Pennsylvania lawmakers weigh AI safety bills with audits and a deployment ban

Pennsylvania lawmakers examined two proposals for policing the most advanced AI systems at an October 19, 2026 House committee hearing. Both would require public safety disclosures and independent audits. One goes further: it would prohibit deployment when a developer’s own assessment finds an unreasonable risk of catastrophic harm. Neither bill has been enacted.

Rep. Melissa Shusterman sponsors HB 2800, the Artificial Intelligence Risk Prevention Act. Rep. Manuel Guzman sponsors HB 2829. Both presented their proposals to the House Communications and Technology Committee.

Revenue and computing power set the boundary

According to the Transparency Coalition’s bill guide, both drafts define large developers as those with annual gross revenue above $500 million. Frontier models must exceed a training threshold of 10²⁶ computing operations.

Both define catastrophic risk using alternative harm thresholds: more than 50 deaths or injuries, or more than $1 billion in damage.

Large developers would publish and follow safety frameworks covering risk thresholds, safeguards, pre-deployment and internal-use reviews, and incident response. Annual reviews would be mandatory; material changes would require publication and justification within 30 days.

Transparency reports would accompany new or substantially modified models, describing intended uses and restrictions. Large developers would add machine-readable risk-assessment summaries. A model or system card—a document describing a model—could satisfy the requirement.

Outside checks and fast incident notices

Annual independent audits would check compliance. Auditors would need frontier-safety expertise and financial independence; payment could not depend on results. Developers would publish audit summaries and redacted reports.

Reporting and worker protections

• HB 2800 sets a 72-hour critical-incident reporting deadline after reasonable belief an incident occurred. HB 2829 also sets a 72-hour deadline.
• Both require notice to authorities within 24 hours for imminent risk of death or serious injury.
• Both protect qualifying whistleblower disclosures against retaliation and require anonymous internal reporting channels at large developers.

When a risk assessment would stop deployment

HB 2829 would prohibit a large developer from deploying a model if its own assessments show an unreasonable risk of catastrophic-scale harm. It also requires test records detailed enough for outside replication and broadens the incident definition to include a model acting on its own.

At the hearing, Transparency Coalition advisor Steve Wimmer recommended explicitly requiring loss-of-control reports to PEMA and the attorney general. The AI accountability and safety advocacy nonprofit supported the direction of both bills.

HB 2800 specifies January 1, 2028 for safety-framework implementation. The coalition flags a possible conflict with its separate effective-date timetable, which depends on publication of PEMA’s temporary regulations.