GAAIA draft bill offers federal AI transparency for state preemption trade

The Great American AI Act (GAAIA), released last month by Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA), is a 269-page discussion draft that aims to establish federal oversight of AI development. Its core trade: for three years, states cannot enforce laws that specifically regulate pre-deployment AI model development (covering any model, not just frontier ones), while post-deployment issues like deepfakes and chatbot disclosure remain state purview. In exchange, large frontier developers—those training models above 10^26 operations and with over $500M in annual revenue—must file published safety frameworks, report incidents to a new Center for AI Standards and Innovation (CAISI) at the Commerce Department, and undergo semi-annual audits by licensed independent verification organizations. Whistleblower protections are the strongest provision and cover the entire industry. However, the bill is purely a "visibility" bill: even if an audit reveals problems, the developer can still deploy, facing only disclosure and penalties. This approach reduces pushback from pro-business factions but leaves safety advocates concerned.

The draft, while well-crafted overall, contains several technical issues. Section 131 raises the maximum fine for mail and wire fraud from $1M to $2M, but adds an AI enhancement clause that caps at $1M – meaning using AI could lower the maximum penalty. A cross-reference points to definitions instead of the intended section, and a citation references a rescinded 2025 executive order. More critically, GAAIA lacks a severability clause; under default doctrine courts would sever unconstitutional parts, risking the entire preemption deal. Canary Institute is compiling a technical corrections annex and will submit it to the sponsors' public inbox. Despite these bugs, the bill represents the most serious federal AI legislation to date, borrowing from aviation safety principles to focus on transparency and incentives rather than blame.