Pennsylvania would set safety rules for powerful AI developers

Large frontier developers would maintain a public safety framework, arrange independent audits, report internal-use risks, and withhold models posing unreasonable catastrophic risk. These duties would fall on a developer whose revenue, combined with its affiliates, exceeded $500 million in the prior year—not automatically on each affiliate. Covered models must be developed, deployed, or operating at least partly in Pennsylvania and trained using more than 10^26 computing operations. The bill would not prohibit lawful AI research, development, testing, or evaluation. The framework requirements would start January 1, 2028. Audits would begin then or 90 days after later qualification. Internal-use summaries would be due every three months unless another reasonable schedule is communicated to both state offices and updates are agreed upon by the agency.

Frontier developers would publish model information and report critical safety incidents to state officials. The public model report would cover every frontier developer; large developers would also summarize catastrophic-risk reviews and results. Public documents could omit information when needed to protect trade secrets, cybersecurity, public safety, national security, or compliance with other laws, but developers would have to explain permitted omissions. State officials would restrict access to internal-use reports, which would not be available under Pennsylvania’s public-records law. A model report would be due before or when a new or substantially modified model is deployed. Every critical safety incident would be reported to the agency and Attorney General within 72 hours after the developer learns enough facts to reasonably believe it occurred. An imminent threat would also be disclosed within 24 hours to an appropriate authority.

Large frontier developers would file disclosures, pay administrative fees, and face state enforcement for listed violations. The filing, fee, and civil-penalty provisions would apply to developers meeting the bill’s $500 million combined-revenue test. Each developer would pay its share of estimated current administrative costs and earlier deficits. The Attorney General could seek penalties or court orders, but the bill would not let private parties sue solely under this act. Current disclosures and fees would be required beginning January 1, 2028. Listed safety violations could carry up to $1 million for a first violation and $3 million for each later violation.

Frontier developers could not block or retaliate against certain workers for reporting serious safety dangers or violations. These protections would cover employees responsible for assessing, managing, or addressing critical safety incidents. Protection would require reasonable cause to believe the information shows a specific, substantial public danger from catastrophic risk or a violation of the act. Large developers would provide an anonymous internal process, monthly status updates to the reporting worker, and regular information to company leaders, except leaders accused of wrongdoing. Developers would give covered workers continuing workplace notice or written notice at least once each year.

The state agency could recognize equivalent federal rules, while Pennsylvania would bar local governments from separately regulating frontier models. The local restriction would cover counties, cities, townships, and other Pennsylvania local governments, but only rules that specifically regulate frontier AI models. Federal compliance would count only for a state requirement the agency specifically finds equivalent or weaker, and only while the developer follows the designated federal rule. Failing to follow a designated federal rule after choosing this route would violate the act. A person would not be covered solely for developing, deploying, or operating a model that is not a frontier model. These provisions would take effect January 1, 2028.