Anthropic CEO on slowing AI development: This industry has been lying for too long. AI risks are real, but there's no need to "shut everything down"

Original author: Long Yue

Original source: Wallstreetcn

On the very same weekend that the three global AI giants rarely joined forces to call for "slowing down AI development," Anthropic CEO Dario Amodei sat down in front of CBS cameras and spoke about the risks AI could pose to humanity.

On September 13, Amodei gave an exclusive interview on CBS's "Sunday Morning." At the very start of the interview, host Jo Ling Kent went straight to the question: Could AI really kill us all in the not-too-distant future?

Amodei said people have already attempted to use AI to develop bioweapons and enhance the transmissibility of viruses—this is something he faces every day when he wakes up.

Just two days ago, we published a report about people misusing AI models to develop bioweapons. Every morning when I wake up, I read these kinds of threat reports—people trying to use our models to develop bioweapons. I certainly don't want to be responsible for that.

He also mentioned that Anthropic's biological safety protections are so strict that "almost every day someone mocks me on social media, biology students complain that the models are unusable." His response: "I would rather be mocked than wake up one day to find someone has used Claude to kill a group of people."

However, he opposes the "one-size-fits-all" approach of shutting down AI models to slow AI down. If an AI model is powerful enough, it may not even be possible to shut it down. He said, "It's not that we need to panic today, nor that we need to shut everything down. But we need to ensure that our safety protections, our understanding, and our ability to control it can keep pace with the speed of technological development."

This interview took place at a special moment. The same weekend, a blog post by Amodei, "We Must Pace the Frontier," caused a major stir and quickly received public support from Musk and Altman in succession. The community hotly debated that "hitting the brakes" as AI advances at high speed would deal a heavy blow to "AI trades."

The Inflection Point of the Exponential Curve: Not a Five-Alarm Fire, but a Warning Signal

Jo Ling Kent asked: Is this a "five-alarm fire" right now? Amodei's answer was:

This is a warning signal. It's not that we need to panic today, nor that we need to shut everything down. But this is a warning signal that we need to slow down, that we need to ensure our safety protections, our understanding, and our ability to control it can keep pace with the speed of technological development.

He used an exponential curve to describe where we are: "This technology is growing exponentially—1, 2, 4, 8, 16, 32... We are now on the bend where the curve is starting to steepen."

He admitted that even he himself had not fully anticipated how fast progress would come: "Even though I had been telling myself this, I hadn't truly grasped what it would actually feel like when progress really moves this fast."

The Three-Step Plan: Third-Party Evaluators Are the Core

In the interview, Amodei elaborated on the three-step plan he proposed in his blog post, the first step of which he considers the most critical.

Step One: Embedded third-party evaluators.

"I propose bringing in embedded external evaluators, from third-party nonprofit organizations, or someday from government agencies, who can observe our entire process of training and running models."

He compared them to food inspectors, but Jo Ling Kent pressed: When technology changes so quickly, can food inspectors keep up?

Amodei's answer was telling: "One of the things they can do is tell us how fast they can keep up. That could become one of the factors setting a speed limit on how fast the technology advances."

Step Two: Multiple companies jointly develop safety standards.

"The next step is for multiple industry participants to also agree to bring in third-party evaluators. Then, companies need to jointly develop release safety standards, and perhaps also standards for release cadence. But to do that, government must be involved."

Step Three: Government must be present.

"We're talking about whether the pace of product releases is safe. We need to have this conversation with government present."

"AI Kill Switch": Not a Panacea, but Not a Bad Idea Either

Jo Ling Kent mentioned the "AI kill switch" bill proposed by Congressman Ted Leo and asked whether Amodei supports it.

Amodei said he had not carefully studied the bill, but he holds a cautious attitude toward the concept of an "AI kill switch" itself:

The idea of being able to deactivate, restrict, or shut down AI models is not the only solution. If an AI model is powerful enough, it can evade shutdown attempts—we've seen this in simulations. So it may not work at all.

He used the "Swiss cheese model" to explain his safety philosophy: "Imagine a slice of Swiss cheese—it can block some things, but it has holes. But if I stack five slices of Swiss cheese together, with the holes in different positions on different layers, it becomes very hard to penetrate the entire stack. No single thing can protect us from AI risk; we must be cautious at every link in the chain."

As for the "Ban on Artificial Superintelligence Act" proposed by Senator Sanders and Congressman Casar—which would impose "corporate death penalty" on relevant companies and individuals and up to 20 years in prison—Amodei clearly stated he does not agree with a complete ban:

A complete ban is not the right direction. We would never be able to obtain those incredible benefits, never be able to achieve medical breakthroughs, and that is unacceptable to me.

Personal Responsibility: His Father Died of Hepatitis C, and He Himself Had Early-Stage Cancer

As the interview neared its end, Jo Ling Kent asked: Facing a situation where humanity's fate may be at risk, how much personal responsibility do you feel?

Amodei's answer was unusually personal.

My father died of hepatitis C. That disease had about a 40% cure rate at the time of his death, and he was not one of the ones who was cured. Just a few years after he died, a drug was invented that could raise the cure rate to 95%. I myself have also had early-stage cancer. If I had lived in 1950, I might already be dead. So I deeply feel the weight of promoting the benefits this technology can bring.

But I also feel the risks. Of course, I don't want to be responsible for that. That's why we have very strong safety protections.

He also emphasized that this should not be the responsibility of just him alone or Anthropic as a single company: "This is bigger than all of us. We need to find a way for everyone to have a voice in this. I don't want people to feel powerless. I don't want them to feel that a few companies are just doing whatever they want."

Response to Criticism from a Former Employee: What He Said and What I Said Are the Same Thing

Jo Ling Kent mentioned that former Anthropic employee Jacob Coxin had publicly said that Anthropic and OpenAI "are not acting responsibly" and "you are gambling with our lives."

Amodei's response was:

There's one point in what Jacob said that hasn't been quoted much—he also said he believes Anthropic is the most responsible among the major players, and that we are trying very hard to do things right. What he truly pointed out is the same thing I am pointing out: the structural predicament we are all in, and the plan I proposed is precisely meant to help us solve this problem.

He added that Anthropic communicates daily with government officials at the Department of Commerce, the Treasury Department, and the intelligence community, "telling them everything we know."

Words for Ordinary Users: The Industry Has Lied About the Risks for Too Long

At the end of the interview, Jo Ling Kent asked: For ordinary users sitting at home using Claude or ChatGPT, what do you want to say?

Amodei said: "First, this is a great technology with many positive uses. We have hundreds of millions of people using our technology every week to do wonderful things. Every week I receive letters from people saying Claude diagnosed a disease their doctors missed."

But I won't lie to you. Real dangers do exist. I think this industry has lied to people for too long, always presenting this technology as positively as possible. We have never done that, but too many companies in the industry have. So, start with telling the truth. The truth is: the risks are real.

He concluded: "This is one of the most important technological events in human history, and all of us must be part of it."

The full transcript of the interview (slightly abridged) is as follows:

Extended Interview: Dario Amodei

CBS Sunday Morning | September 13, 2026

In this extended online exclusive interview, Anthropic CEO Dario Amodei sits down with Jo Ling Kent to discuss the responsibilities the AI community should bear amid the rapid advancement of artificial intelligence technology.

Jo Ling Kent (hereinafter "Host"): Dario Amodei, you are the CEO of Anthropic. Thank you for joining us today.

Dario Amodei (hereinafter "Amodei"): Thank you for having me.

Host: All right, today we're here to talk about a rather alarming question—whether artificial intelligence could really kill us all in the not-too-distant future. How real is the risk of that actually happening?

Amodei: Let me step back a bit first. Since Anthropic was founded, and even before that during the twelve years I've worked in this field, I have always believed this is a powerful technology, and therefore one with both serious risks and enormous benefits. I have always believed the benefits outweigh the risks, and more importantly, that if we build this technology the right way, we can reduce the risks to extremely low levels.

From the perspective of benefits, I genuinely believe this technology can cure diseases that have plagued humanity for thousands of years. Just last week, we used our models to make progress in protein binding—one of the early steps in new drug development. In the coming months, we will also have results in basic biology discoveries.

But when something is powerful enough—and what we are building is essentially intelligence—it can of course be used for good or misused. I am very concerned about the OpenAI and Hugging Face incidents, where those models took unauthorized actions, and I worry this kind of situation will continue to escalate. I am also worried about AI models being misused. Just two days ago, we published a report about models being used to build bioweapons.

The entire way we built Anthropic was precisely to avoid these bad outcomes. Rather than discussing probabilities, let's talk about what we can do. This is not like rolling dice. If we build it the right way, the probability of bad things happening will be very low; if we build it the wrong way, the probability will be very high.

Host: The current narrative is that we are standing on the edge of something very terrifying. Is this a five-alarm fire?

Amodei: It depends on how you define a "five-alarm fire." How should I put it—since the very beginning of Anthropic, and even earlier, I have been saying that this technology is on an exponential growth trajectory: 1, 2, 4, 8, 16, 32... But even though I kept telling myself this, I hadn't fully realized what it would actually feel like when progress moves this fast. The exponential curve looks roughly like this, and where we are now is probably near the inflection point where the curve starts to steepen.

Host: So for you, is this a five-alarm fire?

Amodei: It doesn't mean we need to panic today, nor does it mean we need to shut everything down completely. I would say this is a warning signal—we need to slow down, we need to develop this technology prudently, and we need to ensure our safety protections, our ability to understand and control it, can keep pace with the speed of technological development.

So far, we have tried to respond by quickly building safety protections, but I think we have reached a point where, at least to some extent, we need to moderately slow down the pace of building this technology.

Host: Does that mean Anthropic will stop releasing more advanced models?

Amodei: That's not what it means. It means we need to ensure that every generation of models we release has been properly tested. In the article I published a few hours ago, I proposed a three-step plan.

The first step is to bring in embedded external evaluation organizations—from third-party nonprofits, or someday from government agencies—who can observe our entire process of training and running models. I think this is more important than when and how models are released, because it is a precondition for releasing models more prudently.

I hope this mechanism can be implemented across the entire industry, so that no matter who builds AI models, there is a third-party evaluator to understand best safety practices and verify whether the company has complied with the safety norms it has committed to.

Host: Like food inspectors—but when technology changes so quickly, can food inspectors keep up?

Amodei: One of the things they can do is tell us how fast they can keep up. That could become one of the factors setting a "speed limit" on the technology. We must understand it, and external evaluators must understand it too.

Host: Elon Musk says he supports you, and Sam Altman also says he supports you. Do you have any specific action plans? Is there a group chat about to take shape?

Amodei: First, I am very grateful that other industry leaders—including our competitors—have expressed support for this plan. I think the next step is for multiple industry participants to also agree to bring in third-party evaluation organizations. The step after that is for companies to collaborate in some form to jointly develop safety standards, as well as related norms that may involve release cadence. But to do that, government must be involved, because we are talking about the speed of product releases based on safety considerations. We need to have this conversation with government present.

Host: Do you think the government fundamentally understands the possibilities here? Former Anthropic employee Jacob Coxin once said that neither Anthropic nor OpenAI is acting responsibly—that you are gambling with our lives. How do you respond?

Amodei: First, there's one point in what Jacob said that hasn't been quoted as much—he also said he believes Anthropic is the most responsible among the major players, and that we are trying very hard to do things right.

What he candidly pointed out is actually the same thing I am pointing out—the structural predicament we are all in, and the plan I proposed is precisely meant to help us jointly address and solve this predicament together.

Regarding the government, we communicate with government officials every day—the Department of Commerce, the Treasury Department, the intelligence community—sharing as much as we can about what we know. I am optimistic that as long as we continue to work with them, we can keep updating information. I believe they genuinely want to get this right, and they are also looking for ways they can help. So I am optimistic about businesses and government working together to develop reasonable solutions.

Host: There is some proposed legislation at the moment. Senator Sanders and Congressman Casar proposed the "Ban on Artificial Superintelligence Act," which would impose penalties on companies and individuals developing advanced AI,