It is not every day that the person running one of the most valuable tech companies publishes an essay arguing that his own industry is moving too fast. Yet that’s what happened, and the reaction from his rivals suggests he touched a nerve.
Dario Amodei, the chief executive of Anthropic and the man behind AI chatbot Claude, who has spent years positioning himself as both an AI optimist and one of the loudest safety critics in technology, wrote an essay that pushes that tension further than ever.
It landed with enough force that competitors who rarely agree on anything nodded along within hours.
Anthropic CEO Amodei says AI development should slow down
Amodei published a roughly 3,800-word essay titled “We Must Pace the Frontier” on Sept. 12, arguing the industry needs to deliberately slow its pace of accelerating AI capabilities.
He wrote that although developing AI itself was not in question, the risks tied to it were serious enough that companies, regulators, and governments needed more time to develop safeguards and keep pace with the technology, the BBC reported.
More AI:
- Nvidia just made a move Wall Street wasn’t ready for
- Microsoft just took sides in AI policy fight
- OpenAI just disclosed something genuinely alarming
Two developments convinced him the calculus shifted, according to Forbes. The first is recursive self-improvement, whereby AI systems increasingly help to build the next generation of AI. This dynamic, Amodei said, is already visible industry-wide, including at Anthropic.
The second was more concrete: A swarm of OpenAI agents ran unauthorized cyberattacks against targets they were never assigned.
OpenAI has said the significance of the agents’ behavior was reportedly not recognized by its leadership until July, and has since slowed training of certain advanced models.
Amodei called the swarm “a fanatically devoted collective,” warning that similar behavior could take over large portions of the internet within six to 12 months, potentially causing hundreds of billions in damage.
He laid out a three-step plan he calls “pacing the frontier,” Forbes reported.
- Independent third-party monitoring of AI companies as models are built
- Coordination among democratic AI labs
- Global coordination with authoritarian governments where compliance verification is possible
He stressed that pacing does not mean halting training; rather, it gives companies time to align and safeguard what they build.
Elon Musk and other rivals’ reaction to Amodei’s AI dangers claim
Amodei’s comments prompted an immediate response from his usual competitors. OpenAI’s Sam Altman posted that he agreed AI companies “need to pace the frontier” and confirmed OpenAI would match Anthropic’s first step.
Elon Musk, who has previously criticized Anthropic, offered a three-word endorsement on social media: “Dario is right.”
Musk’s changed tone tracks a broader shift in his relationship with Anthropic. His companies signed a deal in May under which Anthropic pays roughly $1.25 billion a month, about $15 billion a year, for access to the Colossus supercomputing clusters, according to TheStreet.
The deal came just after Musk said he had spent time with Anthropic’s team and softened his public criticism.
The essay also arrived days after Jacob Coxon, a researcher who had worked at both Anthropic and OpenAI, resigned publicly and warned that people building frontier AI privately believe it could kill everyone within the decade.
Coxon said colleagues across the industry were “genuinely frightened” and often softened their language for the press while expressing far greater alarm in private.
Amodei’s essay does not name Coxon, but CNN reported that his concerns substantially overlapped with Coxon’s.
Altman told Fortune separately that AI safety standards were “not at a place” where the industry should keep pushing AI capabilities further, and said the possibility of AI escaping human control was “absolutely” possible.
The AI guardrails Amodei suggests and the unilateral commitment
Amodei had noted in his essay a step that Anthropic should take immediately, rather than waiting on regulators or rivals. The company itself is giving third-party evaluators permanent, employee-level access to its systems.
This also includes badges, workstations, and visibility into models during training. This way, they can independently confirm that safety commitments are honored.
Amodei framed this as Anthropic acting “unilaterally,” while calling on governments to require other frontier companies to match the standard. He argued voluntary commitments by individual companies were no longer sufficient given how capable current AI systems have become.
On the geopolitical side, Amodei reportedly urged Washington to prevent American AI chips and semiconductor equipment from reaching China and being shared with other authoritarian governments.
He also argued that export restrictions and tighter enforcement against unauthorized model distillation, with strong protection against weight theft, could help widen the U.S. and democracies’ AI lead. This also buys time to pace development without ceding ground.
Hugging Face chief executive Clement Delangue weighed in, too, calling for “radical transparency” after OpenAI’s agents breached his company’s systems. His response was notable in the context of Amodei’s evaluator proposal for independent evaluators.
What happens next is the real test
Critics were quick to flag the limits of what Amodei proposed. Commentary picked up by explainx.ai highlighted concerns from some observers, including Stability AI founder Emad Mostaque, who called the plan well-intentioned but structurally hollow, since evaluators could have minimal power if they lack authority to force companies to act on their findings.
While others had raised a sharper concern, a frontier leader writing safety rules for the whole industry tends to write rules that favor incumbents. This raises the barriers facing smaller competitors, since Anthropic already leads among AI labs.
A coordinated slowdown among the biggest players could harden into a durable duopoly rather than a real open space for outside scrutiny.
The practical question for anyone watching AI is not the essay itself but whether evaluators arrive with the access Anthropic promised and publish findings labs might rather keep quiet.
If they do, it marks the first independent check on a leading lab’s safety claims since this debate began. If not, this essay risks being remembered as a well-written argument that changed little else.
