On August 8, 2025, the Executive Office for Immigration Review (EOIR) issued Policy Memorandum 25-40 on the Use of Generative Artificial Intelligence in EOIR Proceedings. It is the agency’s first public policy statement aimed squarely at regulating how AI should and should not be used in immigration court.
I wanted to write briefly about this policy memo and try to put together a clear statement of what the agency’s policy currently is both for attorneys using AI in their legal writing and judges using AI in writing decisions. And I wanted to compare this memo to the agency’s last statement, a June 2024 workforce plan, that said EOIR intended to lean heavily into using AI to make up for budget shortfalls.
This week’s memo echoes warnings that have become familiar in legal circles. Generative AI can fabricate case law, mangle citations, and introduce factual inaccuracies. It cites recent high-profile examples in which judges had to withdraw opinions after AI-generated text slipped into legal filings. It re-emphasizes that EOIR staff cannot use unauthorized AI tools on government devices, and it reminds practitioners that they are responsible for verifying every word in what they file.
The message to attorneys is clear. If you use AI, you are responsible for every hallucinated case or invented fact it produces, and you could face sanctions or discipline if you are not careful.
But EOIR’s Own AI Plans Tell a Different Story
This strict posture toward the private bar contrasts sharply with EOIR’s own embrace of AI in its 2024 Strategic Workforce Plan.
That document was prepared in the face of a budget cut from FY 2023 to FY 2024, despite a surging caseload. It presents AI not as a risk to be minimized, but as a key strategy for closing the gap between resources and workload. Under Goal 3, “Create an Adaptable Organization by Using the Right Technology,” EOIR states:
“The advent of Artificial Intelligence (AI) will play a key role in EOIR human capital endeavors.”
The plan frames AI as a way to maximize and better utilize scarce resources. In plain terms, that could include streamlining the drafting of immigration judge decisions, boilerplate orders, and other adjudicatory documents.
This creates a striking asymmetry. Attorneys are warned about the dangers of AI and held to strict verification standards. But EOIR itself openly contemplates using AI to meet operational goals without any stated commitment in the 2025 memo to disclose when an immigration judge’s decision has been drafted or influenced by AI.
The Transparency Gap
If the agency is going to use AI to help write judicial opinions or guidance, should the parties know? Immigration decisions often hinge on credibility findings, legal nuance, and individualized application of law to fact. If an AI tool is assisting in drafting, the potential for bias, error, or subtle misinterpretation is real. And unlike a brief, there is no adversary to challenge the judge’s “sources.”
The August 2025 memo does say that there is a May 2025 policy prohibiting “unauthorized use” of AI on “Government Furnished Equipment.” But that policy statement hasn’t been made public, which is a bit ironic.* I’ve filed a FOIA request for a copy and will publish it when I receive it. Either way, from the description that policy doesn’t prohibit judges from using AI to write decisions. It apparently only prohibits “unauthorized” AI and only when using “government furnished” computers. Sure, the memo says “improper use” of generative AI by an immigration judge can result in discipline. But the focus here is on whether the use of AI is “authorized,” not so much on whether the result is accurate or the use of AI is disclosed.
And that’s really the biggest concern. Last week’s memo’s silence on internal AI use for decision-writing suggests an uneven application of caution. There is a bright spotlight on the risks posed by attorneys’ AI use (which, to be clear, are quite real), but no parallel discussion of how EOIR will govern its own.
State Bars and Attorney AI Use
Across the country, state bars have been issuing guidance on generative AI.
- California warns that lawyers must maintain the same duties of competence, confidentiality, and supervision when using AI as with any other tool.
- New York reminds attorneys they cannot delegate legal judgment to a machine and must verify all outputs.
- Florida emphasizes protecting client confidentiality and avoiding reliance on undisclosed AI research.
Common themes emerge. Human review, transparency, and ethical accountability. These guidelines apply to lawyers, not courts or agencies.
Judges Using AI: The Next Frontier
Some state court systems are already wrestling with this. The California Judicial Council’s Model Policy for Use of Generative AI by Judges (effective September 1, 2025) requires:
- No use of public AI tools for confidential case information.
- Human review of any AI-assisted drafting.
- Disclosure in the record if AI was used.
That last requirement, disclosure, is precisely what EOIR’s memo does not address. If the agency uses AI to help draft opinions (which the memo suggests may already be happening, just not on government-furnished computers) the parties could be left guessing whether their decision was human-crafted, machine-assisted, or somewhere in between.
Why This Matters
Generative AI is likely not going away. Some believe that with enough human review, AI can reliably make legal research faster and draft text in seconds. But it can also just make stuff up, reinforce bias, or give a false sense of authority. Some AI models have generated hateful text about minority groups, made up facts that weren’t true, and created case law that doesn’t exist.
In immigration court, where the stakes can mean life or death, the standard for transparency should be higher, not lower. If EOIR expects attorneys to be held accountable for AI use, as it absolutely should, then it should hold itself to the same standard.
Looking Ahead
EOIR’s 2024 budget plan shows where the agency wants to go, which is to use AI to stretch resources. EOIR’s 2025 policy memo shows how it wants to control others’ use of the same technology. Those two realities are not necessarily incompatible, but the lack of symmetry raises questions about fairness, due process, and trust in the immigration court system.
If the agency moves forward with AI decision-drafting, it should follow the lead of states like California. Ban confidential data from public AI tools, require human review, and disclose AI use to the parties. Otherwise, we risk a two-tier system of AI rules, one for the people in court and another for the people running it.
AI Rules in Immigration Court: Attorneys vs. EOIR
| Attorneys | EOIR (Agency & Immigration Judges) |
|---|---|
| Must verify every citation, fact, and source generated with AI before filing. | May explore AI to improve efficiency, including drafting opinions and orders, under 2024 Strategic Workforce Plan. |
| Face sanctions if AI-generated content is inaccurate, misleading, or unsupported. | No stated sanctions or penalties for internal AI errors in decision-writing. |
| Prohibited from delegating legal judgment to AI. | No public rule restricting AI from assisting with legal reasoning in decisions. |
| Must protect client confidentiality and avoid inputting sensitive data into public AI tools. | No disclosure requirement if AI is used internally to draft or edit opinions. |
| Bound by state bar ethics opinions, requiring human review and transparency. | Governed internally, with policy memo silent on disclosure or party notification. |
* (this is an aside, but the secret AI policy from May 2025 mentioned in the Owens memo is ironic, because Owens stormed in as Acting Director by excoriating the previous administration for having so-called “Secret Operational Policies.” Owens says in her “secret policies” memo that issuing such policies without public disclosure “may have violated basic principles of administrative law.” It’s a bit rich to now learn that EOIR has this secret AI policy from May 2025 that governs how judges may or may not use AI in writing their decisions, the very thing the Acting Director called unlawful when done by the prior admin).