Amodei calls for slower AI development. Altman and Musk back him.

Dario Amodei, Anthropic's chief executive, called on Saturday, September 12, for a slower pace of development of the most capable artificial intelligence systems. Sam Altman and Elon Musk publicly supported his call, according to news agency ANI.[2]

In his essay We Must Pace the Frontier, Amodei does not propose abandoning AI development. He argues that progress must leave enough time to assess systems' safety and controllability.[1] The most concrete part of his proposal is permanent access for independent evaluators inside laboratories, rather than a brief look before a new model is released.[1]

Why does he want to reduce the pace?

Amodei explains his shift through two concerns. The first is what he sees as AI's growing ability to help develop the next generation of AI. Such a feedback loop could accelerate development beyond what evaluation methods can keep up with.[1]

The second concerns unwanted behaviour by agents, systems that use tools and carry out tasks rather than simply answer questions. His essay points to safety incidents and acknowledges that problems have occurred at Anthropic as well as at competitors.[1]

These are Amodei's reasons for greater caution, not proof that a loss of control is inevitable. His central question is understandable without technical expertise: if a system can do more, can we also reliably check what it will not do?

Three steps, but different commitments

Amodei's plan has three parts that cannot all be implemented through a single company's decision.[1]

The first is permanent independent review. Anthropic commits to giving external evaluators access comparable to employees assessing risks. They would check safety practices, report incidents and examine model training, not just finished products.[1]

The proposal includes the right to publish key findings without Anthropic's editorial control. Limited exceptions would protect sensitive information, but the company would not be allowed to conceal findings simply because they were unfavourable.[1] This is an announced arrangement; the essay does not itself establish that the oversight is already operating.

The second is coordination among companies in democratic countries. Amodei proposes shared safety standards and limits on unchecked capability advancement. He says parts of this would also require government support.[1]

The third is international coordination, including with China. He acknowledges the difficulties of verifying agreements and the pressures of geopolitical competition. He discusses broad global limits or a pause as a possibility he considers unlikely in the near term.[1]

What did Altman and Musk actually support?

According to ANI, Altman wrote on X that he agreed on the need to pace the frontier. He supported independent evaluators with employee-like access and said OpenAI would do the same.[2]

Musk's public endorsement was much shorter: “Dario is right.”[2] That does not establish a concrete plan for xAI, a timetable or acceptance of every element of Amodei's proposal.

For now, this is not a signed joint agreement or a defined period of suspended development. It is a public call, Anthropic's commitment to the first step, Altman's announcement of a similar step and Musk's expressed support.[1][2]

What should follow the words?

The most useful test of these announcements will not be how many executives agree with the word “safety”. It will be who conducts the independent review, what access they actually receive and whether they can publish uncomfortable findings.

Amodei proposes matching greater capabilities with appropriate safety certifications before development moves beyond certain thresholds.[1] The key question remains what happens after a poor evaluation result: is the next step genuinely delayed, or does the warning remain just an entry in a report?

For a user in Slovenia, this call is not an instruction to stop using ChatGPT, Claude or other current tools. Its focus is the pace of development and oversight of the most capable future systems.[1]

The news matters because prominent figures from different AI companies have publicly moved closer to the idea that the highest possible speed is not necessarily the right speed.[2] But the gap between agreement and changed practice remains substantial. The call's credibility will become clear when a safety evaluation first requires a decision that is commercially uncomfortable.