The boss of Anthropic has hit back at claims his company is calling for a ban on overweight Chinese models, days after writing a letter to a rival industry.
Anthropic’s chief executive, Dario Amodei, has moved to retract what he called an allegation circulating in AI circles this week, that his company is seeking a ban on open-weight Chinese models in the US as a way to protect its business.
In a post titled ‘Our position on open-weights models’ on the Anthropic website yesterday (27 July), Amodei wrote that “Anthropic has never advocated the banning of open-weights models”, and said that open-weights models that do not have harmful power are useful to the public that cost nothing more than the computer needed to run them.
This post follows an open letter published on July 24 and hosted on Nvidia’s website, entitled ‘Open Weights and American AI Leadership’, which urged American policymakers not to impose broad restrictions on open weight models. The letter was signed by almost all major technology players including Nvidia, Microsoft, Meta and Hugging Face, and later expanded to include OpenAI and Google – 50 signatories today. Anthropic did not sign.
The letter came amid reports that U.S. officials were considering restrictions on the use of Chinese open-source models by U.S. businesses, first reported by Axios on July 21.
Amodei raised two concerns that he said banning defenders would not address. The first is that authoritarian governments, not just the Chinese Communist Party, can build AI models more powerful than the US and use them for military advantage or domestic repression. He noted that this is a concern whether such models carry open weights or not.
His second concern is that dynamic models, open or closed and regardless of origin, can be misused in cyber attacks or biological attacks. He says the open-weight models carry more risk here because the guardrails are difficult to use once the weights are released, and the released weights cannot be retracted.
That said, Amodei says that preventing American businesses from using such models won’t stop bad actors, who may be legitimate US companies in the first place.
Instead of a ban, Amodei said Anthropic continues to support three measures: to prevent the sale of powerful chips and chipmaking equipment in China and to stop smuggling; directed industrial-scale distillation operations that allow Chinese labs to bridge the gap at the US border without the same computer; and mandatory security assessments, including cyber, biological and compliance risks, on all sufficiently capable models regardless of whether they are open or closed, or where they are built.
He said he agreed with parts of the industry book, including that open standards expand access to the AI economy and strengthen competition, but disputed its suggestion that broader access helps defenders more than attackers.
He cited biological risks as an example, saying that a sufficiently skilled model could help to quickly arm an epidemic-level virus, while building a defense against such an attack is a task of many years or so in the best case scenario.
Amodei said questions about whether open models carry more risk, and whether that risk can be mitigated, should be resolved through pre-release testing rather than preconceived notions.
Don’t miss out on the information you need to succeed. Sign up for Daily BriefSilicon Republic’s digest of must-know sci-tech news.
