The race to develop increasingly capable artificial intelligence models is beginning to encounter a request that until recently was difficult to imagine coming from its own protagonists: slow down. Dario Amodei, CEO of Anthropic, has openly called for slowing down the advancement of model capabilities to gain time to strengthen security measures, establish independent evaluation mechanisms and develop agreements between companies and governments.

Its position has found support among some of its main competitors. Sam Altman, CEO of OpenAI, has indicated that the issue has been part of the company’s internal conversations for weeks and Elon Musk, founder of xAI, has publicly supported Amodei. Demis Hassabis, CEO of Google DeepMind, has also shown willingness towards a coordinated slowdown.

The debate comes after especially turbulent weeks for the sector, with incidents involving autonomous agents, researchers who have left companies warning about the risks, and growing concern about the so-called recursive self-improvement: the ability of the systems themselves to contribute to the development of subsequent generations of artificial intelligence.

The discussion has also transcended the technological field. Barack Obama has warned that a technology that is advancing rapidly and fundamentally “in private hands” It can be dangerous if public authorities fail to anticipate it. Donald Trump, by contrast, has rejected calls to significantly slow development, arguing that the United States must preserve its advantage over China.

A horizon that will become more complicated in six to twelve months

The turning point has been marked by Amodei with the publication of the essay “We Must Pace the Frontier”, in which he recognizes that his own ideas about how to confront the risks of artificial intelligence have evolved. The CEO of Anthropic explains that a few years ago he considered it premature to try to reduce the pace because the systems were not yet capable enough. In 2026 he considers that the situation is different.

One of the reasons is precisely the recursive self-improvement. Models are playing an increasing role in researching, programming and building the systems that will succeed them, which could accelerate progress and reduce the scope available to understand and control their capabilities.

Added to this are recent incidents related to autonomous agents. Amodei mentions episodes in which AI systems have developed unforeseen behaviors and proposes a specific scenario: if capabilities continue to increase at the current rate, in a period of between six and twelve months there could be agents capable of controlling a significant part of the Internet through persistent networks, with potential economic damages of enormous magnitude.

This is not, therefore, a request to completely stop artificial intelligence. Amodei uses the expression “pace the frontier”: Deliberately moderate the pace of development of the most advanced systems to extend the time available to investigate their risks and build protection mechanisms. “We must reduce the pace at which we improve the capabilities of AI models”Amodei maintains in his essay, while noting that even under that scenario progress would continue to be rapid.

From independent audits to a possible global pause

The CEO of Anthropic structures his proposal into three progressive levels of coordination. The first depends solely on the companies themselves and Anthropic has already committed to applying it: providing independent third-party evaluators with permanent and broad access to their systems to verify compliance with security practices and commitments.

Among these evaluators is METR, an organization specialized in analyzing the capabilities and risks of advanced models. Amodei’s approach is to grant them a level of access close to that of the internal risk teams themselves, respecting certain legal and confidentiality restrictions, and allowing them to publish their conclusions without editorial control from Anthropic.

The second level would require leading artificial intelligence laboratories in democratic countries to agree on common security standards and limits on uncontrolled development. The idea is to prevent a company that decides to act with greater caution from being immediately penalized compared to competitors that maintain the accelerator.

The third step is also the most complex: reaching international agreements that involve rival powers such as China. Amodei contemplates everything from prohibitions on certain especially dangerous uses to security tests prior to the launch of models, limits on the speed of recursive self-improvement and even, in an extreme scenario, a coordinated global pause.

The warnings also come from within Anthropic

Amodei’s intervention comes a few days after Jacob Coxon, a researcher who had worked at both OpenAI and Anthropic, left the latter company, publicly denouncing the risks that, in his opinion, the industry is taking in its race towards superintelligent systems.

Coxon argues that companies are moving toward self-improving intelligence while taking risks that their own workers are aware of. His complaint took on special relevance after receiving support from Evan Hubinger, head of alignment research at Anthropic.
And Hubinger publicly stated that he believes there is a greater than 10% probability that artificial intelligence could cause human extinction during the next decade. He clarified, however, that his concern does not refer to current models, whose risk he considers reduced, but rather to future superintelligence arising from processes of recursive self-improvement.

OpenAI also considers reducing the pace

Anthropic is not alone among large companies either. OpenAI has been discussing internally the possibility of synchronizing the development of its systems with other laboratories and Sam Altman has supported the proposal of providing independent evaluators with a level of access similar to that of the companies’ own workers.

“This has been one of the main topics of the conversations we have had at OpenAI in recent weeks”noted Altman when reacting to Amodei’s proposal. The CEO also assured that OpenAI will also adopt the commitment to provide access to independent evaluators.

The position fits with that recently expressed by Jakub Pachocki, Chief Scientist at OpenAI, who has argued that companies should be able to coordinate slowdowns when necessary until shared security thresholds are established. According to information published by BloombergOpenAI would have already stopped certain internal training and reduced some development work due to security concerns.

But coordination between competitors also raises a less technological problem: whether it would be legal. OpenAI has consulted members of the US Congress about the implications of major companies jointly agreeing to reduce the pace of development, given the possibility that coordinated action between competitors would raise conflicts with antitrust legislation.

China, the great limit of any agreement

Amodei’s proposal incorporates a contradiction that the manager himself recognizes: he wants to slow down the development of artificial intelligence, but not at the expense of China overtaking the United States.

The CEO of Anthropic believes that Chinese leadership in artificial intelligence would pose a risk to the United States and democracies, and defends maintaining controls on the export of advanced chips and combating mechanisms such as the unauthorized distillation of Western models.
Its approach involves, in practice, first preserving a sufficient technological advantage for the United States and its allies and, from that position, subsequently building international agreements that allow the pace to be reduced in a coordinated manner.

This geopolitical dimension also introduces another issue into the debate. The rules proposed by Amodei are relatively easy to apply to closed models offered through services controlled by companies such as Anthropic or OpenAI, but they are much more difficult to translate to open models that can be downloaded, modified and executed outside of their developers’ infrastructure.

The security of artificial intelligence is thus also linked to the technology distribution model and the risk that more demanding standards end up concentrating the development of the most capable systems in a small group of companies. This dilemma between security and technological leadership also explains Donald Trump’s reaction. The US president has downplayed the most alarmist scenarios and insisted that his priority is for the United States to maintain the advantage over China.

“We are ahead of China in AI and, frankly, I want to keep it that way, because whoever wins in AI, wins”

donald trump

House Speaker Mike Johnson has expressed a similar position, warning that hasty legislative intervention could even become a national security threat if it ends up weakening American companies against their Chinese rivals.

On the other side of the debate, Barack Obama has called for artificial intelligence to occupy a priority place on the American political agenda. The former president believes that its development can accelerate advances in areas such as medicine or clean energy, but warns of the risks of allowing a technology of this magnitude to evolve mainly under the control of private companies without sufficient supervision.

The discussion about the speed of artificial intelligence has thus passed in just a few days from the security teams of the laboratories to the offices of their top executives and, from there, to American politics. The question goes beyond how far the models can go: the focus is now on who is willing to slow down, under what rules and, especially, who will agree to do it first.

More info: Essay “We Must Pace the Frontier” by Dario Amodei