Anthropic CEO Dario Amodei clarifies stance on open-weight AI after industry backlash

The open letter and the backlash

On 24 July, Nvidia CEO Jensen Huang posted an open letter titled "Open Weights and American AI Leadership" on X, his first post on the platform. The letter, signed by 25 companies including Meta, Microsoft, Hugging Face, Mistral, and Palantir, urged policymakers not to impose broad "premature restrictions" on open-weight AI models. Over the weekend, OpenAI, Google, SpaceX, Amazon, Cohere, and others added their names, leaving Anthropic as the only major AI lab that had not signed.

Silicon Valley noticed. Venture capitalist David Sacks, a former AI adviser to the Trump administration, wrote on X that "the entire tech industry (save for Anthropic) has come out in favor of open source AI." Benchmark general partner Bill Gurley suggested Anthropic's reluctance reflected that open weights compete "with their corporate economic strategy." Kai-Fu Lee, founder of Chinese startup 01.AI, posted that "who DIDN'T sign the Open Weight letter is far more interesting than who did."

Amodei's response

On Sunday, Dario Amodei published a blog post breaking the silence. "Anthropic has never advocated for a ban on open-weights models," he wrote, adding that "open-weights models that don't have dangerous capabilities are a public good." He said his fears about AI do not extend to businesses using open-weight models, even Chinese ones, but rather to authoritarian governments building models more powerful than those in the US to achieve "permanent military superiority" or repress their own people. The Chinese Communist Party, he noted, is the "most capable" such government.

Amodei also warned that open-weight models pose a greater risk in biological attack scenarios because guardrails cannot be applied and, citing a UK AI Security Institute report, once weights are released they cannot be withdrawn.

Three measures instead of a ban

Instead of banning open-weight models, Amodei outlined three policy measures. First, enforcing chip export controls and cracking down on the "rampant smuggling" used to get advanced chips into China, because without US chips, scaling laws prevent China from training models more powerful than American ones. Second, stopping industrial-scale distillation operations that allow Chinese labs to build capable models far more cheaply than training from scratch. Third, mandatory safety testing for all sufficiently capable models, open and closed, regardless of country of origin, with less capable models from startups and academia exempted.

The distillation accusation

The distillation point is commercially charged. Anthropic accused Alibaba's Qwen lab of running the largest distillation campaign ever against Claude, using 25,000 fake accounts for 29 million exchanges. Amodei acknowledged that cracking down on distillation is "challenging" because accounts can often only be identified after substantial distillation has occurred, and called for policy intervention rather than relying on any single company's enforcement.

The Fable 5 precedent

Anthropic's caution has a recent precedent. On 12 June, three days after launching Fable 5 as its most powerful public model, the US government ordered the company to bar foreign nationals from Fable 5 and Mythos 5. Enforcing a nationality rule across a shared cloud service was impractical, so Anthropic switched off access entirely. The episode underscored the company's long-held view that developers lose control of a frontier model the moment its weights are published, a position it has maintained through gated-access programs like Project Glasswing.