Anthropic has rejected accusations that it wants the US government to restrict open-weight AI models, saying such a move would not solve the company’s biggest national security concerns.
In a blog post Monday, Amodei said, “Anthropic has never advocated for a ban on open-weights models,” pushing back against criticism that the company was seeking restrictions to protect its own closed AI products.
The comments came after more than 50 technology companies backed an open letter urging policymakers not to impose broad limits on open-weight models. Companies including Nvidia, Microsoft and Meta signed the letter, while OpenAI later added its support. The signatories argued that open-weight systems improve access, competition and customer control.
Anthropic did not sign the letter, leading some industry critics to accuse the company of favoring restrictions that could disadvantage Chinese AI competitors that have released powerful open-weight models.
Why Anthropic sees greater risk in open weights
Amodei said open-weight models that do not have dangerous capabilities provide public benefits because they allow businesses, developers and researchers to use AI without paying for access beyond the computing costs required to run them.
However, he said two major risks remain.
The first is the possibility that authoritarian governments could develop AI systems more powerful than those built in the US and use them to obtain military advantages or expand domestic surveillance and repression. Amodei said this concern is not limited to open-weight models because the most dangerous systems could be privately developed and kept secret.
The second concern involves misuse of advanced AI for cyberattacks, biological threats and other harmful activities. Amodei argued that open-weight models can create additional challenges because developers cannot easily apply safeguards or monitor how released model weights are used.
“Open-weights models—it does not matter whether they come from China or anywhere else—do potentially present a higher risk than closed models, because it is very difficult to apply guardrails to them or monitor their usage, and once weights are released they cannot be withdrawn,” Amodei wrote.
What Anthropic wants instead
Rather than banning open-weight models, Anthropic is calling for three policy measures. The company supports tighter restrictions on exporting advanced AI chips and chipmaking equipment to China, arguing that access to computing power is a key factor in building frontier AI systems.
The company says that industrial-scale distillation can enable rival developers to reproduce capabilities from leading models with significantly less computing investment. Amodei argued that this practice can help competitors close the gap with leading AI labs without investing the same level of computing resources.
Finally, Anthropic is pushing for mandatory pre-release testing of sufficiently capable AI models for cyber, biological and alignment risks, regardless of whether the models are open or closed.
A policy debate over AI leadership
The disagreement highlights a growing divide over how policymakers should balance AI innovation, competition, national security and safety concerns. Supporters of open-weight models say restricting access could hurt startups, researchers and businesses that rely on flexible AI tools. Critics argue that unrestricted access could make it harder to prevent misuse.
Amodei said some questions about open models should be settled through testing rather than assumptions. He also agreed with parts of the industry letter, including the idea that open models can increase access and competition. However, he disagreed with the argument that wider access automatically helps defenders more than attackers.
The challenge for policymakers is finding a middle ground. Heavy restrictions could slow legitimate research and hurt startups that depend on open models, while ignoring potential risks could make it harder to respond after dangerous capabilities spread.
Read more: Moonshot AI's new Kimi K3 model highlights how Chinese developers are advancing in the open-weight AI race.