On November 19, the Administrative Conference of the United States opened a public consultation on the use of generative AI in federal agency adjudication. Its notice asks how agency staff and adjudicators use the tools, what limits should apply to them and to people appearing before the agencies, and when someone should disclose their use or certify that they checked the output. The notice sets a deadline of 10 a.m. Eastern time on January 19, 2027. The project is gathering information for recommendations on agency practice. The published notice itself solicits comments. It puts no new filing or disclosure requirement into effect. Federal Register notice
The question behind the consultation concerns the record of a decision. A person denied a benefit, charged with violating a rule or seeking some other determination needs to understand the evidence the agency considered and the grounds it gave. A person without a lawyer can also benefit from help finding forms or understanding unfamiliar language. Both situations can involve AI. They call for different safeguards. Using a tool to find the address of a hearing office changes little about the evidence. Using one to summarize a contested interview can change which facts an adjudicator notices. A good policy should make the distinction visible to the people on either side of the file.
What the notice asks agencies and the public to describe#
ACUS studies the efficiency, adequacy and fairness of federal administrative procedures and makes recommendations. Agency adjudication covers processes in which an agency applies law or policy to particular parties or matters. The forms of adjudication vary, and the governing procedures depend on the program. The questions in the notice cover work by adjudicators and other agency personnel, by lawyers and other representatives, and by parties acting on their own. The breadth matters because a model can be used at several points in one case. The people affected by the decision do not all have the same degree of access to the system or its records. ACUS’s adjudication overview
The notice uses the term generative super intelligence for technology it says was previously called generative artificial intelligence. Its focus is systems that generate new material from inputs. The title does not certify a level of reasoning ability. Practical examples in the notice include transcribing a hearing, translating documents, searching for authority, summarizing evidence and preparing notices or decisions. Each use changes a different part of the process. A transcript is expected to preserve what was said. A draft notice conveys a deadline or reason. A summary intentionally leaves material out. Treating these outputs as one category obscures the particular errors that matter and who should catch them. Definition and use cases in the notice
The ten questions invite accounts of current practice, what has worked, where consequences should follow misuse and whether existing professional rules give enough guidance. ACUS also asks about materials agencies and other institutions have already adopted, such as standing orders and practice guides, and whether rules should differ for government lawyers, private lawyers, other representatives and people appearing on their own. That affords room for practical evidence. A comment about a workflow can explain what someone put into the tool, how they validated the output and which consequences actually followed. Those details can inform future recommendations more directly than a general declaration about the technology.
It is worth keeping the adoption statistics in their original frame. The notice cites a Government Accountability Office study that found about a ninefold increase in reported generative AI cases from 2023 to 2024. The GAO found 32 rising to 282 among eleven selected agencies with inventories. Cases were reported in different areas and stages. The study says 84 of the 282 were in an early initiated phase and another 75 were in acquisition or development. This provides evidence of growing federal experimentation. The counts measure reported uses or plans in that sample, and by themselves give no estimate of how many decisions were written with AI or how many participants encountered it. GAO report on generative AI use
The agency has to preserve the evidence trail#
Consider a hypothetical benefits case with a written application, a hearing transcript and several dated supporting documents. An agency employee asks a model to summarize the materials. The draft summary includes the date one document was received and omits a different filing date that would change how the adjudicator reads the timeline. A later written decision could reflect that omission even if the final author never copies a whole sentence of generated prose. The questions worth asking are where the summary came from, how it was checked and whether the omitted evidence can be shown to the person who decided the case. A label reporting AI use answers only the first part of that inquiry.
A policy can distinguish mechanical assistance from tasks that shape the evaluation of a claim. Spelling changes and formatting raise different problems from selecting relevant testimony or drafting a finding of fact. The distinction follows the function more reliably than the name of the product. An assistant may offer both editing and analysis under one button. A legal search tool can suggest citations, some of which may be inapplicable, while a person still decides what to cite. A case file can indicate which parts of an output were independently verified and what source material was considered. That provides a reviewer with more to examine than a list of software names.
The clues should survive the ordinary workflow. If an assistant creates notes that affect a decision and the system removes them when the application closes, the effective record may be less complete than staff expect. If a draft changes repeatedly, a reviewer may need to see which version was actually relied upon. Retention also has costs, particularly when a prompt reproduces private data or privileged information. A workable system can preserve material input, citations, model version and human checks under controls suited to the case, while avoiding unnecessary copies of sensitive records. General knowledge of where software was used gives an auditor an overview. A separate case record needs to explain a particular decision.
Verification has to have a responsible owner. A sign-off can have little force if the person reviewing a summary has neither the original materials nor time to read them. In another arrangement an adjudicator may be able to view the original passage, inspect an extracted timeline and make a correction before signing. Access to the source is essential for comparing what the model selected with what the parties supplied. ACUS’s 2020 statement already addressed the importance of transparency when agencies owe reasons in settings including adjudication. The new project can examine how that principle works with generated drafts and new tools operating on case records. ACUS Statement 20 on agency AI use
There is a similar issue for translations. Imagine a person describes a temporary absence from work and the translation conveys a resignation. Both can fit a short sentence out of context. The person may have a chance to correct the translation if they see what the agency received and can use an independent language service. They are less likely to detect the error if an internal summary replaces the transcript or translated record. The source, which version was checked and how corrections flow into the case should be apparent to the people responsible. The same principle applies to captions and automated transcriptions. These tools can make a case easier to access and introduce consequential errors at the same time.
Timing separates a helpful correction from a late record of a mistake. Suppose the original statement can be reached only by a records request long after a decision. The party may then be able to see what happened, yet have much less time or fewer options to correct the result. An earlier opportunity to examine a disputed passage can allow staff to recheck the file before the decision is issued. The procedure will differ among agencies and case types. A recommendation can still ask agencies to identify when a person first sees the disputed material, how to point out an error and what happens to the decision while the issue is considered.
Duties differ across the people who participate#
For a person representing themselves, a tool can make the first step easier. It can suggest where an agency keeps its forms, explain unfamiliar vocabulary or help write a statement that the person checks against what they actually know. The case for that help is strongest when official forms and deadlines can be directly verified. A requirement to disclose every language correction could burden someone who uses assistance to participate at all. A requirement to verify citations and factual claims targets a different concern. The design question is how to make disclosure meaningful and still give someone with limited resources a reasonable way to understand and comply.
Representatives can have duties to a client and a tribunal, including competence, confidentiality and candor. A familiar risk is a brief that cites a nonexistent authority. A quieter risk is a real authority that supports a different proposition from the one asserted. Verification requires reading the underlying rule or decision with its context, rather than noticing whether a citation looks plausible. A representative who sends material to a software service may also need to understand how it handles that material. ACUS’s question about existing rules is therefore concrete. Will a practitioner know when a duty applies, how to meet it with a particular tool and what agency procedure can examine compliance? ACUS questions on representation and disclosure
The notice points to institutions that have already taken different steps. The US Patent and Trademark Office’s 2024 guidance addresses how existing rules apply when parties and practitioners use AI-based tools. It discusses signing, review of submissions and protection of sensitive information. The Executive Office for Immigration Review issued its own 2025 policy memorandum. These references matter as examples of agencies working within different programs, rules of conduct and filing systems. They give the consultation comparable material. Effects in an intellectual property filing environment and effects in an immigration proceeding should be assessed in their own context before any practice is carried across. USPTO 2024 guidance · EOIR policy memorandum
The agency itself is also a participant with particular responsibilities. An agency lawyer may use a tool when presenting the government’s case. An adjudicator may use one when reviewing facts or writing reasons. A clerk may use one when preparing a notice. A policy can cover all three while acknowledging what each person is accountable for. Even a reduction in staff time can be relevant to fairness if it lets an agency answer parties sooner. It matters whether the benefit comes from easier processing or an unexamined shortcut in considering evidence. The consultation can ask agencies to show how they supervise that workflow and what an affected person can learn about their own case.
It is also useful to compare misuses by conduct and consequence. A person who knowingly submits fabricated evidence poses a different problem from a person who misreads a tool’s translation. Both may need to be addressed, but broad language about AI can collapse the distinction. Existing professional rules may give a way to enforce duties for lawyers. A person appearing alone may have very different knowledge of the rules or means of correcting an error. How an agency gives notice, allows clarification and assesses a sanction belongs in the discussion of misuse, alongside the question of what a tool can do. ACUS questions on consequences and participants
Disclosure needs to tell someone what to check#
A disclosure can be designed around what a person needs to check. It can identify a hearing transcript as machine-generated and indicate that a person checked disputed words. It can say which papers an automated search retrieved and provide the sources corresponding to any claim. A certification can put responsibility for factual and legal assertions on the signer. These statements help people decide whether to ask for a correction or examine a specific record. They have limited value if the only thing disclosed is a vendor name. Asking for an activation log may help a supervisor discover which tools were used. It need not give a party a usable explanation of an individual decision.
Some uses call for different levels of public detail. It may be reasonable to describe the approved tool and its model version in an agency policy, with information about which tasks it can support. A case-level notice can say which substantive step used it. An internal record can preserve more material for an authorized reviewer. Those levels need a connection to one another. A participant who raises a concrete issue should be able to find the right channel and learn what part of an assessment can be checked without revealing another person’s protected records. Data minimization and meaningful review have to be planned together.
The wider federal framework already addresses aspects of agency AI governance. The notice cites the Office of Management and Budget’s memorandum M-25-21 on federal AI use. It discusses oversight and requirements for uses covered by its high-impact category, with scope tied to the application and consequences. An adjudication policy has more specific procedural questions, including where the source record sits and which disclosures help a particular participant. An agency’s inventory and an individual’s record serve different purposes. The inventory can tell the public that a tool is used in a program. To challenge a decision, a party may also need to know what changed in the file and who checked it. OMB memorandum on federal AI governance
There is an opportunity to state what a certification actually covers. A declaration by a representative that citations were verified can answer a focused question about a submission. An agency’s equivalent might document that the deciding official checked cited portions of testimony and reviewed any interpretation that came from a model. The two responsibilities are related and need not have identical wording. A reviewer should be able to distinguish certification about checking facts, confirmation of an administrative step, and an assertion about whether a particular model was ever used. Combining them in one checkbox risks making every signature look alike while the underlying work differs.
The review procedure also has to resolve who can change the record. A party can identify something that looks wrong, a caseworker can check the underlying evidence and an adjudicator can decide how the corrected fact affects the outcome. If one team only owns the tool and another only owns the file, a correction can stop between the two. A useful agency rule would state which office is responsible for receiving a concern, confirming whether the disputed information was used and telling the party how it was resolved. Making these responsibilities clear can prevent a decision from depending on a coordination failure.
Evidence that could improve the eventual recommendations#
The strongest comments would give enough context to compare results. An agency can describe the old workflow, why it tried a new tool, who reviewed the outputs and how often outputs needed correction before use. It can track both time saved and work added, including checking citations or resolving disputed translations. A private representative can explain how a proposed requirement affects verification in a particular forum. A self-represented participant can report whether a tool made instructions clearer and how hard it was to find a person when it did not. Those accounts give useful evidence even if the participants reach different conclusions. Consultation topics and submission information
The request says comments will ordinarily be published. That fact affects the evidence people can share. An agency can describe a workflow without including identifying case details. Researchers can present aggregate error rates and explain how examples were selected. Parties may be able to describe what happened in a hypothetical with the same factual structure, while keeping their own protected records out of a public comment. At the recommendations stage, an important question will be whether examples include failures and all relevant actors, including people who needed language access or could not use one of the offered tools. A record built only from convenient success stories leaves basic questions unanswered. ACUS public comment policy
One useful test would follow a disputed decision across a whole simulated case. Use prepared materials with a clear answer, insert a conflict that requires consulting the original record and see whether a staff member can identify and correct an inaccurate generated summary. Then test whether the correction appears in the copy of the record that the person and a reviewer can see. The outcome is valuable even when the tool saves time overall. It reveals whether the process prevents a common kind of error from becoming an official reason. It also gives someone who reads a future recommendation a way to tell whether the recommended check could work in an ordinary file.
Another test is to measure whether the people who are meant to benefit can actually use the service. A simple dashboard can count sessions without seeing whether someone could find the filing office, understand a deadline or finish a form with a disability or language barrier. A better comparison specifies which task was completed, what access assistance was available and when a person took over. The same standard applies to internal efficiency. A faster first draft is useful if review and correction costs do not erase the time saved. A proposal grounded in these observations can distinguish improved access from a burden moved to the people with the least ability to absorb it.
Agencies could also explain what happens when a tool is withdrawn after it starts to affect casework. A temporary outage may require a way to continue issuing notices. A systematic error can require finding earlier summaries and determining which cases they affected. A supplier update can warrant checking whether a once-reliable translation or extraction changed. These questions concern the life of the procedure beyond procurement or pilot results. It is possible to preserve a manual way to complete the same step, keep a dated record of a material change and measure how long it takes to resume service. Such evidence would make any eventual recommendation easier to apply.
The public notice opens an opportunity to work through these questions before recommending practices to agencies. A good result would let people receive useful help while making consequential use visible in the record. It would leave agencies able to show what material an adjudicator relied on, give parties a way to identify and correct an error, and make clear who continues to own the decision. Those outcomes can be evaluated no matter which model is in fashion by the time recommendations are delivered. The decisive evidence will be whether a person can see and contest a relevant error before the resulting decision becomes harder to change.