Should Government Publishing Standards Include AI-Readable Information?

What “AI-readable” means here

It does not mean a special format for AI. It means information published as structured data in an open format, such as JSON or XML, with named fields, so software does not have to infer meaning from layout, branding or surrounding text. The same data serves search engines, researchers, assistive tools and AI systems alike.

Three layers of a publishing standard

It helps to separate what a standard can ask for.

• Discovery. Can software find the content? Sitemaps, robots.txt and crawl access.
• Description. Does the content say what it is? Page metadata and structured markup.
• Assurance. Can software check that a statement is authentic, unchanged and still current? Signatures, hashes and withdrawal notices.

OMB M-23-22 speaks to the first two. It asks for structure, descriptive metadata and machine-readable content to the extent practicable, directs agencies to sitemaps and robots.txt files, and asks them to consider an API where public interest is significant. It also tells agencies to default to HTML rather than formats such as PDF unless a user need requires otherwise. All of that helps.

Why assurance is the open layer

NIST’s 2024 report on digital content transparency is direct about the limits of description alone: “Metadata on its own is neither tamper-evident nor reliably attributable.” Signing improves that, but depends on key management and a public key infrastructure, which the report identifies as a source of vulnerabilities of its own. A standard that stops at markup would help systems find and label content. It would not tell them whether to trust it.

Where existing standards stand

The 21st Century IDEA Act requires new public-facing federal websites to comply with the General Services Administration’s website standards. Whether those standards already address the attribution of individual statements is a question GSA can answer directly, and a useful first step for any policymaker.

Policy options

Benefits

• Software can tell who published a statement without guessing.
• Agencies state their own authority and jurisdiction instead of leaving a third party to infer them.
• One structure can serve many consumers, not only AI systems.

Implementation burden

Markup is the lightest step: it can often be added in the content system an agency already runs. A feed or API requires a publishing step for each record. Signing adds a key-custody arrangement, whether the agency or a registry holds the keys, and a procedure for rotating and retiring keys. Each step adds maintenance, and the burden grows with the assurance gained.

Drawbacks and open questions

• Scope. Applying a standard to all content would be heavy and likely unnecessary. A standard could name categories, such as emergency notices and election information, where accuracy and currency matter most.
• Format choice. Several open formats could serve. The policy should specify outcomes, such as the fields and how they are verified, and avoid naming a product.
• Consumption. A standard governs publishers. It cannot require an AI system to read what is published.
• Maintenance. An unmaintained feed that drifts from the website is worse than none.

What to resolve first

Which categories of information come first, which fields are required, and who checks conformance. Those are questions for a standards body working with agencies, and a small pilot would give them real answers.