Government AI Accountability: The Missing Policy Discussion About Source Attribution

Where accountability stands today

OMB M-25-21 (April 3, 2025) is the central federal framework. It requires agencies to designate chief AI officers, inventory their AI use cases at least annually and post a public version, and apply minimum risk-management practices to high-impact AI. It calls for “meaningful public transparency into the Federal Government’s use of AI” and lists inaccurate or misleading AI outputs among the risks agencies manage. It states that it does not create rights or obligations for the public.

All of this concerns AI that agencies use or buy. It is the right place for those rules.

The part nobody owns

The public increasingly meets government information through AI systems that agencies neither operate nor control. When those systems misstate an agency’s notice, no framework assigns anyone to notice, measure or report it.

• The Institute for Strategic Dialogue found about 29 percent of chatbot answers to basic voting questions incomplete, inaccurate or outdated.
• The New York Attorney General told voters not to rely on chatbots for voting information and to check official sources.
• CRS reported that fewer than 3 percent of AI-written bill summaries met its standards, and its director stressed keeping a highly skilled human in the loop.

In each case the response was a warning or a human check. Neither is a systematic way to know how an agency is represented.

Options

Benefits

A measurement habit would give agencies facts, not anecdotes. It would show which categories of information are most often wrong, which would tell policymakers where a standard or pilot would matter most.

Implementation burden

Self-monitoring takes staff time to choose questions, run them and keep a log. A shared benchmark takes technical staff and sustained attention, because AI systems change continually and a benchmark must change with them.

Drawbacks and open questions

• Measurement is hard. AI answers vary from one request to the next, and results depend on how questions are worded.
• Speech concerns. Government should not dictate what private AI systems may say. A sound approach concentrates on making government’s own statements clear, dated and verifiable, and leaves other content alone.
• Who decides what is official. A designation scheme needs a transparent rule so it cannot be used to favor one source.
• Scope. The federal government’s own information is only part of the picture. State and local information drives much of the harm in election and emergency settings.

The discussion to have

The existing accountability framework asks, “Is the government using AI responsibly?” A complementary question is, “Can the public tell what the government actually said?” The second question is about attribution. It deserves its own owner, its own measurements and a place in the policy conversation.