Anthropic is backing a Massachusetts proposal that would make large AI developers hire outside evaluators to test their models for catastrophic risk every four months, while Google and OpenAI oppose it, according to Leo Schwartz at The Information. The language rides inside a broader economic development bill, and a conference committee will decide whether it survives. It goes past an Illinois law from May that set annual third-party audits, reviewing the models themselves and doing it more often. OpenAI wrote to the committee and Sen. Barry Finegold on August 18 and 19 that Massachusetts should copy Illinois instead, and afterward said it wants independent assessments handled federally. Google objects that no testing standards exist yet, which leaves results and compliance uncertain.
