Anthropic advocates for governments to possess the authority to prevent hazardous AI implementations.

      Anthropic has released two policy frameworks advocating for governments to gain the legal power to prohibit dangerous AI models, along with financial protections for workers affected by AI in the labor market. The Advanced AI Framework pertains to safety regulations, while the Economic Policy Framework focuses on issues of displacement, capital distribution, and the social safety net.

      The safety framework is the more assertive of the two, proposing that governments be able to halt or discourage the use of models that present a “significant risk of catastrophic harm.” Civil penalties would be based on a company's global annual revenue and would increase with repeated infractions. This would target models that require over 10²⁵ floating-point operations, created by companies generating over $500 million in AI revenue or investing more than $1 billion in AI research and development.

      This criterion includes only a limited number of companies, such as Anthropic, OpenAI, Google DeepMind, xAI, and possibly Meta. Anthropic’s call for regulation applies directly to its operations, setting it apart from competitors that oppose stricter oversight.

      The framework outlines four categories of catastrophic risk: the creation of biological weapons, large-scale cyber vulnerabilities, loss of control over autonomous systems, and AI that can self-automate its own research and development. Anthropic cited its own experience with Mythos Preview, which identified numerous critical vulnerabilities across major operating systems and browsers, as evidence that these risks are real.

      Frontier developers would need to test their models, publish summary results, undergo independent evaluations, maintain security measures, and regularly release risk assessments. Anthropic argues that transparency alone is no longer adequate, stating, “AI capabilities are going to improve rapidly over the coming months,” and governance must keep up.

      Regarding preemption, Anthropic criticized the White House's initiative to block state-level AI laws, asserting that Congress should not override state laws unless it creates a federal law that meets or exceeds the proposed framework's strength. They advocate for a “surgical” preemption approach, allowing states to regulate issues such as child safety and consumer protection beyond federal safety laws.

      The economic framework, while less detailed, addresses concerns about job displacement raised by recent graduates. It suggests measures such as capital accounts, wage insurance, tax incentives, and an enhanced social safety net. The intention is to ensure that the financial advantages of AI are “broadly shared,” though specific funding and implementation details are reserved for later discussions.

      The timing of this announcement is strategic, as Anthropic is currently engaged in legal battles with the Pentagon over its listing as a supply chain risk, preparing for an initial public offering, and observing Congress's negotiations that could exchange state AI preemption for online safety legislation. By releasing its own framework, Anthropic aims to influence the rapidly evolving regulatory conversation in which it has had little input thus far.