High-risk AI developers would follow federal standards and share models

Covered developers would follow applicable federal standards and give the AI Safety Board access to their models. This covers developers of models that perform, or could be modified to perform, at high levels on tasks posing serious risks to national security, the national economy, or public health or safety. Standards violations could draw fines up to $250,000 per violation, with each continuing day counted separately. Access would include the files needed to run the model. Access would be due at least 45 calendar days before introducing a covered model into trade across state or national borders.

Covered developers would publish, file, and follow safety plans. The duty would cover each AI system that a high-risk AI developer creates, substantially modifies, or uses to train or evaluate other models. Plans would assess capabilities or risks, list steps to reduce those risks throughout the model’s life, and name the responsible corporate officer. Developers would file plans with the Commerce Secretary.

Covered AI companies and infrastructure operators would report confirmed AI safety and security incidents. Reporting would cover developers and providers of high-risk AI models, plus critical infrastructure operators using AI in industrial controls or other operating technology. For covered AI companies, reporting would apply throughout development, testing, and operation, even for models never intended for public release.

The National Institute of Standards and Technology would maintain a public database of AI incidents and flaws. Shared reports would be shielded from public disclosure unless affected parties consent to being named in the database. The agency would consult the developer or provider before publishing incident information. The database would be due within one year of enactment.