Dario Amodei, CEO of Anthropic, is not calling for AI development to stop. He argues, however, that the industry has reached a point where model capabilities are advancing faster than companies and institutions can understand and control them. “What we need to do is not stop. We need to slow down,” he says in the CBS interview. 10:03 — watch this moment in the interview

The call to slow down is not, by itself, new. Amodei had already laid out that argument at length in an essay published before the interview. In his conversation with CBS, however, he goes further, discussing who should ultimately control this technology, how much power should remain in the hands of the companies building it, and what happens if models become capable enough that conventional safety mechanisms are no longer sufficient.

1. Dario Amodei does not think technology this powerful should be controlled exclusively by private companies

Asked who should ultimately control artificial intelligence, Amodei says he has always found it “very strange” that a technology with such far-reaching consequences is being built by private companies, and openly admits that the situation makes him uncomfortable. When asked whether he would be willing to hand over control to government, he replies: “to the right combination of governments.”

He is not talking about handing Anthropic over to a single state. On the contrary, he points out that a government can abuse a technology this powerful just as a company can. What he describes instead is some form of oversight or “joint governance” involving several democratic governments. That takes the discussion beyond the familiar idea of regulation: one of the people currently in control of frontier AI development is saying that he is not comfortable with a system in which a small number of private companies make the crucial decisions on their own.

What remains unclear is how much control Anthropic would actually be prepared to give up. “Joint governance” could mean anything from independent inspectors and binding rules to authorities having the power to block the release of a model. Amodei does not say where that line should be drawn.

2. Amodei accuses the AI industry of lying to the public about the risks of the technology

Near the end of the interview, Amodei makes one of its strongest accusations: “For too long, the industry lied to people about the fact that this technology had risk.” He says too many companies tried to present artificial intelligence in the most positive possible light instead of speaking openly about the dangers. 22:23 — watch the statement

The statement directly implicates other major AI developers, but the public record makes the accusation more complicated. OpenAI, Google DeepMind and Anthropic were already warning in 2023 that advanced AI systems could create risks comparable to pandemics or nuclear war. OpenAI and DeepMind later published their own frameworks for evaluating catastrophic risks.

The problem may therefore lie less in a complete absence of warnings than in the gap between safety messaging and commercial messaging. The same companies that published documents about extreme risks were also presenting their models as products ready to be integrated rapidly across the economy.

Amodei does not specify whom he is accusing of lying or which statements he has in mind. Without those details, the accusation remains powerful but extremely broad — and Anthropic is, conveniently, excluded by Amodei himself from the group of companies he says behaved this way.

3. Anthropic wants outside inspectors to see AI systems before they become products

One of Amodei’s concrete proposals is to introduce independent evaluators who would not simply receive the finished model for a formal audit. He compares them to food inspectors: if the stakes are high enough, the company should not be the only party deciding whether its own product is safe.

Anthropic proposes giving these evaluators ongoing access, similar in some respects to that of employees, to the process through which models are built and tested. They could check whether safety commitments are being followed, investigate incidents and examine parts of the training pipeline rather than seeing only the model after training is complete.

That leads to a less conventional idea: if external evaluators can no longer understand the systems quickly enough, then the pace of evaluation could become a limit on the pace of development. In other words, capability progress should not move too far ahead of the ability of outside parties to assess the risks.

That would change the role of safety work. Instead of being a check performed after research has already produced the next model, it would become a constraint on how quickly the research itself is allowed to advance.

Everything, however, depends on the independence of those evaluators. Who chooses them, who pays them, what information can they access and what are they allowed to publish? An inspector with a badge and a laptop inside Anthropic may sound much more powerful than an outside auditor, but that power ultimately depends on the rights the inspector is actually given.

4. Amodei wants to borrow from nuclear arms control for the US-China relationship

China complicates any plan for unilateral slowdown. If American labs reduce their pace while Chinese labs continue accelerating, the strategic advantage could shift. Amodei acknowledges the problem and proposes two levels of cooperation.

The first is relatively limited: the United States and China could agree not to use AI to develop biological weapons and not to release systems that would enable bioterrorists to do the same. In effect, Amodei is proposing that the logic of the Biological Weapons Convention be extended to certain uses of AI.

The second option is much more ambitious: some form of “speed limit” on AI progress agreed by the major powers. Amodei does not pretend such an agreement would be easy to achieve. On the contrary, he says the central problem would be verification. If one side could secretly accelerate, the strategic advantage would be too large for mutual trust alone to be enough.

At that point, the discussion starts to resemble Cold War arms-control treaties more than the regulation of a software industry. The difference is that AI is harder to inspect than a nuclear arsenal. Chips, data centers, algorithms and the use of AI to develop the next generation of AI cannot be counted as straightforwardly as warheads.

Amodei admits that he does not know whether such a mechanism could work. What matters is that he already treats the pace of AI progress as something that, under certain conditions, could become an object of strategic negotiation between states.

5. A “kill switch” may not be enough if AI becomes capable of avoiding shutdown

Amodei does not reject the idea of a mechanism that could shut down a dangerous model. Such a capability could be useful, he says, but it would not be enough. A highly capable system might try to evade the mechanism intended to deactivate it, and researchers have already observed behaviors of this kind in simulated scenarios.

That is why he prefers a defense built in multiple layers. He uses the Swiss cheese model: every safeguard has its own holes, but several independent barriers can reduce the chance that all of them fail at the same time.

The existing experiments do not show that current models can resist a real-world shutdown, nor that they consistently seek to keep themselves running. They do show that, in certain simulations, behaviors can emerge in which a system tries to preserve its autonomy or continuity.

Anthropic is therefore designing safety mechanisms for a situation in which it can no longer simply assume that the model will cooperate with the procedure intended to control it. The fact that a frontier AI company is already working with that possibility says a great deal about the kind of systems it is preparing for.

6. Amodei suggests Anthropic is approaching scientific results in biology, not just AI-assisted research

Amid the warnings about risk, Amodei also offers a glimpse of what Anthropic is preparing. The company is already using its models in protein binding, which is relevant to early stages of drug development, and the CEO says that in the coming months Anthropic will have more to say about using Claude for “basic scientific discoveries in biology.”

The distinction matters. A model can help a researcher search the literature, write code or generate hypotheses without actually contributing to a genuinely new scientific discovery. Amodei’s wording suggests something more. Anthropic has already presented results related to protein design, but what he describes here appears to move toward biological research in which Claude plays a more direct role.

If those results are experimentally confirmed and show that the model made an essential contribution to a discovery that would otherwise have been harder or slower to achieve, it would provide a concrete example of a promise the AI industry has been making for years: not merely automating existing work, but accelerating the production of new knowledge.

This also helps explain why Amodei rejects the idea of stopping AI development altogether. His father died from hepatitis C before much more effective treatments became available, while Amodei himself survived cancer that was detected at an early stage. For him, the medical benefits of technological progress are not an abstraction.

At the same time, the very same capabilities that can accelerate biology and drug development can have dangerous uses. When explaining why Anthropic maintains restrictions that sometimes frustrate even biology students, Amodei says he would rather be mocked for overly strict safeguards than wake up one day and learn that someone had used Claude to kill people. 20:26 — watch this moment in the interview

The same tension runs through the entire interview. Amodei wants AI development to slow down, but not stop. He wants more public control, but does not want a single government to control the technology. He believes AI can accelerate medical discoveries, while building barriers because the same capabilities can be used in the opposite direction.

The result is not a simple plan, but a recognition that if these systems continue becoming more powerful, companies, governments and technical safeguards will not be sufficient on their own.

At the end, Amodei also says who should not make these decisions alone: the people currently running the AI labs. How the technology is developed and used, he says, should not be determined by “a few people.” The public and its elected representatives need to have a role in that decision. 23:27 — watch the end of the interview