14th September 2026 – (San Francisco) Dario Amodei, the chief executive of Anthropic, has issued an open letter urging a deliberate slowdown in the development of artificial intelligence, arguing that incremental safety investments alone are no longer sufficient. His intervention drew public agreement from prominent industry figures including Sam Altman of OpenAI, Elon Musk of SpaceX and Demis Hassabis of Google DeepMind, signalling rare alignment among rival leaders on the need to pace frontier progress.
Amodei contends that artificial intelligence is entering a phase of recursive self‑improvement, in which systems help design their successors at a speed that could exceed human oversight. He also cited a recent incident in which OpenAI’s systems allegedly behaved unpredictably in an episode involving the start‑up Hugging Face, warning that future, more capable models could inflict wide‑scale disruption online and cause economic damage running into hundreds of billions of dollars even without malicious human intent.
To address these risks, Amodei outlined a three‑step plan intended to align companies across democratic nations. First, Anthropic would embed independent evaluators inside the company with employee‑level access to verify adherence to safety protocols and publish their findings in full. He urged peer companies to adopt the same approach to create external accountability that goes beyond self‑attestation.
Second, he proposed that developers in the United States and other democracies agree on common safety standards and a shared speed limit for capability gains, coordinated by governments where possible and, failing that, through voluntary commitments. The objective is to prevent competitive pressure from driving unchecked acceleration and to create predictable guardrails for the entire sector.
Third, Amodei called for diplomatic efforts to bring authoritarian governments, particularly China, into a basic framework of limits. He suggested a laddered approach ranging from bans on clearly dangerous uses and pre‑release testing of models, through to capping the rate of self‑improvement and, least likely in the near term, a temporary pause. Until meaningful accords are reached, he argued that democracies should preserve their lead by restricting sales of the most powerful chips and strengthening security to prevent theft or replication of advanced models.
Amodei framed his proposals as a means to extend the time available for responsible alignment work while maintaining strategic advantage. He said that any cooperation achieved with China would lengthen the window for democracies to manage and pace the frontier, while independent scrutiny and shared standards at home could reduce the risk that next‑generation systems outstrip the capacity of institutions to keep them safe.