Amodei clarifies: Anthropic opposes open-weight bans, not open weights

Anthropic CEO Dario Amodei published a blog post this week clarifying his position on open-weight AI models: he sees them as risky when they reach certain capability thresholds, but Anthropic has "never advocated for a ban." The distinction matters because critics have accused him of talking down open-source releases to protect Anthropic's business from cheaper competitors.

"Models that don't have dangerous capabilities are a public good," Amodei wrote. He also dismissed the idea of a US ban on Chinese models as ineffective, since "bad actors are unlikely to be legitimate US businesses." That line reads as a direct response to reports that US authorities are weighing restrictions on American companies using Chinese open-weight models.

What Amodei actually wants

The blog post outlines three concrete policy goals. First, stricter chip export controls so China cannot train models more powerful than US labs without US hardware. Second, action against industrial distillation, a technique China reportedly uses to train capable models by learning from the outputs of American frontier systems. Third, mandatory safety tests for all powerful models, regardless of origin or license.

On safety testing, Amodei sees momentum. He points to recent proposals that would exempt smaller models from startups and researchers, suggesting a growing consensus that regulation should target capability, not licensing structure.

Critics push back on each point. Cutting off chip sales to China, they argue, mainly accelerates China's domestic chip industry. Distillation is a legitimate technique used throughout the field. And Anthropic's repeated warnings about its own models' capabilities have already helped justify a US ban, making the "I never called for a ban" framing sound like a technicality.

The two nightmare scenarios

Amodei laid out two specific fears. The first: authoritarian states, especially China, could build more powerful models than the US and use them for military superiority or mass surveillance. The second: sufficiently capable open models could be misused for cyberattacks or bioweapon development.

He argues open weights are riskier than closed APIs because they lack safeguards and cannot be recalled once released. A study by the UK AI Security Institute found open-weight models trail frontier models by only a few months on cybersecurity benchmarks. That gap is closing.

“Sufficiently capable models could weaponize viruses quickly while building defenses would take years even in the best case.”

— Dario Amodei

Amodei acknowledges that open models also empower defenders, but he sees a strong asymmetry on biological threats specifically. Attack timelines compress faster than defense timelines.

The business conflict no one can ignore

Anthropic charges per token for Claude API access. Meta, DeepSeek and other labs release capable models for free. The conflict of interest is obvious, and critics have not been shy about pointing it out.

OpenAI signed a recent letter opposing US restrictions on Chinese open-weight models, but reportedly continues to lobby against open weights internally. Anthropic did not sign. Amodei's blog post reads partly as an attempt to separate two positions: "regulate dangerous capabilities" versus "ban open-source licensing." Whether the distinction holds depends on where the capability threshold lands.

What happens next

The blog post lands at a specific policy moment. US authorities are actively considering restrictions on Chinese AI models, and major tech companies are lobbying on both sides. Amodei's argument is that capability thresholds, not open versus closed licensing, should determine what gets regulated. The next moves will come from policymakers, not labs.

Source: The Decoder / Matthias Bastian