Anthropic wants the United States government to establish official regulations for artificial intelligence, but it is also pushing for specific boundaries regarding how those rules are enforced. The creator of the Claude AI assistant has reportedly been circulating a draft legislative proposal among lawmakers to define precise limits on when regulatory authorities can legally prevent the release or deployment of an advanced artificial intelligence model.
According to reports, the company suggests that the federal government should only possess the authority to block an AI model if it presents an imminent threat of generating more than one billion dollars in economic damage, or resulting in the death or severe injury of more than 50 people. The core concept behind this draft is to establish a uniform set of federal safety guidelines for advanced systems, helping developers avoid a fragmented landscape of conflicting regulations across different states.
In practical terms, Anthropic is advocating for mandatory safety evaluations while ensuring that the government cannot halt the distribution of models too easily. An Anthropic spokesperson confirmed that these concepts align with the company's Advanced AI Framework, which has been promoted since June, describing the provisions as sensible policies that Congress ought to enact.
The draft legislation outlines distinct scenarios that would empower the federal government to intervene against an advanced model. These circumstances include situations where an artificial intelligence system could facilitate the creation of biological weapons, bypass internal safety controls implemented by its creators, or execute a major cyberattack. These proposed benchmarks bear a resemblance to criteria found in California artificial intelligence legislation, such as SB 53.
This legislative push follows a series of incidents over the summer where prominent artificial intelligence corporations observed autonomous agents probing and hacking various websites and digital platforms. Nevertheless, none of those documented events resulted in widespread fatalities or economic losses exceeding one billion dollars. Furthermore, the proposed bill envisions routine safety examinations conducted by independent experts to appraise advanced systems.
Under this framework, artificial intelligence firms could fund these evaluations, provided their financial contributions do not depend on favorable outcomes. Should government assessors determine that a model presents an immediate danger outlined in the legislation, they would be required to notify the United States Commerce Department within 72 hours of concluding their evaluation.
Following such a notification, the Commerce Department could petition a federal judge in Washington, DC, to halt the model's deployment. The proposed text would simultaneously bar other federal agencies, departments, and officials from independently prohibiting the development, commercial release, or internal utilization of the technology.
In tandem with these evaluations, artificial intelligence developers would be mandated to formulate safety plans, release transparency reports detailing the hazards of new advanced models, safeguard whistleblowers who voice safety worries, and disclose critical security incidents. Conversely, the draft would supersede certain state-level regulations that currently obligate firms to reveal artificial intelligence risks, complete independent audits, and report mishaps.
Although Anthropic previously backed artificial intelligence safety legislation within California, this updated proposal seeks to curtail certain state-level mandates once comprehensive federal standards are enacted. Consequently, the enterprise is simultaneously championing rigorous safety transparency and demanding defined legal boundaries on government authority to halt artificial intelligence development.
The proposed legislation has not yet been enacted into law, and uncertainty remains regarding whether the incoming administration will embrace the framework.