'Anthropic Has Never Advocated For A Ban On Open-Weights Models': Dario Amodei Sets The Record Straight

Anthropic CEO Dario Amodei has moved to clarify the company's position on open-weights AI models, pushing back against recent claims that Anthropic favours restricting them, particularly those built in China. In a new blog post, Amodei laid out a detailed rebuttal to the idea that Anthropic sees such bans as useful policy, while also outlining what he believes are the real risks tied to powerful AI systems, open or closed.

Why did Amodei feel compelled to respond?

The clarification comes amid a broader debate in Washington over whether US officials should restrict American companies from using Chinese open-weights models. Several tech companies had signed a letter defending open-weights models more broadly, and in that context, some critics accused Anthropic of quietly pushing for a ban to shield its own commercial interests. Amodei rejected that characterisation outright, stating plainly that Anthropic has never advocated for a ban on open-weights models. He argued that open-weights models without dangerous capabilities function as a public good, offering value to businesses, developers, and researchers at little more than the cost of the compute needed to run them.

The two scenarios Amodei says actually worry him

Rather than the models' licensing structure, Amodei said his concern centres on two specific risks he has held for years and had already detailed in an earlier essay. His primary worry is that authoritarian governments, with the Chinese Communist Party as the most capable example, could build AI systems more powerful than anything developed in the US and use them to lock in lasting military dominance or intensify repression domestically.

He noted that this concern is shared within the US government, pointing to warnings from Vice President JD Vance and language from the Intelligence Community's 2026 Annual Threat Assessment about other global powers challenging American economic and security advantages through AI progress. Notably, Amodei argued that whether such a model is released openly, or even used by American firms, is beside the point, since the most dangerous version of this scenario would involve a model trained in secret and handed directly to state military or surveillance agencies rather than released at all.

His second concern is that powerful models, regardless of origin, could be misused for cyberattacks or biological attacks, or could carry serious alignment flaws. He acknowledged that open-weights models can pose a somewhat higher risk here simply because they are harder to monitor or restrict once released, and cannot be recalled. But he maintained that banning US businesses from using such models does little to address this risk, since bad actors are rarely legitimate American companies to begin with, and such a ban would mainly serve to protect US AI firms from competition, something he said was never Anthropic's intention.

What does Amodei recommend?

In place of a blanket ban, Amodei outlined three measures he said Anthropic has consistently supported. The first is tightening controls on advanced chips and chipmaking equipment to China while cracking down on smuggling, arguing that scaling laws mean China cannot out-build the US on frontier models without access to American chips. The second is targeting industrial-scale distillation operations, which he said allow China to extract stronger capabilities from a limited chip base and close the gap with the US frontier, even if such operations are not fundamentally an open-weights issue. The third is mandatory pre-release safety testing for all sufficiently capable models, open or closed, to directly evaluate cyber, biological, and alignment risks rather than assuming those risks in advance. He noted this idea already has some momentum, citing recent moves by the Trump administration and industry proposals along similar lines, though he cautioned that such testing would need global buy-in, including from China, to be truly effective.

Where he agrees, and disagrees, with the open letter

Amodei said he agreed with much of the open letter's broader case for open-weights models, including that they expand access to the AI economy, strengthen competition, and give customers more control. However, he pushed back on the letter's assumption that open-weights models automatically make safeguards easier to build or that broad access to AI capabilities helps defenders more than attackers. He pointed specifically to biology as an area where he believes the opposite may be true, warning that sufficiently capable models could potentially help weaponise dangerous pathogens far faster than defensive countermeasures could realistically be developed. He argued that such questions should be settled through rigorous testing rather than assumed one way or the other.