H.R. 8094: AI Foundation Model Transparency Act of 2026

H.R. 8094 would create a new federal transparency regime for large AI foundation models, with the Federal Trade Commission writing the rules not later than 1 year after enactment. The bill is aimed at "covered entities" that provide a foundation model meeting at least one trigger: the model shows high-risk performance; has more than 10,000,000 monthly users; has more than 10,000,000 monthly download instances; or was trained using more than 10^26 integer or floating point operations. The bill defines a "foundation model" as an AI model trained on broad data, generally using self-supervision, with at least 1,000,000,000 parameters, designed for generality of output, and generally usable across many contexts and tasks.

The core idea is disclosure. Covered entities would have to publish information on their own websites in a human-readable, consumer-friendly format, while also providing information for a central FTC website in a machine-readable format. Required disclosures are broad: summaries of training data sources and collection methods; whether data is retained during inference; demographic and language composition of data; intended uses; model limits and risks; version and release date; knowledge cutoff date; supported languages; adverse incident monitoring and response procedures; and how the company aligns the model with the NIST AI Risk Management Framework, government-approved consensus technical standards, or its own internal guardrails.

The bill also pushes companies to disclose testing results for some of the most sensitive use cases. That includes evaluations tied to medical and health risks; biological, chemical, radiological, and nuclear weapons risks; national security; cybersecurity; critical infrastructure; elections; law enforcement; financial and housing decisions; education; employment and hiring; public services; and effects on vulnerable populations, specifically minors under age 18 and seniors age 65 or older. It also requires disclosure of computational power used to train and operate the model, which is notable because compute thresholds are already part of the bill's coverage test.

The measure is not absolute. Fully open-source models are exempt. Companies building downstream models can rely partly on a base model's disclosure by linking to its URL, but they still must comply if they make significant changes, retrain, or adapt the system. Small businesses and new businesses operating for less than 1 year get help: the FTC must provide guidance and machine-readable templates, plus a 3-month grace period from penalties, during which a technically proficient representative must meet with them to assist compliance. Enforcement would run through the FTC Act as an unfair or deceptive practice, but the FTC must give at least 14 days' notice before taking action. The regulations would take effect 90 days after promulgation, and the FTC would have to review and update the requirements every year starting 1 year after the initial rules are issued.