Image: Anthropic / anthropic.com Anthropic's Dario Amodei: 'We Never Advocated for a Ban on Open-Weights Models'
Responding to accusations of opposing open-source AI, Anthropic's CEO clarified the company's actual position: open-weight models are a public good, with one serious exception.
Anthropic CEO Dario Amodei published a statement yesterday clarifying the company’s position on open-weight AI models, after weeks of industry pressure suggesting Anthropic was quietly trying to restrict open-source AI development.
“Anthropic has never advocated for a ban on open-weights models,” Amodei wrote. “Open-weights models that don’t have dangerous capabilities are a public good.”
The statement came in direct response to a letter organized by Nvidia’s Jensen Huang and signed by roughly 50 companies, including Microsoft, Meta, OpenAI, and Google, in support of open-weight AI development. Anthropic wasn’t on the letter. Critics read the absence as implied opposition.
What Amodei actually thinks
The concern isn’t that models are open. It’s about what happens when very capable models become freely available to adversary governments.
Amodei specifically names the Chinese Communist Party as the most capable threat, though he frames the risk as authoritarian governments generally. His worry: a state-backed program could use open-weight U.S. models as training data for distillation, effectively compressing years of expensive frontier research into cheap copies. That closes the AI capability gap without requiring independent research.
The endgame in this framing isn’t that someone downloads a model and does something dangerous with it directly. It’s that systematic distillation lets an adversary build a competitive AI ecosystem without the cost of building it from scratch.
What he’s recommending instead
Amodei proposes three measures rather than a blanket ban:
Tighter chip restrictions. Export controls on advanced chips to China already exist, but Amodei argues enforcement needs to improve, specifically around smuggling. Chips are the bottleneck; cutting off access to them limits how much any distillation operation can actually produce.
Formal rules against industrial-scale distillation. Using large models to train smaller ones is standard practice. Doing it at state scale to systematically close the capability gap is different. Amodei wants policy that distinguishes between the two.
Mandatory safety testing for all capable models. This would apply regardless of whether models are open or closed, and regardless of country of origin. He argues that even China might cooperate on this specific measure because mutual interest in preventing AI-enabled biological weapons is strong enough to overcome other tensions.
The context that makes this interesting
Anthropic had its own brush with export controls in June. Fable 5, Anthropic’s most capable model, was briefly pulled from international distribution under U.S. restrictions before Anthropic negotiated a framework for controlled release. Being on the receiving end of export rules gave Amodei a different perspective than most AI executives writing about this topic.
The statement is also well-timed. Congress is actively working on AI legislation, and the industry is lobbying hard on open-weight model rules. A clear, public position from Anthropic keeps the company from being misread as an ally of the most restrictive proposals.
For developers, the practical implication is limited. Nothing Amodei is proposing would restrict open-weight models in common use today. The proposals target state-scale distillation operations, not individual developers downloading weights. The debate matters for the regulatory environment around AI development broadly, but the code you’re writing today isn’t affected by it.
Sources: Anthropic — TechCrunch — CNBC