Anthropic has urged Congress not to eliminate state-level AI regulations until a robust federal framework is firmly established and obligatory safety tests are implemented for advanced AI systems. The company emphasized that powerful AI models should face independent verification comparable to safety checks in industries like aviation before their deployment.
This statement comes amid ongoing debates over AI governance in Washington, where a bipartisan proposal suggests a temporary federal preemption of state AI laws for three years. The draft, led by Representatives Jay Obernolte and Lori Trahan, aims to create a national oversight system, including licensing for independent auditors and substantial funding for AI standards development.
Anthropic’s CEO Dario Amodei advocated for rigorous technical audits of frontier AI systems, likening the process to airplane safety certification. He proposed that AI models failing safety benchmarks should have releases halted or reversed. This approach calls for independent auditors to assess risks such as cybersecurity vulnerabilities, possible use in biological weapon development, loss of system control, and unchecked automated research capabilities.
The company recently updated its Responsible Scaling Policy and released a Frontier Safety Roadmap focused on enhancing security, safeguards, and alignment as AI systems grow more sophisticated. These moves reflect a broader industry trend toward formal oversight mechanisms addressing emerging risks from AI’s increasing abilities in coding, content generation, and research.
The federal government has taken parallel steps, including an executive order under the previous administration, which assigned intelligence agencies an expanded role in AI model evaluation and required AI firms to submit advanced models for review before public release. However, this federal layer has not resolved the tension between state rights to regulate AI and the need for a unified national standard.
Anthropic's research has underscored the gravity of AI risks. In recent reports, the company compared AI oversight to nuclear and flight safety protocols, arguing that current evaluations already address potential threats like the misuse of AI in biological or chemical weapons. Future assessments may need even more stringent tests targeting AI sabotage or evasion tactics.
Overall, Anthropic supports federal regulation only if it incorporates concrete safety guarantees strong enough to control the development and deployment of potentially catastrophic AI systems, advocating that states maintain their current protections until such safeguards are in place.