The Role of Congress in Examining AI-Readable Government Publishing Standards

Congress as a user and a publisher

Congress has direct experience with the problem. On June 25, 2026, the Congressional Research Service director told the House Administration Committee that CRS had tested six AI models over two years on about 1,000 bills, and that fewer than 3 percent of the results met its standards for accuracy, coherence, relevance and objectivity. She named the risks as outdated information, hallucinations, bias and distortions. Members at the hearing urged vigilance about the data going into these models.

Congress is also a publisher. It sets its own publishing practices for legislation, committee materials and reports, and legislative-branch publishers could serve as an early test of any approach.

What is already before Congress

• S. 4915, the AI Labeling Act of 2026. Would require labeling and machine-readable disclosures for AI-generated images, video and audio, and would direct NIST, with the FTC, to convene a working group on standards. Referred to the Senate Commerce Committee.
• The REAL Act. Would require agencies and officials to label AI-generated material they publish.
• The OPEN Government Data Act. Already law (Title II of Public Law 115-435). It makes public data assets machine-readable, with metadata and a data inventory, and is oriented to data assets rather than individual statements.

None of these addresses the question in this series: whether an official statement should carry its authority, jurisdiction, provenance, time and withdrawal status in a form software can verify.

Tools available to Congress

Why a staged approach fits

The sequence that serves the public best is measure, pilot, guidance, standard, and only then statute. A new Congress convenes in January 2027, and committees organize their agendas then. That makes the months ahead a natural time to ask for assessments rather than mandates.

Benefits of congressional attention

• The accuracy of public information is not a partisan matter, and members of both parties have raised AI accuracy concerns.
• Hearings and reviews bring state and local officials into the conversation, which federal memoranda cannot.
• A modest first step, such as a GAO review, creates a factual record for everyone.

Implementation burden

Hearings and report language mainly take committee staff time. A pilot would require agency participation and a measurement plan. A statute would require the most drafting and the most care about technical detail.

Drawbacks and open questions

• Technical detail in statute. Legislating a specific technical approach can lock in a design that does not hold up. Statutes work best when they set goals and assign standards work to expert bodies.
• Election sensitivity. Election information is among the most important categories and the most sensitive. Any approach must be visibly neutral across parties and candidates.
• Overlap. Provenance is already being discussed in the AI labeling bills. Policymakers should coordinate so that labeling of AI-generated content and verification of official statements fit together.
• Reach. Most government information originates at the state and local level, where Congress can inform but not direct.