Researcher Stephen Casper dissects New York RAISE Act amendments point by point

New York State Senators Gounardes and Fahy introduced a revised version of the RAISE Act (Responsible AI Safety and Education Act), amending Section 1420 of the commerce law. AI safety researcher Stephen Casper posted a point-by-point thread on October 31, calling the update substantial and long-awaited by many in the field. He listed the 11 changes he appreciates most, flagged what's still missing, attached detailed notes, and said he'd welcome a direct conversation with the sponsoring senators.

Confirmed

• The compute threshold for frontier models lowered to 10^25 operations; Casper considers this necessary if frontier developers' read of current trends is credible
• The definition of "catastrophic risk" broadened to deaths of 50+ people or over $1 billion in losses from "one or more events"; the "critical safety incident" definition expanded to include some near misses
• A new state office established to issue regulations and orders; large frontier model developers must designate an internal executive responsible for compliance
• Third-party safety assessments shift from optional to mandatory; the bill specifies in the most concrete legislative language to date what assessors can access inside companies, allows assessors to contest company redactions, and requires the state to license assessors
• More transparency and reporting required for internal model deployments, including research uses
• New whistleblower protections added; the state gains emergency powers to restrict model training, inference, or automated development; the state attorney general is explicitly authorized to sue for injunctions to enforce the law

Casper's suggested fixes

• Borrow from the Sarbanes-Oxley Act (post-Enron legislation): require frontier developers to retain records of catastrophic risks, critical safety incidents (including potential ones), and compliance documentation for at least 5 years
• Contracts between assessors and frontier developers should be public (with necessary redactions)
• Under the current draft, even lying to the public about an imminent catastrophic risk or defying an order to shut down a high-risk AI system carries no criminal penalties such as imprisonment
• Compute counting needs clarity on integer vs. floating-point operations, and he offers concrete suggestions; assessor access rights, while the most specific in any comparable bill, still need refinement and a dispute-resolution process

Why it matters

The RAISE Act is among the most operational frontier-AI regulatory bills at the US state level; its revisions on compute thresholds, mandatory assessments, and emergency powers will directly shape frontier developers' compliance obligations, and Casper's technical suggestions offer a reference for the enforcement details of such legislation.

2026-10-31 ~ 2026-10-31 · 14 related posts

• [source] NY RAISE Act update broadens 'catastrophic risk' definition for frontier models — StephenLCasper · 2026-10-31
• [source] Researcher reads NY RAISE Act update, flags strengths and missing fixes — StephenLCasper · 2026-10-31
• [source] NY RAISE Act revision: 11 key updates, from 10^25 threshold to whistleblower protections — StephenLCasper · 2026-10-31
• NY RAISE Act revision lowers frontier compute threshold to 10^25 operations — StephenLCasper · 2026-10-31
• NY RAISE Act revision details evaluator access and internal deployment transparency — StephenLCasper · 2026-10-31
• NY RAISE Act revision mandates compliance officers and third-party assessments — StephenLCasper · 2026-10-31
• NY RAISE Act revision lets evaluators dispute redactions, grants state emergency powers — StephenLCasper · 2026-10-31
• NY AI bill details: state can emergency-restrict training and inference — StephenLCasper · 2026-10-31
• NY RAISE Act revision adds whistleblower protections — StephenLCasper · 2026-10-31
• Casper flags ambiguity in how RAISE Act counts integer and floating-point operations — StephenLCasper · 2026-10-31
• Casper: NY RAISE Act most specific yet on evaluator access, still needs work — StephenLCasper · 2026-10-31
• Casper proposes Sarbanes-Oxley-style document preservation for AI safety in NY RAISE Act — StephenLCasper · 2026-10-31
• AI safety researcher flags draft AI Act gaps: no jail time even for lying about catastrophic risks — StephenLCasper · 2026-10-31
• Casper: RAISE Act drafts lack public auditor contracts and criminal penalties for lying — StephenLCasper · 2026-10-31