New York already has its own mechanism to examine the artificial intelligence that companies put into circulation. Gov. Kathy Hochul presented a state program that instructs independent researchers to review the work of the developers, instead of accepting the evaluations that companies make of themselves. The announcement was made public on Tuesday and is presented as the first of its kind in the United States.
Reasoning is direct: self-regulation is not enough. Hochul summed it up in a phrase that accompanies the initiative.
“The Stakes Are Too High to Simply Take AI Companies at Their Word”.
The risks are too high to simply accept the word from AI companies.
This is how the State AI tests will work
The work lies in a center based at Binghamton University, whose creation the Governor’s Office had anticipated in the State Statement of January. Its name circulates with two different denominations in the documentation released so far, without official confirmation on which will be the definitive one.
The direction corresponds to Jeannette Wing, Executive Vice President of Research at Columbia University, and Greg Morrisett, Vice Chancellor of Cornell Tech. More than 80 scientists and researchers will participate in the program throughout the state.
Computational fuel comes from Empire AI, the state’s $500 million public-private partnership. The center’s initial funding combines a donation of 30 million Tom Secunda, co-founder of Bloomberg and chairman of the Council of Empire AI, with an investment of 25 million SUNY and New York State.
Tasks go beyond issuing a verdict on a specific model. The center will test advanced systems, study its risks, and gather evidence for public office to decide with more than part reports. It will also analyze responsible uses in medicine and scientific discovery.
The weak point: only what is published is evaluated
There is a limit that should be clear. Researchers only agree to what AI companies publicly disclose; The models that are still in development within the laboratories are out of reach. State AI tests will not be able to review internal versions before they see the light.
This gap is covered in part by the Raise Act, signed by Hochul on December 19, 2025, and with entry into force scheduled for January 1, 2027. From that date, the great developers must register in the State, publish frameworks of Security and transparency reports and reporting critical security incidents within 72 hours.
The standard targets frontier model developers with incomes over $500 million and leaves out universities and centers dedicated to academic research. The boundary model definition is tied to a computing threshold of 10^26 training operations, or to distillates of those models with costs above $5 million.
Wing warned in the presentation about the difficulty of detecting and preventing damage when swarms of self-contained agents operate without control. That kind of scenario is just what the center wants to document with its own evidence.
Mandatory registration and pulse with Washington
The calendar is not without friction. Trump’s executive order on federal AI standards, which bets on a “minimumly burdensome” framework and calls on the Department of Justice to challenge state laws that contradict it, arrived eight days before Hochul signed the RAISE Act. challenge, and his final text had not yet been published when the analysis of the Skadden office was disseminated.
Among the security advocates in AI there are those who consider that the amendments to the Raise Act weakened the original version of the state legislature. Meanwhile, Democrats are preparing a legislative AI offensive for the next Congress, with citations to sector executives and proposals on AI security, online minor protection and export controls.
For now, it is not known which models or which companies will go through the AI tests of the center, nor what will be its definitive official name. What is defined is the change of logic: the state stops trusting self-regulation and begins to build its own verification capacity.
Read also
- Take-Two strengthens its technical and legal defenses against AI
- Gemini 4 Carbon arrives at AI Studio and expands its global access
- Anthropic launches Claude Sheets and brings the office to Claude
Source: Engadget News
