Federal Artificial Intelligence Transparency Act

The bill defines an "AI system" as any computational process that performs tasks normally requiring human cognition. Developers of systems used by 100,000 or more people, or by any federal agency, must file an annual disclosure: the categories of training data used, the broad demographic distribution of that data, the system's intended use, and any known failure modes.

A new office within the National Institute of Standards and Technology, the Office of AI Auditing, is authorized to spot-check disclosures, request additional documentation, and refer false statements to the Justice Department. The bill creates no general prohibition on AI use. Applications classified as "high-risk" (criminal sentencing, hiring decisions affecting more than 50 employees, medical diagnosis, or election infrastructure) must be registered with the OAA before deployment.

The bill is silent on civil penalties; enforcement comes through existing FTC and DOJ authorities. State preemption is explicit: states cannot impose conflicting disclosure regimes on systems already covered.

Disagreement runs along two main axes. The first is whether the disclosure framework is too narrow or too broad. Industry groups argue that "training data categories" is undefined and will force trade-secret disclosures by the back door. Consumer advocates argue that without underlying data access, audits will be unable to detect bias and the disclosures become marketing material.

The second is preemption. State attorneys general from both parties have opposed the preemption clause, arguing it freezes state innovation on AI policy at a moment when federal frameworks are still in flux. Bill sponsors counter that a 50-state regulatory patchwork would push AI development overseas. They argue federal disclosure should function as a floor that states can build on. Both readings of the text are textually plausible.