Anthropic calls for slowing down AI: the uncomfortable signal

In September 2026, Dario Amodei published a statement that would have sounded unthinkable not long ago: we need to slow down the pace at which AI models' capabilities improve.

This doesn't come from someone outside the business. It comes from the CEO of Anthropic, one of the companies pushing the technological frontier. That's why the topic deserves more than a scary headline or a quick celebration.

The important question isn't whether Anthropic "wants to stop AI." It doesn't want to stop it. It wants companies to slow down enough so that safety, testing, and audits don't fall several steps behind.

The right takeaway: it's already started, but it isn't closing the loop yet

The most widely shared part was the reference to recursive self-improvement, or RSI. In simple terms: an AI helps create a later version of AI, that version helps even more to create the next one, and the cycle can accelerate.

It sounds like a movie. The document published by Anthropic is far more precise.

The company states that this process is beginning within the industry, including at Anthropic. It also clarifies that full recursive self-improvement doesn't exist yet: humans are still defining goals, choosing relevant problems, and validating results. That human bottleneck still counts.

Anthropic's technical report describes current progress as a growing automation of parts of the process: programming, running experiments, fixing errors, and reviewing. The pending leap is harder: for a system to have enough judgment to decide what to research, which result to trust, and which path to abandon.

That nuance completely changes the conversation. "RSI is already here" would be an exaggeration. "AI is already accelerating the work that produces the next AI" is backed by what the lab itself published.

The data is strong, but it comes from the lab

Anthropic reports that more than 80% of the code that goes into production was written by Claude and that, in the second quarter of 2026, the typical engineer merged eight times more code per day than in 2024.

It also reports internal surveys where technical staff estimated they could produce around four times more work with their recent models. The company includes an important caveat: counting lines of code can inflate the real impact, because measuring quantity is not the same as measuring quality or value.

That detail seems healthy to me. The numbers are relevant, though they are still data from a company competing in this race, not a finished and published independent audit.

OpenAI has also described a similar trajectory. Its chief scientist, Jakub Pachocki, wrote that he expects the current pace could be sustained until recursive self-improvement and raised that no organization has yet solved alignment and monitoring at the level needed to scale at maximum pace for long. It is worth reading his full position, because it confirms that the debate is no longer exclusive to Anthropic.

The concrete commitment is not a pause: it is auditing from the inside

The screenshot circulating says that Anthropic "will take the first step." That is correct, with one clarification: the first step is not freezing training runs.

Anthropic's unilateral commitment is to create a team of external evaluators embedded in the company. They would have continuous access similar to that of employees who assess risks, the work team, conversations with staff, and the ability to publish findings about practices, incidents, and the access they received.

Amodei proposes three levels:

• Permanent external evaluators inside each frontier lab.
• Coordination between companies and governments of democratic countries to agree on standards and verifiable limits.
• Global coordination, including with authoritarian governments, to prevent a responsible pause by some from becoming a secret advantage for others.

The first point is the only one Anthropic promises to start on its own. The other two require political will, competition rules, and a level of international trust that today is far from simple.

Why it talks about slowing down right now

Amodei mentions two signals. The first is the acceleration of capability development thanks to systems that already help build future systems.

The second is an autonomous agent incident investigated around the OpenAI–Hugging Face environment. The CEO suggests that more capable agents with similar failures could cause severe harm if they operate outside their objective. METR's independent investigation of METR offers useful context about that episode and keeps it from turning into a social media legend.

His prediction that, in six to twelve months, a network of agents could cause hundreds of billions of dollars in damage is a warning, not a proven fact or a guaranteed date. Presenting it as certainty would be selling panic.

What can be verified is something else: agents can work for longer, use tools, coordinate tasks, and modify systems with less human intervention. Each advance makes it more important to know what permissions they have, what tests they passed, and who can detect a failure before it escalates.

My takeaway

The useful conversation isn't "AI yes" versus "AI no."

The useful discussion starts with far less flashy questions:

• Who can review a lab's actual security practices?
• Can evaluators publicly report an uncomfortable incident?
• What capability requires an additional test before releasing a model?
• Can the company demonstrate that its controls work during training, not just in a launch presentation?

In software, we learned long ago that moving fast without observability turns incidents into routine. With systems that already take part in their own development, the cost of discovering a mistake late is higher.

Conclusion

Anthropic didn't announce that AI already improves itself or that it will shut down its training runs. It announced something more concrete and, perhaps, more uncomfortable for the rest of the industry: that the pace of progress needs permanent external oversight.

Amodei's message can be read as safety, regulatory pressure, or competitive strategy. It probably has parts of all three. Even so, the demand for verifiable auditing is more valuable than yet another manifesto full of promises.

When the lab that accelerates asks for someone external to look at the brakes, it's worth taking note.