Anthropic CEO Dario Amodei said on July 27, 2026 that "Anthropic has never advocated for a ban on open-weights models," responding to reports that some US officials were considering barring US companies from using Chinese open-weights models. Instead, he restated three measures the company supports: keeping powerful chips away from China, cracking down on industrial-scale distillation and requiring safety tests for every sufficiently capable model, open or closed. For developers who run open-weights models on their own hardware or in their own clouds, the debate shows that model choice now carries policy risk.

Key Facts#

  • Statement: Amodei published the post on July 27, 2026, and edited it on July 28 to credit AE Studio as a collaborator on research he cited.
  • Trigger: He pointed to reports that some US officials were considering a ban on US companies using Chinese open-weights models, to a letter from many tech companies supporting open-weights models, and to accusations that Anthropic wanted a ban to protect its business.
  • Position: Anthropic has never advocated a ban, and Amodei described open-weights models without dangerous capabilities as a public good.
  • Primary concern: Authoritarian governments, with the Chinese Communist Party (CCP) the most capable among them, building AI more powerful than America's and using it for military superiority or repression.
  • Secondary concern: Misuse for cyber or biological attacks and serious alignment problems. Open-weights models may carry higher risk because guardrails are hard to apply, usage is hard to monitor and released weights cannot be withdrawn.
  • Three measures: No sales of powerful chips or chipmaking equipment to China, plus a crackdown on smuggling; action against industrial-scale distillation; and mandatory safety testing for all sufficiently capable models, with less capable models exempt.

What Happened#

Amodei argued that a ban would miss both of his concerns. On the first, he said it is irrelevant whether an authoritarian government's models are released openly or used by US businesses; the most dangerous model might be trained in secret and handed only to the People's Liberation Army and China's Ministry of State Security. On the second, he accepted that open weights can raise misuse risk but said that stopping US businesses from using them does nothing about it, because bad actors are unlikely to be legitimate US businesses. A ban would mainly protect US AI companies from competition, which he said has never been his goal.

Testing was the center of his proposal. Amodei called mandatory pre-release testing for cyber, biological and alignment risks "close to a consensus," and said he was encouraged both by recent moves in that direction by the Trump administration and by industry proposals that would test the most capable models regardless of their country of origin or openness, while exempting less capable models from startups and academia. Whether open models pose extra risk, he wrote, should be established by testing rather than assumed, and research such as the modular training work by AE Studio and Anthropic may help make open-weights models safer. To be effective, testing would have to be global, including China, which he considered possible at least for preventing AI-enabled biological weapons.

He also responded directly to the industry letter. Amodei agreed that open weights widen access to the AI economy, strengthen competition for some uses and give customers more control, and that distillation concerns should be handled through targeted legal and commercial frameworks. He disagreed with its claims that open-weights models necessarily make safeguards easier to build or that broad access necessarily helps defenders more than attackers, warning that biology in particular may favor attackers.

Background#

The statement followed months of Anthropic policy work focused on China. In February 2026 the company reported industrial-scale distillation campaigns by DeepSeek, Moonshot and MiniMax, and in May its 2028 scenarios paper urged tighter chip export controls and action against distillation. Amodei acknowledged that many companies carrying out distillation also release open-weights models, but argued that state backing, not openness, is the real issue. DeepSeek, one of the labs Anthropic named, publishes open-weight models such as its V4 family.

His testing proposal is consistent with Anthropic's Advanced AI Framework, published in 2026, which asks governments to require testing, independent evaluation and security programs for the most capable models. Its rules would apply only to models trained with more than 10^25 FLOPs by companies with more than $500 million in AI-related revenue or more than $1 billion in AI R&D spending, so smaller developers would fall outside it.

Why It Matters for Developers#

Many .NET teams now run open-weights models through Ollama, ONNX Runtime or Foundry Local, as covered in our local AI guide. Even without a ban, the debate points to practical steps.

  • Inventory model provenance. Record the publisher, country of origin, license, version and file hashes of every open-weights model you deploy. A rule that targets models by origin would otherwise force a scramble to find out what you run.
  • Make replacement measurable. Put model calls behind IChatClient from Microsoft.Extensions.AI and keep evaluation suites, as in evaluating AI applications in .NET, so swapping one open model for another becomes a measured migration rather than a guess.
  • Treat weights as supply chain artifacts. Pin versions, verify checksums and control where weights come from, using the practices in the software supply chain security guide.
  • Watch for testing requirements. If mandatory pre-release testing applies to capable open models, publishers' test results may become part of your due diligence, while smaller models may be exempt.
  • Respect distillation terms. Fine-tuning an open model on outputs from a closed model can breach the provider's terms, and distillation has become a policy issue in its own right.

What's Next#

Amodei's post described officials as considering a ban, not adopting one, and it gave no indication of when or whether a decision would come. Whether US policy ends up targeting Chinese open-weights models by origin, by capability or through testing is the key open question. Amodei's hope that China could join a global testing regime, at least on biological risks, is his own expectation rather than an announced plan. Anthropic, for its part, continues to argue for chip controls, anti-distillation measures and testing of all sufficiently capable models instead of a categorical ban.

Sources#