New York Is struggling to track its own AI use

This article was originally published by New York Focus.

In New York’s first-ever effort to track its own use of artificial intelligence, only 13 of the state’s roughly 100 agencies publicly reported using the technology in a way that could “directly impact the public” — fueling concerns that the state’s AI transparency guidelines are too ambiguous for effective oversight.

The disclosures were the result of an inventory under the state’s 2024 policy governing AI use, part of Governor Kathy Hochul’s efforts to boost transparency around the technology. But several high-profile AI programs do not appear in the inventory, New York Focus found. 

Agencies began self-reporting their AI use to the state’s Office of Information Technology Services, or ITS, last March, and the results were published on December 30. As agencies prepare to start their second round of reporting in the coming months, critics say, the inconsistencies raise questions about the state’s ability to effectively track and disclose its use of the rapidly growing technology.

The state broadly defines AI as any machine-based system that directly impacts the public by making predictions, recommendations, or decisions. This includes analyses used by law enforcement and to determine eligibility for public benefits, and natural language processing.

The 13 agencies in the inventory reported 19 computer programs in total, ranging from common apps like Google Translate — which the Department of State uses to translate forms — to AI companion robots deployed by the Office for the Aging to combat loneliness among seniors. 

In a February budget hearing, ITS Director Dru Rai acknowledged the 19 inventory disclosures seemed low, but said he wasn’t sure if that was because submissions were withheld for safety or privacy concerns. “That’s why I would guess the list is smaller than it should be,” he said. A note published with the inventory mentions that “back-office or purely internal AI tools” were not included. A disclaimer also warns the disclosures haven’t been “independently validated” by ITS. 

Reporting guidelines allow agencies to request their AI tools be withheld from disclosure for privacy or safety concerns, but it’s unclear how many submissions were omitted from the 2025 inventory for those reasons.

Several programs that have been publicized elsewhere did not appear in the inventory. For instance, the inventory includes two AI tools that the Department of Motor Vehicles uses to verify documents and translate forms, but does not include AI-powered facial recognition technology that the department has been using since 2010 to detect fraud. The DMV did not respond to requests for comment.

The New York State Police and the state’s Department of Homeland Security and Emergency Services, meanwhile, currently use Dataminr — an AI-powered threat monitor that has received backlash for surveilling protesters — according to state contract records. Only the latter agency’s use of the platform appears in the AI inventory. The New York State Police doesn’t appear on the inventory at all, though reports have shown that it has heavily invested in various forms of AI over the past decade.

Spokespeople for the State Police and ITS did not clarify whether the State Police submitted programs to ITS for review.

In an email, a State Police spokesperson did not outline which programs the agency was using and wrote that “any systems that State Police use with artificial intelligence components are used in accordance with the State’s Acceptable AI Use Policy,” also noting that the department uses Dataminr “for notifications only, not for criminal investigations.”

The state Office of Children and Family Services is also absent from the inventory despite having publicized its subcontractor’s use of AI software to monitor and train its caseworkers. The reporting mandate should cover contractors that work with state agencies. The inventory also did not include proprietary AI chatbots used by at least seven county social services departments, which work closely with the state’s Office of Temporary and Disability Assistance. The chatbots have helped county departments manage increased call volumes as a result of eligibility changes to public benefit programs like the Supplemental Nutrition Assistance Program.

OCFS did not provide comment. In an email, a spokesperson for Onondaga County said that it did not participate in the 2025 inventory, but plans to participate in the upcoming reporting round. The spokesperson noted that it has notified state officials at OTDA about its AI chatbot. The six other county social services departments — Chenango, Clinton, Greene, Jefferson, Oswego, and Orleans — did not respond to requests for comment. 

The state’s policy is ambiguous, said Clara Langevin, senior manager of AI policy at the Federation of American Scientists. 

“I could see state agencies having difficulties interpreting this guidance,” she said. “Internal uses of AI that may seem routine, such as prioritizing grant or permit applications, flagging cases for investigation, or helping triage child welfare cases, could reasonably be considered to affect the public.” 

“Without additional guidance or examples, agencies could reach different conclusions about whether these systems fall within the reporting requirements,” she said.

ITS updated its AI policy in June, but those changes still leave questions. For instance, the inventory reporting requirement previously applied to “state entities” but has been narrowed to “state agencies.” But the agency hasn’t clarified whether other types of government bodies, like the Metropolitan Transportation Authority, are subject to reporting. ITS did not respond to a request for comment on the updates.

The reporting gaps didn’t come as a surprise to state Senator Kristen Gonzalez, who has advocated for stricter, more transparent AI safeguards.

Gonzalez pointed to an April 2025 State Comptroller audit, which found that the state’s AI governance policy as of late 2024 contained inadequate guidance and left agencies ill-equipped to fully and accurately identify the technology due to “non-existent training.” The same report predicted the upcoming inventory would not identify “all AI use cases.”  (ITS updated its policy in March 2025 in response to the comptroller’s findings). 

ITS spokesperson Scott Reif told New York Focus that the agency recently provided additional documentation and training to agencies to improve future reporting, but did not provide more details. He also did not say how many agencies submitted AI programs for review or specify which agencies are required to participate.

Gonzalez blamed Hochul, who she said watered down AI transparency provisions originally proposed by lawmakers a few years ago. Gonzalez introduced the LOADing Act, which oversees AI use by state agencies, in 2023; the final version Hochul signed in late 2024 allows the state to redact details in the public version of the inventory. “When trust in government is at an all time low, transparency has to be a priority,” Gonzalez said.

Other states have also struggled with incomplete AI inventories. California reported that none of its government agencies used AI in a high-risk way last year, despite public controversies with several programs.

Both Gonzalez and Langevin pointed to the federal government’s more successful efforts at an AI inventory. Langevin outlined the federal government’s more “robust” reporting criteria, which requires agencies to determine whether the AI being used could negatively impact people’s civil rights, privacy, health, or safety. Federal agencies also go through a more detailed multi-stage review process, she noted, where AI systems are evaluated for higher-impact uses. 

In an email, Reif defended the agency. “New York State has an extraordinary record when it comes to promoting the safe, responsible, and transparent deployment of artificial intelligence technologies,” he wrote. “ITS takes its responsibility in this area very seriously and we remain fully engaged with our agency partners in supporting the safe, responsible, and transparent adoption of AI.”

While the reporting mandate applies to some local governments that work closely with state agencies, like county social services departments, it doesn’t appear to cover other cases like the state’s Division of Criminal Justice Services. The agency runs LETECH, a popular grant program that provides state funds to local counties for a wide array of pricey and controversial police tech, ranging from drones to gunshot detection devices. Earlier this month, Hochul announced $75 million in new LETECH funding. 

AI-powered license plate readers funded by the state program have sparked intense community backlash across the state, including in Suffolk County, Ithaca, Newburgh, and Saranac Lake. The company behind the cameras, Flock Safety, has been under heightened scrutiny nationwide amid the federal government’s ongoing immigration crackdown. Local police departments have accessed the camera feeds to aid agencies like Immigrations and Customs Enforcement.

Data gathered from the tech devices feed into criminal intelligence hubs maintained by DCJS across the state, but DCJS does not appear in the AI inventory. Aaron Cagwin, director of public information for the criminal justice agency, said that ITS determined it “did not employ any technology products that met the state’s definition of AI.”

“Decisions about how grant funds are used are made by the individual agencies within program guidelines,” he wrote. “While DCJS generally monitors emerging technologies and trends in law enforcement, it does not currently track or evaluate the use of AI technologies by grant recipients.” 

Hochul has emphasized the importance of AI disclosure in maintaining public trust, mitigating risk, and cementing the state’s trailblazing reputation. “New York is once again leading the nation in setting a strong and sensible standard for frontier AI safety,” Hochul said in December, while announcing new reporting guidelines for AI developers. 

“Governor Hochul remains committed to ensuring that NY remains at the forefront of the AI revolution,” Hochul spokesperson Kassie White wrote in an email. “That includes better serving New Yorkers by both expanding AI education and training to the NYS workforce as well as ensuring that agencies are transparent about the uses of AI systems that directly impact the public.”

Several of the AI programs that were disclosed in the inventory have been embroiled in lawsuits. Eightfold, an AI-powered hiring platform contracted by the Labor Department to help job seekers, is being sued for potentially illegal screening practices. AI-generated narration used in instructional materials by the state’s Higher Education Services Corporation is part of a class action suit alleging voice theft. Software used by the state’s Office for People With Developmental Disabilities to detect distracted driving was at the center of a $4.25 million privacy settlement in 2025. 

In an email, the Labor Department said it is using Eightfold to “augment the job matching the Department has always done, helping make the employment search smarter and broader for jobseekers and businesses.” Lisa Pink, a public information officer for DOL, noted the agency completed a risk assessment to ensure that Eightfold complies with the state’s AI policy, but did not say whether it was taking additional precautions or received direction from ITS on how to move forward in light of the lawsuit. HESC and OPWDD did not respond to requests for comment.

When Gonzalez pressed ITS Director Rai during a budget hearing in February about how his agency was managing potential risks like Eightfold, he put the onus on the Labor Department. “We are there to help, but they are responsible for taking the risk,” he said.

The recently updated ITS guidelines require agencies to conduct risk assessments of AI tools largely on their own and eliminated a provision that would have required certain state agencies — including the Labor Department — to conduct those assessments in conjunction with ITS. 

The next reporting round starts this fall and will be posted online on December 30. Langevin called accurate AI inventories critical to any meaningful oversight. She urged New York to adopt a clearer definition for reportable AI use, outline specific procedures, and provide sector-specific guidance to ensure all agencies are well-equipped to govern their AI systems as the technology becomes increasingly ubiquitous.

“You can’t meaningfully evaluate the benefits or risks of government AI if you don’t know where it’s being used,” she said. “Inventories enable oversight bodies, researchers, and the public to ask informed questions, identify patterns across agencies, and evaluate whether governance frameworks are working as intended.”