The ‘AI Slowdown’ Is Good For Tech CEOs, Not For You

AI Industry Leaders Call for Coordinated Slowdown in Frontier Model Development

Proposals for a temporary slowdown in frontier AI research gained traction this summer after initial calls from activist groups were joined by executives at leading AI companies and some elected officials. The suggestions include coordinated limits on the most advanced model training runs, with access potentially restricted to a group of democratic nations. Proponents argue this would allow time for public discussion of governance and safety measures.

Anthropic CEO Dario Amodei wrote that society should have input on technology use and that additional time for deliberation would be beneficial. Former Anthropic and OpenAI researcher Jacob Coxon posted on social media that some developers believe advanced AI systems could pose severe risks by the end of the decade. His statements received widespread attention and were referenced by policymakers in several countries. Former colleagues at the two companies publicly supported parts of his assessment, and Coxon appeared on news programs to elaborate.

Both Anthropic and OpenAI are preparing for initial public offerings that analysts have described as potentially among the largest in technology history. The companies currently lead in closed-source model development. Supporters of the slowdown proposals state that the measures address safety and alignment challenges. Critics contend that any resulting regulatory framework would favor firms with substantial legal and compliance resources, making it harder for smaller or open-weight developers to compete at the frontier level.

Closed-source systems are hosted centrally by their developers and can incorporate usage restrictions or monitoring. Open-weight models, released by organizations including Google, NVIDIA, and several Chinese firms such as Kimi, Moonshot, and Deepseek, can be downloaded and run locally. Once distributed, these models lack centralized controls that could enforce future usage rules. Some lawmakers have proposed restrictions specifically on models originating from certain countries, citing national security considerations.

NVIDIA’s acquisition of the open model platform Hugging Face was presented by the company as support for decentralized development. The deal has been cited by advocates of open approaches as evidence that such models can continue to advance without the same oversight structures applied to frontier labs. Proponents of open models argue that local deployment allows greater user control and experimentation. Others note that the absence of built-in restrictions increases the potential for misuse, a point raised in discussions of both safety and export controls.

Public opinion surveys have shown higher levels of concern than enthusiasm regarding rapid AI progress. Experts have cited concrete issues including bias in training data, potential for generating misleading content, and challenges in verifying model outputs. Separate analyses have examined whether open-weight releases could accelerate capability diffusion to actors outside established labs. These considerations form part of the stated rationale for pacing advanced research, alongside arguments that competition among multiple developers has historically driven performance gains.

The debate centers on whether temporary limits on the largest training runs would primarily serve safety objectives or create barriers that consolidate market position among the current leaders. Both perspectives are represented in statements from company executives, researchers, and policymakers. Implementation details, such as the scope of any oversight body and criteria for permitted research, remain under discussion.