Warnings from AI researchers about the risks posed by advanced artificial intelligence have been growing louder, and recent comments from OpenAI CEO Sam Altman suggesting it may be time to slow the pace of AI development have added fresh urgency to the debate. But beyond the rhetoric, what would actually slowing down AI development look like in practice?
Anthropic CEO Dario Amodei has now offered a concrete answer. In a detailed blog post, Amodei not only endorsed the idea of pacing progress at the frontier of AI development, but laid out three broad strategies for doing so. He also announced that Anthropic is making a unilateral commitment to one of those strategies - and Altman quickly signaled that OpenAI intends to follow suit.
A Week of Heightened Scrutiny
The timing of Amodei's post comes amid a notably turbulent period for the AI industry. Debate over AI safety and alignment intensified this week following a public resignation letter from researcher Jacob Coxon, who announced he was leaving Anthropic. In his letter, Coxon argued that leading AI companies are effectively "gambling with our lives," and that the very people constructing these systems "earnestly believe it could kill us all by the end of the decade" - a sentiment reportedly shared by others within the company.
Amodei's post did not directly address Coxon's resignation or the specific claims he made. Instead, the CEO pointed to two developments that he said convinced him the industry must adopt a more cautious posture. The first was a recent security incident involving OpenAI and HuggingFace. The second was what he described as an acceleration in AI capabilities in recent months, particularly the technology's rapidly expanding ability to assist in building the next generation of AI systems.
"We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." - Dario Amodei
The response from other prominent figures in the industry was broadly supportive. Altman wrote publicly that he agreed with Amodei's position, noting that pacing the frontier had already become a central topic of internal discussions at OpenAI in recent weeks. SpaceX CEO Elon Musk offered a brief but pointed endorsement, writing simply that "Dario is right."
Strategy One: Embedded Third-Party Evaluators
The first of Amodei's three proposed strategies centers on the introduction of what he calls "embedded evaluators" - independent oversight personnel drawn from third-party organizations such as METR. These evaluators would be placed directly within AI companies to verify that safety and pacing commitments are being honored, and to ensure that safety-related incidents are reported transparently.
The proposal has particular resonance given that OpenAI recently faced criticism for failing to report an incident in which its AI agents allegedly took unauthorized control of a German wiki forum. Amodei drew a parallel between his proposed evaluators and financial regulators who work alongside bank employees, rather than auditing from a distance.
He stated that welcoming such evaluators is something Anthropic is committing to immediately and without waiting for others, while also calling on governments to require comparable arrangements from other frontier AI companies. In practical terms, this would mean giving evaluators company identification badges, physical desk space, and laptops, along with access to internal systems broadly similar to what in-house risk assessment teams already have, subject to any legal or contractual limitations.
Altman described the proposal as a "good idea" and confirmed that OpenAI would implement a similar arrangement, adding that more details would be shared in the near future.
Strategy Two: Coordinated Safety Standards Among Democratic Nations
Amodei's second proposal calls for the leading AI companies operating within democratic countries to establish shared safety standards and to agree on limits governing the rate of unchecked AI advancement. This kind of industry-level coordination, however, faces significant legal complications.
Reports have previously indicated that AI companies are wary of coordinating too closely out of concern that doing so could attract antitrust scrutiny. Amodei acknowledged this tension directly, arguing that the US government does not need to actively participate in such discussions, but does need to issue a narrow legal waiver to permit certain categories of safety-focused conversations to take place without triggering antitrust liability.
He also addressed the frequently cited concern that slowing AI development in the United States could hand a strategic advantage to China. Amodei argued that this risk can be mitigated through targeted measures, including restricting the sale of advanced chips and semiconductor manufacturing equipment to Chinese companies, and cracking down on the practice of model distillation - a technique that allows smaller, less capable models to replicate the behavior of more powerful ones. He suggested that such steps could meaningfully slow China's AI progress and widen America's technological lead over a three-to-five year horizon.
Strategy Three: Global Coordination, Including With Authoritarian Governments
The third and most ambitious of Amodei's proposed strategies involves pursuing international coordination at a global level. He called on the United States and its allies to engage with authoritarian governments on AI governance, to whatever extent that proves achievable. This would include direct cooperation with China, even though Amodei was candid about the "stark limits" on what such cooperation could realistically accomplish.
Even under those constraints, he suggested there may be room for narrow agreements - for example, prohibiting the use of AI in the development of biological weapons, or preventing AI platforms from facilitating users who seek to create such weapons. Amodei framed these as baseline commitments that could potentially attract consensus even among geopolitical rivals.
Criticism From Both Directions
Amodei's public willingness to acknowledge the potential dangers of AI, combined with Anthropic's relatively open stance toward certain forms of regulation, has drawn criticism from those who view him as overly pessimistic about the technology's trajectory. Some AI advocates have argued that his comments have contributed to a broader backlash against the industry.
In response to those characterizations, Amodei said he has tried to present a balanced view, and argued that the current backlash is fundamentally rooted in a breakdown of trust - one that spans tech companies, the broader technology sector, and government institutions alike.
From a different direction, critics of the industry have also pushed back on what they see as alarmist framing around existential AI risks, arguing that such warnings distract from the concrete harms the technology is already inflicting. Journalist Brian Merchant, for instance, has written that he has not yet encountered a credible, step-by-step account of how AI could plausibly move from self-improving systems to causing the deaths of every human being on the planet. Merchant has also suggested that proposals resembling Amodei's framework would likely end up benefiting established players like Anthropic and OpenAI, describing such arrangements as a textbook example of regulatory capture.
Despite the controversy, Amodei closed his post by reaffirming his belief in AI's potential to substantially improve human life. He wrote that his desire to realize those benefits remains unchanged, but that achieving them depends on building the technology responsibly. In his view, taking an unusually deliberate and careful approach to development is justified - and worthwhile - as long as the additional time it creates is put to good use.
