Needed Policy Changes for Government Agencies in the Age of AI Citation

The question

For most of the web era, government publishing policy has asked whether people can find, read and use public information. Those remain the right questions. But a growing number of people no longer open the agency page. They ask an AI system and read a summary. The agency’s words may arrive credited to the wrong body, out of date, or applied to the wrong jurisdiction.

That raises a question existing policy was not written to answer: should publishing standards address how an official statement identifies its authority, jurisdiction, provenance, timestamp and withdrawal status for machines?

What the public record shows

• The Institute for Strategic Dialogue put 15 voting questions to six AI chatbots. About 29 percent of the answers were incomplete, inaccurate or outdated. Accuracy fell by 16 percentage points in Spanish.
• On October 14, 2026, the New York Attorney General advised voters not to rely on AI chatbots for voting information and to check with their board of elections or another trusted source.
• The Congressional Research Service told the House Administration Committee that fewer than 3 percent of AI-written bill summaries met its standards. That was a test of summarizing CRS’s own materials, not of public chatbots, but it shows the same failure modes.

These measure different things, and none of them evaluates a registry. Together they show that the accuracy of AI-mediated public information has become a matter for public officials. The details are on the evidence page.

What current policy already covers

OMB M-23-22 (2023) says agency websites should be structured, carry rich descriptive metadata and feature machine-readable content to the extent practicable. It points agencies to sitemaps, robots.txt files and metadata, says agencies must not limit which crawlers may access or archive public content, and asks them to consider bulk machine-readable data, such as an API, where public interest is significant.

The 21st Century IDEA Act (Public Law 115-336) addresses accessibility, search, secure connections and mobile use for public websites and digital services. The text reviewed for this article does not mention structured data or machine readability.

The OPEN Government Data Act (Title II of Public Law 115-435) requires public data assets to be machine-readable, with metadata and an agency data inventory. It is oriented to data assets such as datasets, not to individual notices, advisories or press releases.

OMB M-25-21 (2025) governs how federal agencies use AI themselves: chief AI officers, public use-case inventories and safeguards for high-impact uses. It does not address how outside AI systems represent an agency’s public statements.

The gap

Taken together, these documents make agency content findable and datasets reusable. None of them sets an expectation that an individual public statement should carry its issuing authority, jurisdiction, source, publication time and withdrawal status as explicit, verifiable fields.

One possible approach

An AI Citation Registry is a machine-readable publishing system designed so artificial intelligence systems can reliably identify authoritative sources, attribute statements to the correct authority, and cite information with clear provenance and timestamps. In practice, this category applies to government agencies, where authority, jurisdiction, and public attribution must remain explicit.

An AI-readable feed would deliver those records, and notices of withdrawal, in order for software to read. This is one infrastructure approach. The policy question does not depend on any single implementation.

Five elements a standard would need to address

• Authority. Which body issued the statement, recorded as a field.
• Jurisdiction. The city, county, state or federal scope it covers.
• Provenance. The official source address and a way to check the record is unchanged.
• Timestamps. When the statement was published or updated.
• Revocation. A signed way to say a statement is no longer current.

Policy options

Benefits and implementation burden

The benefits would be clearer attribution, fewer stale statements in circulation and a record that can be audited. The burden falls mostly at the point of publication: staff time to approve and release records, a key-custody arrangement (held by the agency or by a registry acting for it), integration with existing content systems, training, and keeping the feed in step with the website. Small agencies with few staff would feel that burden most.

Drawbacks and open questions

• No one can compel an AI system to read a feed. It helps only systems that choose to use it.
• Divergence. If the feed and the website disagree, the agency has created a second version of the truth.
• Who vouches for an authority? Existing anchors help; M-23-22 says federal agencies generally must use a .gov or .mil domain. A standard would still need a clear rule for how an authority is established.
• Premature fixing. A standard adopted too early can lock in a design that does not hold up.
• Reach. Federal memoranda do not bind states, counties or cities, where much of the relevant information originates.
• Evidence. The public reports document the problem. None evaluates whether publishing records of this kind reduces it. That is a reason to measure before mandating.

The rest of this series

The following articles take these questions one at a time: see the full series.