Дарио Амодеи (Anthropic): почему мы не выступаем за запрет open-weights моделей

On July 27, 2026, Dario Amodei, CEO of Anthropic, published a statement on the company blog: Anthropic has never supported a ban on open-weights models, including Chinese ones, and considers such bans useless for US national security.

Why the dispute arose

The dispute flared up over reports that US officials are considering banning American companies from using Chinese open-weights models. In response, many tech companies signed a letter in support of open weights, and some commentators accused Anthropic of wanting to lobby for a ban to protect its own business. Amodei directly rejected this.

• Date of the statement — July 27, 2026, author — Dario Amodei, CEO of Anthropic
• Occasion — reports of a possible ban on Chinese open-weights models for US companies
• Many tech companies signed an open letter in support of open-weights models
• Anthropic was accused of lobbying for a ban to protect its business — Amodei rejected this
• He laid out his position earlier in the essay "The Adolescence of Technology" (January 2026)

According to Amodei, open-weights models of safe capabilities are a public good: they cost nothing beyond the compute to run, and they benefit businesses, developers, and researchers.

What Anthropic is actually afraid of

Amodei named two "nightmare scenarios" that protectionist bans do not solve. The first and main one is the risk that authoritarian governments, above all the Chinese Communist Party (CCP), will build models more powerful than American ones and use them for lasting military superiority or deep repression of their own population.

This concern is shared within the US government. Vice President J.D. Vance warned a year earlier in Paris that "authoritarian regimes have stolen and used AI to strengthen their military, intelligence, and surveillance capabilities."

"Other world powers are showing steady progress in AI, which challenges US economic competitiveness and national security advantages," the US intelligence community's

Annual Threat Assessment 2026 states.

The second scenario is the misuse of powerful models for cyberattacks or biological attacks, and alignment problems. Here open weights really are riskier than closed ones: it is hard to apply protective restrictions to them, and once released, weights cannot be recalled.

What Anthropic proposes instead of a ban

Instead of bans, Amodei backed three measures that Anthropic has consistently advocated for. Not selling China powerful chips and the equipment to produce them, and also strictly cracking down on mass smuggling and evasion schemes. Because of scaling laws, China without American chips will not be able to build models more powerful than the US's. By Amodei's logic, export controls hit the first threat directly and indirectly complicate the second, slowing down model training outside US jurisdiction.

The third measure is fighting industrial distillation. Distillation is far more compute-efficient than training from scratch, and it lets China partially bypass chip restrictions, obtaining models better than its chip count would otherwise allow.

"Anthropic has never supported a ban on open-weights models," —

Dario Amodei, CEO of Anthropic, on the company blog.

What this means

Anthropic has publicly distanced itself from the idea of banning open models and shifted the national-security conversation to chip export controls and distillation — that is, to what actually limits rivals' capabilities, rather than shielding American companies from competition. The debate over banning specifically Chinese models for American business is, in effect, being dismissed as a false target.

Frequently asked questions

Does Anthropic support banning open-weights models?

No. Dario Amodei stated directly that the company has never proposed such a ban and considers it a useless measure for national security.

What measures does Anthropic support?

Three: not selling China powerful chips and equipment, cracking down on chip smuggling, and fighting industrial distillation of models.