Anthropic’s US AI regulation proposal criticized for limiting public oversight

The company behind Claude has proposed a draft regulation for advanced artificial intelligence systems in the United States. The proposal includes external assessments, risk disclosure, and a judicial process to suspend models in cases of imminent danger. However, the Americans for Responsible Innovation (ARI) organization argues that the proposal gives companies the power to set their own safety standards and fails to grant the government sufficient authority to oversee these systems.

Criticism of the proposal

ARI contends that the government should establish mandatory criteria and have the capacity to test and inspect AI platforms. According to the organization, the draft would allow developers to create and adjust their own rules, thereby weakening public protection. Additionally, the proposal would replace some state requirements with federal standards, potentially undermining existing state-level safeguards.

Mechanism to halt AI models

Under the proposal, the Department of Commerce would need to petition a federal court in Washington to halt the development, deployment, or internal use of an AI model. The petition must demonstrate an imminent risk and explain why less drastic measures would be insufficient. The thresholds for intervention are set at the risk of death or serious injury to more than 50 people or property damage exceeding $1 billion. Scenarios include biological weapons, cyberattacks, loss of AI control, and automated research that amplifies these risks.

Dispute over state laws

Federal preemption allows national laws to override state requirements on issues such as transparency, external assessments, and incident reporting. ARI warns that this could weaken existing state protections and argues that federal rules should be more stringent and adaptable to technological advancements. Anthropic’s own Advanced AI Framework, published in June, already conditioned the replacement of state regulations on the existence of a robust federal regime, while preserving state authority in areas not explicitly covered.

Anthropic’s response

Anthony Cimino, the company’s federal affairs director, announced on Sunday, 18, that the language regarding state laws would be revised. In a post on X, he stated that the goal is to preserve a role for states in areas such as evaluator standards and the ability of labs to shut down models. Cimino also advocated for balancing the need for swift government action in emergencies with protections against the abuse of power to halt AI systems. The company is discussing stricter incident reporting requirements and ways to prevent evaluator selection based on result convenience.

Context and other proposals

The debate over who should oversee the AI sector was brought to public attention by leaders from Anthropic, OpenAI, and Google DeepMind. The FRONTIER Act, introduced in July by Representatives Jay Obernolte and Lori Trahan, proposes licensing evaluators and allows the Secretary of Commerce to issue emergency orders to restrict AI models. For users, key concerns include risk disclosure and the removal of dangerous products from the market. The draft has no legal force and does not create obligations in Brazil.