own bar counsel or malpractice carrier before changing firm policy.

References

1. American Bar Association (2024). Formal Opinion 512.
2. Executive Office for Immigration Review (2025). Policy Memorandum PM 25-40.
3. The Florida Bar (2024). Ethics Advisory Opinion 24-1.
4. State Bar of Texas Professional Ethics Committee (2025). Opinion No. 705.
5. New York State Bar Association (2024). Report and Recommendations of the Task Force on Artificial Intelligence.
6. The State Bar of California (2026). Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence. Status subject to change — confirm current adoption status before citing as binding.
7. Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024). "Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models." arXiv:2401.01301.
8. Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2024). "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools." arXiv:2405.20362.
9. American Immigration Lawyers Association (2023). "Visalaw.ai and AILA Unveil Gen."