Government AI Bias Rules Leave Major Gaps in Oversight

OMB Memo M-26-04 requires federal AI systems to be truthful and objective, but vendor self-evaluation and contract loopholes undermine its enforcement.

Government AI Bias Rules Leave Major Gaps in Oversight

The Office of Management and Budget released Memo M-26-04 in December 2025, requiring that large language models used by federal agencies be "truthful" and prioritize "historical accuracy, scientific inquiry, and objectivity." Writing for Lawfare, Merve Hickok — President and Policy Director of the Center for AI and Digital Policy — argues that the mandate contains critical gaps that undermine its stated goals.

The most glaring weakness is the reliance on vendor self-evaluation. Under the memo, companies supplying AI tools to the government assess their own systems for bias and accuracy. There are no standardized model cards, no independent evaluation requirements, and no consistent benchmarks for what constitutes truthful output in public sector AI applications.

Existing government contracts present another problem. The memo requires modifications only "to the extent practicable," language Hickok identifies as a significant loophole. Contracts already in place — including xAI's Grok deployment, which has produced conspiratorial and inflammatory outputs — can continue operating under their original terms. Agencies face little pressure to renegotiate or impose new standards on current vendors.

The federal government also missed its November 2025 deadline for completing AI use case inventories, a prerequisite for understanding where government automation tools are actually deployed. Without that accounting, enforcing bias standards across government and public sector agencies becomes guesswork. For government AI governance to move beyond aspiration, independent auditing and enforceable procurement standards will need to match the scope of what the memo promises.