Pennsylvania Lawmakers Push Catastrophic-Risk AI Bills Into a Purple-State Test
On Monday, the Pennsylvania House Communications and Technology Committee spent the morning on a subject state legislatures have mostly left to California, New York, and Illinois: what happens if a frontier AI model causes mass harm, and who finds out first. House Bill 2800, from Rep. Melissa Shusterman, and House Bill 2829, from Rep. Manuel Guzman, would require the largest AI developers to publish safety frameworks, report serious incidents to the state's Attorney General and emergency management agency, and protect employees who blow the whistle on catastrophic risks. Neither bill has had a vote. The hearing was the first public airing of either one.
A template, not an experiment
What's notable isn't the substance, which is almost a copy of California's SB 53, the Transparency in Frontier Artificial Intelligence Act that Gov. Gavin Newsom signed in September 2025. That law requires large frontier developers to publish an annual framework describing how they identify and mitigate catastrophic risk, report critical safety incidents within 15 days (24 hours if there's imminent danger), and shield whistleblowers from retaliation. Pennsylvania's HB2800 reproduces that structure almost exactly, down to routing incident reports to a state emergency-management agency rather than a tech-specific regulator. That's a sign of how this legislative category now works: once one state writes the template, others don't redesign it, they localize it. New York's and Illinois's versions followed the same pattern earlier this year. Pennsylvania is simply the next adopter, and arguably the most politically exposed one, since it's a competitive swing state rather than a reliably Democratic one.
A quiet hearing, for now
What coverage of Monday's hearing doesn't mention is as telling as what it does. When California debated SB 53, the record shows an aggressive, well-funded industry campaign against it, with OpenAI and venture firms lobbying hard for a lighter bill. Reporting on Pennsylvania's hearing names no comparable opposition: witnesses were described only as AI researchers and safety advocates, and neither bill's committee record shows testimony from a frontier lab, a trade association, or a venture-backed opponent. That could mean industry hasn't mobilized here yet, that a House hearing this early in the process isn't worth the lobbying spend, or simply that local coverage didn't capture it. All 14 listed co-sponsors of HB2800 are Democrats, and the bill now sits in a chamber the party holds narrowly, en route to a Senate historically less receptive to new tech regulation. Whether this becomes a fight like California's, or quietly stalls the way most state AI bills do, is the real open question the hearing didn't resolve.
The sponsors leaned on two local incidents to make the case for urgency. The first: a fabricated homicide tip submitted to Philadelphia police in July, which lawmakers attributed to an AI system. The second: OpenAI's own disclosure last month that one of its models had interacted with several U.S. government websites, which the bill's backers framed as evidence that even leading labs can't fully account for what their systems do once deployed. Neither incident, as described, rises to the mass-casualty scenarios the bill's 'catastrophic risk' language is built around — which is itself a useful data point: Pennsylvania's lawmakers are building a disaster-response bill using the only AI incidents they currently have, which are closer to nuisance and security-adjacent events than the biological or infrastructure-scale harms the statute's text invokes.
"These systems are getting stronger every month — the companies themselves say so. Washington hasn't written a rule. That leaves it to the states, and without guardrails, there's no guarantee anyone can control what these systems do." — Rep. Manuel Guzman
Not everyone pushing for the bill thinks it goes far enough. Steve Wimmer of the Transparency Coalition, an advocacy group that has backed similar laws elsewhere, told the committee the bill should explicitly require developers to report any loss of control over a model directly to the Attorney General and the Pennsylvania Emergency Management Agency, not just file it inside a broader incident report. That's a meaningful gap: SB 53's California version and its state cousins generally bundle 'loss of control' into the same reporting bucket as other safety incidents, which critics argue lets a genuinely novel failure mode get buried alongside routine disclosures. If Pennsylvania's final language doesn't separate the two, it will have reproduced California's structure, including the same weakness advocacy groups have flagged in the original.
The polling numbers lawmakers cited, 86% support for independent AI safety standards and 73% concern about AI as a threat to human survival in a statewide Quinnipiac survey, are consistent with national polling on the subject and suggest the politics may be less lopsided than Pennsylvania's divided government implies. But polling support for AI safety in the abstract has not reliably translated into votes when it collides with a legislature's appetite for data-center jobs and investment, a tension this same committee has navigated all year on more industry-friendly AI bills. HB2800 and HB2829 still need a committee vote, two floor votes in the House, and Senate concurrence before they reach the governor's desk. California's law took over a year, several amendments, and one prior veto to get there. Pennsylvania's version is only a few weeks past introduction, and this week's hearing was its first real test of whether that path looks any shorter the second time a state tries it.