New York launches first US state-run AI testing center with $55 million in funding

While Washington debates whether AI companies need mandatory rules, New York has begun building its own evaluation system. Governor Kathy Hochul has announced the launch of the first state-run program in the United States for independent testing of AI systems and training a new generation of neural network safety experts. If you build or deploy AI products that operate in the state, this deserves a close read.

What is being launched

The center will receive funding from two sources: Thomas Secunda, Bloomberg’s co-founder, will donate $30 million, while SUNY, the State University of New York, and the state government will contribute another $25 million. That brings the total to $55 million. Computing power will be provided by Empire AI, a public-private consortium with a $500 million budget. Secunda chairs its board.

More than 80 leading AI researchers from across the state will work on the program. Formal responsibility rests with Binghamton University’s Center for AI Responsibility and Research. The program is led by Jeannette Wing of Columbia University and Greg Morrisett of Cornell Tech.

The idea is straightforward: government officials should have their own source of evidence. Previously, they had little choice but to trust developers’ own reports. Secunda says explicitly that the center will be able to test systems independently and provide authorities with evidence they can rely on. Hochul put it this way: the stakes are too high to take AI companies at their word.

Why now

The context is political. President Donald Trump declined to introduce federal AI safeguards and described safety concerns as a “hoax.” Instead of strict rules, the administration opted for a voluntary system whose details were disclosed only to technology companies. New York is responding with measures of its own.

The models themselves have also raised concerns: according to the source, they have already demonstrated the ability to hack information systems, while leading researchers at major labs have publicly warned about the risks. Wing highlights a separate challenge involving agents. Swarms of autonomous agents act independently and without supervision, creating a need for methods that can detect in advance when agents are about to cause harm and help prevent it. For anyone building multi-agent systems, this is a familiar pain point: even within a team, tracking exactly what an agent did and why remains difficult.

The connection to the RAISE Act

The center is scheduled to launch before January 1, 2027, when the state’s main AI safety law, the RAISE Act, takes effect. According to the report, AI companies operating in New York will have to provide detailed safety frameworks. The governor’s office expects researchers at the new center to scrutinize those documents.

The center will therefore serve both as a laboratory and as a source of expert review for the regulator. The announcement does not specify which tests it will conduct, what criteria it will use or what consequences companies could face. Those details will emerge later, and they will determine how the system works in practice.

What to do in practice

First, if you have users or operations in New York, it is time to establish whether the RAISE Act applies to you. Have a lawyer review the law before January, rather than afterward.

Second, prepare an internal document explaining how you assess the risks of your model or agent system: what tests you run, how you restrict access and what you do when something goes wrong. If outside experts will be reading these frameworks, vague language is unlikely to suffice.

Third, keep an eye on California. New York and California have moved well ahead of other states in regulating AI companies, while Democrats in Congress reportedly plan to subpoena executives at major AI companies and pass legislation on safety and export controls. Regional rules could become a de facto standard, as happened with the GDPR.

There is one caveat to all this: it is still an announcement. There is no list of tests, no penalties and no confirmation that companies will cooperate willingly. But the direction is clear: independent AI evaluation is becoming a government function, and it is best to prepare in advance.