Anthropic Splits From Google, OpenAI Over State AI Safety Bill

Anthropic is at odds with other big tech firms over a Massachusetts Senate proposal requiring big AI developers to hire independent evaluators to assess catastrophic risks posed by their models every four months.

The Massachusetts proposal could set a new standard for AI regulation, which state legislators across the country are pursuing aggressively even as congressional lawmakers drag their feet on passing federal legislation. The proposal also comes as concerns are mounting about the dangers of rogue AI agents in the wake of the OpenAI–Hugging Face attack, when hundreds of agents figured out how to communicate with each other and coordinate an attack on Hugging Face.

Anthropic and safety groups support the AI language in the legislature, which is opposed by companies such as Google and OpenAI.

The legislation is tucked into a broader economic development bill before the state upper house. A conference committee of the two Massachusetts chambers is now meeting to finalize the bill, which will determine whether the AI provision advances into the final text.

Previous AI-related state bills have established new floors for safety regulation that were later reflected in other states’ efforts. For instance, an Illinois bill passed in May instituted the concept of annual third-party compliance audits.

The Massachusetts proposal would likely have a similar impact on other states. It goes further than the Illinois bill by requiring independent evaluations of companies’ models themselves, as well as increasing the frequency of audits. On July 24, after the Massachusetts Senate passed its version, Anthropic head of U.S. state and local policy Cesar Fernandez wrote in a post on X that it “raises the bar on AI safety while ensuring that innovation is able to continue.” Anthropic has previously been a vocal supporter of other state-level AI safety bills.

Anthropic is joined by AI safety organizations including Encode and Secure AI Project, which have engaged lawmakers in Massachusetts and recently sent a letter of support to the members of the conference committee, Secure AI Project policy director Scott Wisor told The Information.

Meanwhile, OpenAI has voiced its objection to the Massachusetts language. In previously unreported letters sent on Aug. 18 and 19 to the conference committee and Sen. Barry Finegold, who has spearheaded the AI language, OpenAI policy executives wrote that Massachusetts should consider a requirement similar to the Illinois bill, and added that “fragmented oversight” across the country would represent an “unhappy result.”

“A confusing jumble of rules will unnecessarily delay getting the most advanced cyber models into the hands of state and local officials before bad actors with their own powerful AI break in and exploit vulnerabilities,” wrote Donnie Fowler, the company’s head of U.S. state policy and partnerships, and Ann O’Leary, its global affairs vice president.

OpenAI and Anthropic are not the only AI giants active on the ground. In an interview with The Information, Finegold said he has also met with representatives from Google and Meta Platforms, though he declined to comment on their conversations.

Google is concerned about states setting up additional testing requirements beyond the audits laid out in the Illinois bill, according to a person familiar with the company’s thinking. The lack of standards for testing models creates uncertainty about results, making compliance difficult until those standards are established, the person said. Google has previously withheld products in specific jurisdictions due to regulations—for example, it delayed offering its AI Overviews in many European countries until this summer—raising the possibility that the same scenario could happen with its models across different states.

Though most states, including Massachusetts, have wrapped up their legislative sessions for the year, the conference committee’s work reconciling the economic development bill is expected to pick up after Labor Day, according to Finegold and political strategists working in the state. Finegold said he continues to engage with tech companies on the bill. He traveled to California in August to meet with representatives from OpenAI and Anthropic.

The companies’ interest in pending legislation from Massachusetts reflects the rising importance of states’ roles in advancing AI policy. While OpenAI has previously advocated for federal preemption of state rules to prevent a fragmented patchwork of laws, it has since softened its position, including backing the Illinois law that included third-party audits.

In a July blog post, OpenAI’s chief global affairs officer, Chris Lehane, argued for what he dubbed “reverse federalism,” or states establishing a “shared direction through a common framework.”

Some AI safety advocates saw OpenAI’s rejection of the Massachusetts AI provision as backtracking. “This ongoing fight looks an awful lot like regressing to the mean,” Encode founder Sneha Revanur wrote on X in August. In an interview, Encode policy adviser Ben Snyder, who has worked with Massachusetts lawmakers on the bill, said he was optimistic OpenAI would change its position. “They don’t want to be seen as a company that is opposing these measures, and I think for good reason,” Snyder said.

Whether other companies, including Google and Meta, will lobby more aggressively against the language remains to be seen. Wisor said he expects to see more opposition emerge in the fall as the conference committee works on the language.

Another industry group, the Andreessen Horowitz and Y Combinator–affiliated American Innovators Network, is urging the committee to reject the AI provision. “The Senate bill includes several additional compliance requirements that will create substantial barriers to scaling for small companies while making it more difficult to use open-weight models in the Commonwealth,” Executive Director Jeremy Kudon said in a statement.

But even if the AI provision doesn’t make it into the final version of the economic bill, it could reappear as other states take up their own attempts at regulating the sector. On Aug. 30, the California legislature advanced a bill that would establish a framework for independent evaluation organizations, though it doesn’t set up requirements for audits of AI companies. The bill still needs to be signed by Gov. Gavin Newsom.

Finegold said he has been in conversations with lawmakers from other states about advancing new legislation.

“As the technology evolves, I continue to see each state leapfrogging the other state to make sure that we’re ahead of this,” Finegold said.

Leo Schwartz is a reporter covering the intersection of tech and politics. He can be reached at [email protected] or on Signal at leomschwartz.24