Policy series · Article 5 of 8
Should Government Agencies Establish Policies for AI Citation Registries?
Federal memoranda do not reach most governments. The decisions that matter often sit with a city, a county, a state agency or a public information office. This article sets out what an internal policy might cover and what an agency should weigh before adopting one.
Why an internal policy
Publishing through a registry or AI-readable feed changes who must do what at the moment a statement is released. Without a written policy, those decisions get made case by case, and the feed can drift from the website. A policy settles responsibilities in advance.
What a policy would cover
- Scope. Which categories go first. Emergency notices, election information, public-health advisories and benefit deadlines are common candidates because accuracy and currency matter most.
- Authority to release. Who may approve a record. Ideally the same people who approve a public release today.
- Single source of truth. Records are generated from the finalized public statement, not from a separate drafting channel. Drafts and internal material never enter the feed.
- Timing. When a record is released relative to the website and other channels.
- Corrections and withdrawals. Who may withdraw a record, how fast, and how a corrected statement is published as a new record.
- Key custody. Who holds the signing keys (the agency, a registry acting for it, or both), how they are rotated and retired, and what happens if a key is lost or exposed. If a registry signs on the agency’s behalf, the policy also says how the registry confirms that a submission comes from the agency.
- Records retention. How the feed aligns with the agency’s existing retention schedule.
- Monitoring. A routine check of how AI systems answer a sample of questions about the agency, kept as a log.
- Review. A date to revisit the policy and retire it if it is not working.
Options for an agency
| Option | What it would do | Main trade-off |
|---|---|---|
| Wait | Take no action until a standard or clear precedent exists. | Statements remain open to the misattribution and staleness already documented. |
| Pilot one category | Start with emergency or election information in one department. | Limited reach, but the lowest burden and a clear way to learn. |
| Adopt agency-wide | Apply the policy to all public statements. | Heaviest workload, and hard to sustain with a small staff. |
| Join a shared effort | Adopt a model policy developed with peer agencies or an association. | Depends on someone developing and maintaining it. |
A way to measure
An agency can test before and after. Ask several AI systems the same set of questions about the agency’s notices, record the answers, and repeat after publishing records for a few months. This does not prove cause, because AI systems change for many reasons. It does give the agency its own evidence instead of relying on assertions from anyone, including vendors.
Benefits
- The agency states its own authority, jurisdiction and timing, rather than leaving AI systems to infer them.
- Withdrawals and corrections have a defined path.
- The monitoring log gives leadership a factual picture of how the agency is represented.
Implementation burden
The principal burden is a workflow step for public information staff on every record, plus training and, where the agency holds keys, a custody procedure that survives staff turnover. Legal review of the policy and of any retention implications is prudent. Agencies with a single public information officer will need to be selective about scope.
Drawbacks and open questions
- Divergence. If the website changes and the feed does not, the agency has two versions. The policy must prevent this, preferably by generating both from one source.
- Uncertain uptake. The benefit depends on AI systems choosing to read the feed. Monitoring will show whether that is happening.
- Key loss. Lost or exposed keys are a real operational risk. Whoever holds them needs a tested recovery plan.
- Scope creep. A feed that tries to carry everything becomes unmaintainable. Narrow beats broad.
A reasonable first step
A one-category pilot with a monitoring log and a review date is the smallest commitment that produces evidence. It also lets the agency adjust the policy before applying it more widely.
Sources
- OMB Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (April 3, 2025)
- OMB Memorandum M-23-22, Delivering a Digital-First Public Experience (September 22, 2023)
- Foundations for Evidence-Based Policymaking Act of 2018, Public Law 115-435 (Title II, OPEN Government Data Act)
- New York Attorney General, warning about AI-generated election information (October 14, 2026)
Sources were read in the original or, where noted, through a published summary. Statuses and dates are as of October 18, 2026.