Anthropic CEO Dario Amodei Clarifies His Stance: Not Against Open Source, but Advocates Strict Controls on Compute, Distillation, and Pre-Release Testing

Amid reports that U.S. officials are weighing restrictions on domestic companies using open-source AI models from certain countries, Anthropic co-founder and CEO Dario Amodei has issued a public clarification: Anthropic has never argued for banning open-source models.

The statement comes as debate over open-source AI has intensified across the industry. A number of technology companies have backed open development through public letters, while some critics have accused Anthropic of trying to use government action to weaken competitors and protect its own commercial position. Amodei rejected that framing, saying open models can be highly valuable public goods when they do not cross into dangerous capability levels. In his view, they can offer lower-cost tools that help businesses, developers, and researchers move faster.

He also argued that broad, protectionist-style bans would do little to address real national security concerns. Instead, such measures could end up functioning as tools for market concentration and industry lock-in, which he said is neither Anthropic’s position nor its objective.

Amodei’s concerns are centered on two different kinds of risk

In explaining his position, Amodei separated AI-related security concerns into two main categories rather than treating the open-source question as the central issue.

1. More advanced AI in the hands of authoritarian states

The first concern is that authoritarian governments could use stronger AI systems than those available to the United States to gain a lasting military edge or deepen domestic surveillance capabilities.

According to Amodei, whether a model is open source, or whether U.S. companies are allowed to use it, is not the decisive factor in that scenario. The more serious danger, he suggested, would come from systems developed in secret and deployed only for military or intelligence purposes.

2. Misuse of powerful models for cyberattacks or biological threats

The second concern involves high-capability AI being abused for large-scale cyber operations or for work related to biological weapons.

Here, Amodei acknowledged that open models can create tougher safety challenges because they typically do not come with ongoing monitoring or protective controls once released. After distribution, they cannot realistically be pulled back. Even so, he said a simple ban on law-abiding domestic companies using open models would not stop determined malicious actors, and therefore would not meaningfully reduce the underlying threat.

Anthropic’s preferred policy approach

Rather than focusing on an outright open-source crackdown, Amodei pointed to three policy areas that Anthropic believes would be more practical and effective.

Keep tight control over critical compute

The first recommendation is to continue restricting access to advanced chips and the manufacturing equipment needed to produce them, while also stepping up enforcement against smuggling and other methods used to evade export rules. The goal is to block high-risk actors from obtaining the computing power required to train frontier-level systems in the first place.

Address large-scale cross-border model distillation

Amodei also warned about organizations using distillation and similar low-cost techniques to extract capabilities from leading closed models at scale, dramatically shortening their own development timelines.

He argued that when such activity is used to bypass compute restrictions or violate usage terms, governments and companies should have clearer legal and commercial tools to respond. In his view, this is a more concrete governance problem than the broad question of whether open-source AI should exist.

Require strong pre-release safety testing

The third proposal is mandatory testing before release for any model with sufficiently advanced capabilities, regardless of whether it is open or closed.

Amodei said these evaluations should be comprehensive and should cover areas such as cybersecurity, biosecurity, and alignment-related risks. Safety, he argued, should be judged through rigorous empirical testing rather than assumption or branding.

Shifting the debate away from “open versus closed”

Amodei’s broader message is that the AI sector should move past a simplistic fight between open-source and closed-source camps. In his view, the more urgent questions involve compute controls, the governance of improper distillation practices, and the creation of credible safety testing standards that can be applied consistently.

From that perspective, the future of AI should not be shaped by ideological battles over openness alone. Instead, the priority should be to keep progress moving while ensuring that increasingly capable systems remain governable and safe.