AI Foundation Model Transparency Act of 2026
Sponsored By: Representative Beyer
Introduced
Summary
This bill would force greater transparency for foundation AI models by making companies disclose training data, testing, risks, and operational details to the Federal Trade Commission and the public. It focuses on clear, machine-readable and human-readable disclosures, help for small businesses, and FTC enforcement for noncompliance.
Show full summary
- Companies that meet "covered entity" thresholds would need to submit detailed information for each foundation model, including training data sources and retention, broad data composition, model size, knowledge cutoff, performance on benchmarks, languages supported, computational power, risk monitoring, and data governance. Enforcement would be handled by the FTC as an unfair or deceptive practice.
- Developers and downstream services that build on base models must publish the base model’s transparency URL and follow rules when they significantly change or retrain models. Fully open-source models are exempt from these rules. The bill defines a foundation model as one with at least 1 billion parameters and sets coverage thresholds like 10 million monthly users.
- Small and new businesses get a machine-readable template, a qualified technical representative for help, and a three-month grace period without penalties while they comply.
Your PRIA Score
Personalized for You
How does this bill affect your finances?
Sign up for a PRIA Policy Scan to see your personalized alignment score for this bill and every other piece of legislation we track. We analyze your financial profile against policy provisions to show you exactly what matters to your wallet.
Bill Overview
Analyzed Economic Effects
5 provisions identified: 2 benefits, 0 costs, 3 mixed.
New transparency rules for AI models
If enacted, covered entities would have to give the FTC detailed information about each foundation model and publish key items for the public. Disclosures would appear on the company website and in a central FTC machine‑readable repository with a published URL. Required items would include training data summaries, data governance, intended uses and limits, version and knowledge cutoff dates, monitoring and incident response, supported languages, alignment efforts, benchmark performance and precautions for many high‑risk areas, and compute used. Some sensitive items could be redacted but each redaction must be identified and justified. Downstream models must link to a compliant base model’s disclosure and must follow rules when they materially change the base.
Help for small and new AI businesses
If enacted, the FTC would be required to help small and new businesses that become covered entities. The help would include published guidance and a machine‑readable disclosure template developed with NIST. A qualifying small or new business would get one three‑month penalty‑free grace period after becoming covered. The FTC would assign a technical representative to meet with the business several times during that grace period to help with compliance.
FTC schedule and resources for developers
If enacted, the FTC would have to write transparency rules within one year and update them at least annually. The rules would start 90 days after the FTC issues them. The FTC would assess the rules within one year and then every year after that. The FTC would also create a public Foundation Model Resources Page within one year to help developers who are not covered, listing recommended resources such as the NIST AI Risk Management Framework.
FTC enforcement and notice rules
If enacted, breaking the new transparency rules would count as an FTC Act violation. The FTC could use its usual enforcement powers and remedies under the FTC Act. The Commission would have to give covered entities notice that they are covered at least 14 days before starting enforcement action. The FTC must also report to Congress on rule implementation within one year.
Who counts as a foundation model
If enacted, the bill would define what a "foundation model" is and who is a covered entity. A foundation model would usually have at least 1,000,000,000 parameters. A provider would be covered if the model is high‑risk, has over 10,000,000 monthly users, over 10,000,000 monthly downloads, or used more than 10^26 operations to train. Fully open‑source models would be exempt if they meet the Commission’s test. The FTC could change those numeric thresholds by rule.
Sponsors & CoSponsors
Sponsor
Beyer
VA • D
Cosponsors
Lawler
NY • R
Sponsored 3/26/2026
Jacobs
CA • D
Sponsored 3/26/2026
Roll Call Votes
No roll call votes available for this bill.
View on Congress.govTake It Personal
Get Your Personalized Policy View
Start a Free Government Policy Watch to see how policy affects your household, then upgrade to PRIA Full Coverage for year-round monitoring.
Already have an account? Sign in