Anthropic is backing a Massachusetts proposal to require large AI developers to conduct independent catastrophic risk assessments every four months. OpenAI and Google are opposing it, arguing for annual audits instead.
TL;DR
- Anthropic supports a Massachusetts bill mandating strict AI safety audits every 4 months.
- OpenAI and Google prefer annual audits, citing compliance burdens and patchwork regulation.
- This is the first time major AI labs have publicly disagreed on specific binding AI safety rules.
What happened
A Massachusetts state Senate proposal, part of a larger economic development bill, would require large AI developers to commission independent evaluators to assess catastrophic risks in their models every four months. Anthropic is backing the proposal, describing it as the most robust AI safety legislation in the nation. OpenAI and Google are opposing it, advocating for annual third-party audits instead.
The disagreement centers around the cadence and independence of these audits. Anthropic wants independent risk assessments every four to six months, while OpenAI favors an approach similar to Illinois' annual third-party audits. Google has not publicly detailed its stance but is aligned with OpenAI's position.
This is the first time major AI labs have publicly taken opposing sides on a specific binding rule, marking a visible break in the industry's previously united front on AI safety regulations.
Why it matters
The four-month cadence is significant because frontier models are revised several times a year. An annual audit could assess a model that has already been superseded twice, making the four-month evaluation more aligned with the release cycle. This could set a new standard for AI safety evaluations if adopted.
OpenAI's objection highlights a real problem: aggressive state-by-state regulation can lead to a patchwork of inconsistent rules, delaying the release of beneficial models. This could impact AI development and deployment, particularly for tools used in cybersecurity defense. However, the absence of a proposed federal alternative weakens their argument.
Anthropic's stance could be seen as commercially convenient, as compliance costs could burden smaller competitors while being absorbable for a company already running internal evaluations. Nevertheless, the rule could simultaneously be good policy and beneficial for Anthropic's market identity centered around safety.
Key facts
- Massachusetts proposal mandates independent catastrophic risk assessments every four months for large AI developers.
- Anthropic supports the proposal, describing it as the most robust AI safety legislation in the nation.
- OpenAI and Google oppose the proposal, advocating for annual third-party audits.
- The disagreement centers around the cadence and independence of AI safety audits.
- This is the first public disagreement among major AI labs on specific binding AI safety rules.
- Frontier models are revised several times a year, making the four-month cadence more aligned with the release cycle.
- OpenAI's objection highlights the potential for a patchwork of inconsistent state regulations.
- Anthropic's stance could convert its safety-focused identity into a competitive moat.
Context
State law has become the primary venue for AI regulation in the United States due to the absence of federal legislation. The outcome in Massachusetts is likely to influence other states, similar to how California's privacy law set a template a decade ago. This development is significant as it marks the first time major AI labs have publicly disagreed on specific binding rules, potentially impacting the industry's ability to present a united front to legislators.
The debate highlights the tension between ensuring AI safety and avoiding overly burdensome regulations that could delay the release of beneficial models. It also underscores the need for a federal framework to preempt conflicting state rules, which could provide clearer guidelines for AI developers and ensure consistent safety standards across the country.
