Anthropic CEO: no ban on open AI models, but chip exports and distillation need limits

On July 27, Dario Amodei, CEO of Anthropic, released a statement responding to criticism of his company's decision not to sign a joint statement issued by Microsoft, NVIDIA, and other technology companies on July 24. That joint statement opposed what the signatories characterized as excessive regulation of open-weight AI models (models whose code and weights are publicly available). Amodei's statement emphasized that Anthropic has never advocated for banning open-weight models as a category. However, his position diverges from the joint statement's framing of open models as broadly beneficial.

Amodei identified what he called "the most serious national security concerns": the risk that authoritarian regimes, particularly the Chinese Communist Party, will develop more powerful AI models than the United States and use them for military purposes, and the risk that open-weight models will be misused for cyberattacks or bioweapon development. Regarding the joint statement's core argument—that open-weight models make safety features easier to develop and give cybersecurity defenders an advantage—Amodei stated he could not agree with that framing.

Instead of an outright ban on open models, Amodei proposed three specific measures. First, he called for restricting the export of U.S.-made AI chips to China, arguing that limiting Chinese access to these chips would slow the development of advanced AI models within China and reduce both military and dual-use risks. Second, he advocated for cracking down on "industrial-scale distillation," a technique in which the outputs of a high-performance AI model are used to train another model efficiently. Amodei noted that many Chinese companies use distillation to develop models with fewer computational resources, and while many of these companies publicly release open-weight models, he assessed that dual-use risks outweigh the importance of open-weighting in that context. Third, he proposed mandating safety testing on all sufficiently advanced AI models—open or closed—to evaluate their potential for being misused in cyberattacks or biological weapons development.

Amodei concluded by restating his position: "I and Anthropic's stance is that we have never advocated for a ban on the open-weight model category from the past to the present. Rather, we should focus on ensuring that powerful chips do not fall into the hands of authoritarian forces, preventing industrial-scale distillation, and requiring safety testing on all models with sufficient capability, regardless of whether they are open or closed."

Amodei's July 27 statement attempts to reposition Anthropic's stance in a debate over how strictly to regulate open-source AI models. On July 24, Microsoft, NVIDIA, and other tech companies released a joint declaration opposing excessive regulation of open-weight models, framing such restrictions as economically and innovatively harmful. Anthropic's refusal to sign prompted critics to suggest the company was seeking a de facto ban to protect its own closed, proprietary models from competition. Amodei's response concedes that he agrees with "most" of the joint statement but draws a sharper distinction between what he opposes and what he supports: rather than opposing open models as a category, he identifies China's potential military advantage in AI and the dual-use risks of open models (cyberattacks, bioweapons) as the actual national security concern. His three-point framework—chip export controls, distillation restrictions, and universal safety testing—attempts to address those risks without calling for an outright ban, thereby distancing Anthropic from the accusation of competitive self-interest.

No comments yet. Be the first to share your thoughts!