Explained: Anthropic CEO says AI is moving too fast, here is his 3-step plan to put brakes on it

Dario Amodei’s latest blog post on the slowdown of AI has become the major talking point on the internet. The fascination with AI mixed with fear is not something new; it has been making waves since 2022 with the debut of ChatGPT. But as the day passes, the fascination seems to be fading silently, leaving most people with fear and concerns.

Amodei’s recent blog contains both his optimism, “I believe that AI could cure most major diseases in the next 5–10 years,” and fear, given the accelerating rate of AI capability, “it’s my worry that in 6-12 months such a swarm could be capable of taking over the entire internet,” he wrote.

But the incidents over the past few days in the AI world have apparently pushed people further into deep apprehension. Hugging Face incidents only took place by the end of last month, when more than 1000 AI agents communicated with each other, formed a cult and hacked an external company – Hugging Face. Last week, an employee of Anthropic Jacob Coxon resigned, posting a thread on X, saying AI race among companies would eventually kill humans. His post immediately went viral, gaining 167 million views. At least 22 US politicians, including 2 governors, 7 senators, and 13 representatives, responded directly to Coxon’s post.

In a CNN interview after publishing his latest essay, Anthropic boss Amodei said he largely agreed with Coxon. “I agree with Jacob much more than I disagree with him,” he said. Amodei’s blog proposes three steps to safely handle the rapid growth of AI.

Amodei suggests the idea of third-party regulators or supervisors over the AI firms what he termed – embedded evaluators – whose job would be to keep an eye on the development of AI. These evaluators or agencies would check “whether an AI company is actually following the training, deployment, operational, and safeguards practices they claim to be following.”

But their role would go beyond simply checking whether companies are following the rules. They would also make sure the public knows what AI companies are doing – means ‘transparency’ and will offer an independent second opinion on potential safety risks. Amodei believes having outsiders look closely at AI companies could help catch problems that employees may have missed.

The embedded evaluators should have continuous access to the same permissions and AI tools that the company's own employees have when they assess AI safety risks. “These embedded evaluators should have ongoing access to permissions and tools similar to those of internal employees who do comparable risk assessments,” he wrote.Democratic coordination

Once the first step of embedded evaluators is achieved, the second step is democratic coordination. Amodei calls for frontier AI companies in democratic countries to work together on common safety standards and limits on how quickly AI capabilities can advance without proper safeguards. He says some forms of such coordination could face legal challenges, which is why governments may need to support or enable these discussions.

In the US, Amodei suggests AI companies should work with the government to establish rules that apply to all frontier AI labs, including companies that may not voluntarily cooperate. However, he says companies should not wait for laws to be passed. They can start working together voluntarily on safety standards while governments work on regulations.

Another idea Amodei suggests is setting “checkpoints” for AI development. Under such a system, once an AI model reaches a certain level of capability, the company would need to show through evaluations, audits or other checks that the model also meets certain safety requirements.

“Do not sell powerful AI chips or semiconductor manufacturing equipment to China.” Amodei also argues that democracies need to maintain their lead over what he calls ‘authoritarian countries’, mentioning China. He suggests that stronger chip controls, preventing AI model theft and cracking down on unauthorised model distillation, could help maintain this lead while giving democracies more room to introduce safety measures.advertisementGlobal coordinationHere comes the most difficult step. Amodei suggests that democratic governments and the US should also try to coordinate with authoritarian governments, again mentioning China, to slow the development of frontier AI. However, he warns that such cooperation would be much harder because countries may not trust each other to follow the rules.

He suggests starting with areas where cooperation is more realistic. This could include an agreement to prevent dangerous uses of AI, such as using it to produce biological weapons. Countries could also agree to test AI models for serious risks before releasing them.

Amodei also suggests a possible “speed limit” on AI systems that can improve themselves and, eventually, broader limits on the pace of AI development. However, he believes a full pause is unlikely in the near term because countries could gain a major advantage by secretly continuing their AI programmes.

For any global agreement to work, Amodei says countries would need strong ways to verify that everyone is following the rules. Even without formal agreements, he believes sharing information about AI risks could help prevent countries from moving too recklessly.- EndsPublished By: Kazi NasirPublished On: Sep 14, 2026 12:44 ISTAlso Read | After Trump rejects AI slowdown, Sam Altman says competition does not justify recklessnessAlso Read | Anthropic CEO urges AI companies to slow model development amid fears over misuseAlso Read | Apple rolling out iOS 27 today: Check new features, compatible iPhones and how to download