Anthropic CEO Dario Amodei said governments should have the power to block dangerous artificial intelligence systems during an interview this week [1, 2].
The request highlights a growing tension between the rapid pace of AI development and the ability of regulators to prevent catastrophic failures. As AI capabilities expand, industry leaders are increasingly signaling that voluntary safety commitments may not be sufficient to protect the public [2, 6].
Speaking on ABC News Live Prime, Amodei said advanced AI could outpace current societal guardrails [1, 6]. He said government intervention is necessary to ensure that dangerous systems are not deployed before adequate safety measures are in place [2, 6].
Amodei's position emphasizes a shift toward formal regulation rather than industry self-policing. The call for government authority to halt the release of specific models suggests that the risks associated with frontier AI are now viewed as a matter of national or global security [2, 4].
This push for regulation comes as AI companies face scrutiny over job displacement and the societal impact of automation [7]. Amodei said the speed of these technological shifts creates a precarious environment where the window for implementing safety protocols is shrinking [6].
By advocating for the power to block systems, Amodei is positioning the U.S. government as the final arbiter of what is too dangerous for public release [2]. This approach seeks to create a legal framework where safety is a prerequisite for deployment, rather than an afterthought [4, 6].
“Governments should have the power to block dangerous AI systems.”
Amodei's call for government 'kill-switch' authority represents a strategic pivot in the AI industry. By inviting regulation, Anthropic and other safety-focused labs may be seeking to establish high entry barriers that prioritize safety and stability over the 'move fast and break things' ethos of earlier tech cycles. This move signals that the industry recognizes a threshold where the potential for systemic harm outweighs the benefits of rapid, unregulated iteration.
