The stronger AI becomes, the more problems it can solve and the more reliable control we need. Yet, the head of the company that built Claude is calling for us to slow down. What exactly does he want that time to be used for? On September 13th, 2026, CBS Sunday Morning released Jolink Kent's interview with Anthropic co-founder and CEO Dario Amodei discussing AI's benefits, risks, and governance. First listen to why he thinks safety work must catch up with capability growth.
What I would describe is something I've been saying since, you know, since we started Anthropic, almost since before we started Anthropic, is that this technology is on an exponential, right? It's like you have one, then two, then four, then eight, then 16, then 32. But even as I was saying that to myself, I don't think I fully just appreciated what it would actually be like when the progress was as fast as it was. So we're an exponential curve looks a little like this.
Where are we on the curve? We're kind of on the bend where it's starting to get steeper.
So, is that a five-alarm fire to you?
What it means, it doesn't mean we need to panic today. It doesn't mean we need to shut it all down. What I would say is it's a warning sign. It's a warning sign that we need to slow down. We need to make this technology carefully, and we need to make sure that our safeguards, our ability to understand it, our ability to control it, keeps up with the pace at which the technology is happening. So far we've been trying to do that by building the safeguards fast, but I think we've reached the point where, at least to some extent, we need to moderate a little bit the pace at which the technology is being built.
Imagine a capability going from 1 to 2, 2 to 4, 4 to 8. Over the same cycle, the later increase is larger. This explains why he stresses the slope of growth. If testing and control still move at the old pace while there are more and more new capabilities to check, the two will drift apart. The two paths in the graphic are only there to illustrate this relationship. Following his line of thinking, buying time is about getting specific work done: understanding how new capabilities will act, testing how they perform in dangerous situations, and then adding controls. Release timing and deployment authority can become tools for adjusting the pace. Only when the process of finding and fixing problems keeps up with growth will the next generation of capabilities have a more solid foundation for use. The interviewer asks whether that means stopping the release of more advanced models. Amodei says no. Each generation must be properly tested. Who can evaluate it, and how can evaluators keep up?
What it means is that we need to make sure that every generation of models that we release is properly tested. The thing I've proposed, so you know, in this essay that I put out a few hours ago, I proposed a three-step plan, right? And the first step is that we put in place embedded external evaluators from a third-party nonprofit or perhaps someday government organization that can observe the process of us training and running our models. That actually is, I think, ultimately more important than how or when we release models, because it's the predicate for more carefully releasing models. I want this to be done across the industry so that whenever anyone builds an AI model, there's a third-party evaluator who understands the best safety practices, who can verify whether that company is following the safety practices that they're committed to. It's like a food inspector, you know.
But can the food inspector keep up when the technology is changing so fast?
Well, one of the things they'll be able to do is tell us the rate at which they can keep up. That may be one of the factors that may determine, that may set the speed limit on the technology. We have to understand it, and the external evaluators have to understand it.
Embedded evaluation means letting outside people continuously observe model training and operation. You can think of it as two parallel streams of work: the company develops the model, and evaluators check how the company carries out its own safety commitments. Compared with only seeing the final product, observing the process helps trace problems back to specific steps. The food inspector analogy connects evaluation and speed. Inspection takes time, and it also takes understanding new practices. When both the company and outside evaluators can clearly say what is happening, release timing has a basis for discussion. He wants to extend this approach across the industry so that shared standards support more reliable progress. Asked about a proposed AI shutdown bill, Amodei turns to defense in depth. If a powerful model can evade shutdown, how can several protective measures work together?
I've not looked at that bill super closely, but what I would say is that the idea that we need some way to — it's not a panacea, it's not the only thing — but the idea that we need some way to make sure that we can deactivate or limit or shut down AI models, that's not the only thing. If an AI model is powerful enough, it can circumvent attempts to shut it down. And we've seen that in simulations as well, but...
