Innovation & Tech

New York Creates AI Unit as RAISE Penalties Reach 15%

New York’s ASSET Unit will enforce the RAISE Act, which requires 72-hour incident reports, annual audits and compute-cost penalties from covered frontier-model developers.

By Nathan Brooks

5 min read

New York Creates Dedicated Unit to Enforce AI Safety Law
New York Creates Dedicated Unit to Enforce AI Safety Lawfrabre / pixabay

What's News

  • Civil penalties can reach 5% of total frontier-model compute costs for a first violation and 15% for subsequent violations.
  • Safety incidents must be reported to the Division of Homeland Security and Emergency Services within 72 hours.
  • Critical harm covers death or serious injury to at least 100 people, or at least $1 billion in damages.
  • Covered large developers must spend more than $100 million in aggregate on frontier-model training.
  • The ASSET Unit will begin work before the Jan. 1, 2027 effective date listed in the attorney general’s announcement.

By Nathan Brooks

New York State Senate Bill 2025-S6953A sets civil penalties of up to 5% of total frontier-model compute costs for a first violation of the Responsible AI Safety Education Act, and up to 15% for subsequent violations.

Attorney General Letitia James announced Tuesday that her office will create the AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit to lead enforcement of the law. The announcement described the measure as a pioneering state statute with mandatory safety, transparency and incident-reporting duties.

Cristy Phillips, formerly Special Counsel for the Economic Justice Division, will lead the unit. The office said it is “currently building the ASSET Unit team.”

The unit will operate inside the Attorney General’s Economic Justice Division. It will begin work before the announcement’s Jan. 1, 2027 effective date.

Who qualifies under the RAISE Act?

The text of New York State Senate Bill 2025-S6953A defines a “frontier model” as an AI system trained with more than 10^26 computational operations and a compute cost exceeding $100 million. It also covers models derived from those systems through knowledge distillation.

The law defines a “large developer” as any entity that has trained at least one frontier model costing over $5 million and spent more than $100 million in aggregate on frontier-model training. Accredited colleges and universities conducting academic research are excluded.

What must developers do before deployment?

The bill imposes these safety, transparency and operating duties:

  • Deployment protocol. Before deploying a frontier model, covered developers must implement a written safety and security protocol specifying protections against “critical harm.” The bill defines critical harm as death or serious injury to 100 or more people, or at least $1 billion in damages, caused or materially enabled by the model.
  • Disclosure and access. Developers must publish a redacted version of the protocol and transmit it to the Division of Homeland Security and Emergency Services. The division or attorney general can obtain the unredacted protocol on request, with redactions only as federal law requires.
  • Testing records. Developers must retain detailed testing records sufficient for third-party replication.
  • Annual audit. Covered developers must undergo an annual independent third-party audit of compliance. They must publish a redacted report and send it to the Division of Homeland Security and Emergency Services. The division or attorney general can obtain the unredacted report on request, and only federal law can require redactions.
  • Incident reporting. Each “safety incident” must reach the Division of Homeland Security and Emergency Services within 72 hours, or three days. Covered events include autonomous model behavior, theft or escape of model weights, critical control failures and unauthorized use.
  • Risk test. Developers may not deploy a frontier model if deployment would create an unreasonable risk of critical harm.
  • Whistleblower protection. The law bars retaliation against employees who disclose safety concerns to the developer or the Attorney General. Civil penalties reach $10,000 per employee for retaliation.

What changes for frontier-model developers?

For a covered developer, the bill turns model safety into a documented, audited operating requirement rather than a voluntary policy. Compliance starts before deployment and continues through annual audits, record retention and incident reporting.

The 72-hour incident window leaves three days to notify New York. The repeat-offense penalty is three times the first-offense rate, while retaliation can cost up to $10,000 per employee.

For companies near the training thresholds, the commercial question is whether they qualify as large developers at all. StateScoop said the current thresholds capture only the largest models, leaving the qualifying company count unclear.

Which other laws will ASSET coordinate?

James’s announcement says ASSET will coordinate implementation of:

  • SHIELD Act: data security requirements.
  • Chatbot safety law: takes effect June 2027 and bans unsafe features for children.
  • General Business Law Article 47: suicidal-ideation detection and AI disclosure for chatbots.
  • Penal Law §245.15: criminalizes nonconsensual intimate imagery, including AI-generated imagery.
  • Synthetic Performer Disclosure Act: disclosure of AI use in advertising.

Beyond enforcement, ASSET will develop policy proposals and provide technical expertise to other Attorney General divisions. It will also work closely with the existing Bureau of Internet and Technology (BIT).

Why is New York moving now?

A November report from the National Association of Chief Information Officers found that at least 46 states have identified someone responsible for directing AI initiatives. Yet 56% lack dedicated AI budgets, and 75% cite insufficient staffing as a major challenge.

StateScoop reported that California Governor Gavin Newsom recently ordered accelerated development of an independent AI safety auditing system. StateScoop described New York’s approach as a dedicated attorney general team focused on enforcing AI laws and investigating potential misconduct by technology companies.

Can federal policy displace New York’s rules?

In 2025, President Trump issued an executive order barring states from enforcing their own AI laws. StateScoop said the move drew widespread backlash from lawmakers and advocacy groups. The dispute over state authority remains unresolved.

Separately, James led a bipartisan coalition urging Congress to reject federal legislation that would preempt state regulations, according to her office. According to StateScoop, a coalition of state lawmakers warned against the “Great American AI Act” discussion draft in June 2026.

The proposal would authorize a federal AI standards center and preempt state laws. The coalition submitted a letter to Congress urging opposition, StateScoop reported. James has consistently argued that states must retain the ability to act where federal legislation stalls, her office said.

What remains unresolved?

The Attorney General’s announcement lists Jan. 1, 2027, as the effective date. The bill says the law takes effect 90 days after becoming law on Dec. 19, 2025. Neither the announcement nor the bill explains the discrepancy.

StateScoop said the 10^26-operations and $100 million compute thresholds capture only the largest current models. It remains unclear how many companies will qualify as large developers.

The effective-date conflict adds urgency to a practical test: whether New York can turn a high-penalty statute into verifiable rules. That will depend on staffing, audit standards and enforcement choices the Attorney General’s office has not specified.

Source: epyhia.com

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,426 articles

Related articles

« Previous articleNext article »