New Delhi: For years, Anthropic occupied an unusual position in the AI industry: a company that publicly acknowledged its own technology could be dangerous while arguing that heavy regulation would do more harm than good. That position has now changed.
In an essay titled ‘Policy on the AI Exponential’, published Thursday, Anthropic CEO Dario Amodei argues that binding federal regulation of AI in the US is no longer optional—and that the window to get it right is narrowing fast.
The essay, running to over 5,000 words, covers everything from drug approval timelines to autonomous weapons to the risk of an AI-enabled seizure of political power. It comes alongside Anthropic releasing a legislative proposal on frontier model testing and a policy framework for job displacement, which the company says it will back financially.
The shift matters because Anthropic is not a peripheral player. Its Claude models are among the most widely used in the world, and Amodei has been one of the more credible voices on AI risk in an industry not known for candour.
Also Read: How India uses AI to empower the next billion users
What changed
The catalyst, according to Amodei, is Anthropic’s own technology.
Claude Mythos Preview—a frontier AI model not available to the public and currently being tested by a small number of organisations under Anthropic’s Project Glasswing—demonstrated risks to financial infrastructure, critical systems and national security serious enough, Amodei writes, to settle a debate that had previously felt theoretical.
“Its broader significance is that it proves beyond doubt that AI models are now tools of global and national strategic consequence,” he writes.
Until recently, Amodei and Anthropic had focused their policy efforts on transparency: pushing for disclosure requirements, safety reporting standards and incident reporting mandates. They backed California’s SB 53, New York’s RAISE Act, and Illinois’s SB 315 (landmark state laws designed to regulate developers of “frontier” AI models) on that basis.
The argument was that good regulation requires good information, and that legislating too early risked creating compliance burdens that missed the actual risks entirely.
That argument, Amodei now says, no longer holds. The risks have arrived. The shape is clear enough. It is time to act.
What he is proposing
The regulatory model Amodei has in mind is the Federal Aviation Administration—a US government agency within the department of transportation with the authority to ground aircraft that do not meet safety standards.
Applied to AI, this would mean mandatory third-party testing for models above a compute threshold across four specific risk categories: cybersecurity, biological weapons, loss of control of AI systems, and automated research that could accelerate any of those risks.
Governments would have the power to block or reverse deployment where a model is found to present unacceptable danger, he suggests.
He is explicit that this is not the endpoint, though. “There may come a time, perhaps relatively soon, when we need to go beyond this,” he writes, when the most powerful AI systems look less like airplanes and more like weaponisable nuclear material.
But he argues against getting ahead of current evidence. Design for today’s dangers, he says, while building the infrastructure to respond faster to whatever comes next.
The jobs question
On labour, Amodei treads carefully. He has previously warned about AI-driven job displacement and has faced criticism for it—both from those who think he is being alarmist and from those who think the CEO of an AI company has no business raising the alarm.
He addresses this directly. “Enduring job displacement is undesirable and dangerous, and we should do everything we can to minimise or prevent it, not to bring it about,” he writes.
He argues that AI differs from previous automation waves because it replicates human cognition broadly rather than automating specific tasks—and that the usual mechanisms of economic recovery, comparative advantage and Jevons paradox may simply be overwhelmed by the pace of change.
His proposals range from the modest—better government data collection on AI’s labour market effects—to the structural: wage insurance for workers who take lower-paying jobs, retention tax incentives for employers and, if displacement proves large and lasting, long-term income support financed through taxes on AI companies and increases in capital gains tax.
The power question
The most striking section of the essay deals with political control.
Amodei argues that AI could enable a seizure of power—by a government, a company or another actor—that existing legal and constitutional frameworks are not equipped to prevent.
Autonomous weapons systems that follow orders without human accountability, surveillance infrastructure that makes mass analysis of private data trivial, AI capabilities that accumulate faster than democratic oversight can track: these are the scenarios he has in mind.
“AI will soon become so capable that I worry it cannot safely be fully entrusted to either governments or companies,” he writes.
His proposals here include banning domestic use of fully autonomous weapons, closing the data broker loophole that currently allows bulk purchase of private data for surveillance and—in a provision that would face significant legal and political resistance—establishing a right for any citizen facing government legal or regulatory action to access AI at least as capable as what the government is using against them.
The geopolitical frame
Amodei’s international proposals are built on a single premise: that AI is not a trade technology but a strategic one, closer in kind to nuclear weapons than to semiconductors or software.
“A nation that possesses powerful AI facing one without it—or even facing one that is behind in AI by three years—could be the equivalent of an army of World War II Marines facing an army of medieval swordsmen,” he writes.
His prescription is a formal coalition of democracies that would share semiconductor supply chains internally, coordinate AI safety standards and restrict access for adversaries.
He endorses tightening US chip export controls on China—controls that he credits as a significant contributor to the American lead in AI—and points to pending legislation MATCH and OVERWATCH as early steps.
The two legislations are aimed at restricting the flow of advanced AI chips and semiconductor manufacturing equipment to foreign adversaries like China.
The context
The essay lands at a moment when the AI industry’s relationship with the government is shifting rapidly. The Trump administration in the US has moved towards a greater federal role in AI oversight, though Amodei argues the steps taken so far do not go far enough. Several of his peers, who spent years resisting regulation, have begun making similar arguments.
Amodei closes with a pointed rejection of the framing that public anxiety about AI is essentially a communication problem.
“People are worried about AI because they correctly perceive that its risks are real,” he writes, “not because AI CEOs have been insufficiently Panglossian”.
Whether governments move fast enough to meet the moment, he suggests, is the question that matters. The technology, in his telling, will not wait.
(Edited by Nida Fatima Siddiqui)
Also Read: Anthropic’s Claude Mythos: The AI model that India cannot access but cannot ignore either