AI Foundation Model Transparency Act of 2023

This bill directs the Federal Trade Commission (FTC) to create standards requiring companies that provide certain large AI foundation models to disclose information about their training data and algorithms, making some of that information public. It aims to improve transparency for copyright enforcement and consumer protection.

Sec. 3, Foundation model data source and training transparency: Within 9 months, the FTC must write regulations setting standards for transparency of foundation models by covered entities, covering training data, model documentation, data collection during use, and model operations. The FTC must also issue guidance. It must consult with NIST, the Copyright Office, and others. The standards must require covered entities to submit certain information to the FTC and make other information publicly available, including on their websites and a central FTC website, in machine-readable format. The FTC will consider requiring details like training data sources, size and composition, data governance, labeling, intended purposes and risks, alignment with NIST's AI Risk Management Framework, performance on benchmarks (e.g., for health, elections, hiring), and computational power used. The FTC may consider alternative rules for open-source models or models built on others. Regulations apply 90 days after issuance. The FTC must update standards at least every 2 years. Violations are treated as unfair or deceptive acts under FTC law, with enforcement powers. A report to Congress is due in 2 years. Authorized funding is $10 million for fiscal year 2025 and $3 million yearly after. A covered entity is defined as a person or corporation under FTC jurisdiction that provides a foundation model with over 100,000 monthly outputs or 30,000 monthly users. A foundation model is an AI model trained on broad data, using self-supervision, with at least 1 billion parameters, applicable across many contexts, and capable of high-risk tasks.