Amodei open weights stance rejects ban while warning on AI risks

Anthropic CEO Dario Amodei used a new company blog post to restate his position on Amodei open weights: he says Anthropic is not seeking a ban, while arguing that the most capable open-weight AI models create security risks that current policy does not adequately address. The intervention lands as U.S. officials are reportedly considering measures that would make it harder for American companies to use Chinese open-weight models, and as major tech companies have signed a letter opposing that kind of restriction.

Amodei wrote that “Anthropic has never advocated for a ban on open-weights models,” and said models without dangerous capabilities can be a public good. His line is narrower than an outright anti-open-source position, but it still puts Anthropic on the restrictive side of a fight that has become commercially and politically charged across the AI industry.

What is Amodei saying about open weights?

Amodei’s argument is that open-weight releases become harder to control once a model is published: safeguards can be removed, and the model cannot be recalled. He said the risk rises when a model has capabilities that could be used in cyber operations, biological threats, military systems or surveillance.

He described two worst-case outcomes. In one, authoritarian governments, particularly China, develop more capable models than the United States and use them for military advantage or surveillance. In the other, powerful open models are used by attackers for cyber or biological harm. The Decoder reported that AI companies already see users attempting such activity on closed models as well, based on usage data.

Amodei said open models can also help defenders, but argued that the balance looks different for biological threats. In his view, a sufficiently capable model could speed up attempts to weaponize viruses, while defensive work would still take years even under favorable conditions.

What policies does Anthropic’s CEO want?

Amodei called for tighter chip export controls intended to prevent China from training more powerful models than U.S. companies without access to U.S. chips. He also urged action against industrial distillation, a technique he says China uses to partly work around chip limits and benefit from U.S. AI progress.

Those claims remain disputed. Critics cited by The Decoder argue that restricting U.S. chip sales may accelerate China’s domestic chip industry. Others say distillation is a legitimate AI development technique, a point that has become more prominent as AI labs restrict rivals from training on their systems while using broad datasets themselves.

Amodei also backed mandatory safety testing for powerful models regardless of origin or license. He pointed to recent proposals that would exempt less capable models used by startups and researchers, suggesting that he sees momentum around rules aimed at capability thresholds rather than licensing labels alone.

Why the fight matters for AI companies

The open-weight debate is no longer only a research licensing issue. It affects cloud vendors, model providers, app developers and companies building on lower-cost models from China and elsewhere. Open-weight systems can reduce dependency on closed model APIs, which makes them a commercial threat to companies selling proprietary frontier access.

That commercial context is why Anthropic’s position is drawing scrutiny. Critics have accused the company of criticizing open-weight models to defend its own business from cheaper competition. OpenAI has also reportedly continued lobbying against open weights, even though it signed the industry letter opposing restrictions on Chinese open-weight models.

Anthropic did not disclose any new model, pricing change or customer commitment in connection with Amodei’s post. The company’s stated position is regulatory: keep less capable open models available, require safety tests for powerful models, and use export controls to slow Chinese AI progress.

This story draws on original reporting from The Decoder.