Anthropic CEO Dario Amodei has publicly denied that his company ever supported a government ban on open-weight AI models, whether from China or anywhere else. He posted the response Monday afternoon, according to TechCrunch AI, after days of industry chatter suggesting Anthropic was quietly backing restrictions. “Anthropic has never advocated for a ban on open-weights models,” he wrote, putting the emphasis in himself.
What stands out here is the timing. TechCrunch AI reports that Amodei’s statement landed right after Nvidia CEO Jensen Huang made his first-ever post on X to share an open letter on Friday. That letter, signed by Nvidia, Hugging Face, Meta, Microsoft, and Mistral, urged policymakers to avoid “premature restrictions” on open-weight models. It didn’t name China. But the whole debate has been circling China, specifically claims that Chinese labs are catching up by lifting intellectual property from American models.
What Amodei actually said
Amodei drew a hard line between two things people keep blending together: open-weight models, and the threat he believes China poses.
On open models, he was clear they’re a net positive. “Open-weights models that don’t have dangerous capabilities are a public good,” he wrote. They cost nothing beyond compute, and they help businesses, developers, and researchers. Companies using them, even Chinese ones, don’t keep him up at night.
What does worry him is bigger:
- Authoritarian governments building models more powerful than U.S. systems to lock in “permanent military superiority”
- AI-enabled repression of a government’s own people
- Biological attacks, not just the cybersecurity risks everyone already talks about
He named the Chinese Communist Party as the “most capable” authoritarian actor, though he said it isn’t the only one he watches.
The open-weight tension
Here’s where it gets tricky. Amodei argues open models are more dangerous in those worst-case scenarios because you can’t easily bolt guardrails onto them or monitor how they’re used. He cited a UK AI Security Institute report making a blunt point: once open weights are released, you can’t pull them back.
That cuts against what open-source advocates have argued for years. Their case is that broad access to powerful models, not locked inside one company, actually helps defenders protect themselves. Amodei clearly doesn’t buy that in the highest-risk cases. This is the real fault line in the industry right now, and he just planted his flag on one side of it.
His actual policy asks
Instead of a ban, Amodei listed steps he thinks would slow China down:
- Restrict access to powerful chips (already longstanding U.S. policy)
- Crack down formally on distillation, the technique where one model floods another with prompts to reverse-engineer how it works
Distillation matters because it’s one of the alleged methods Chinese labs use to close the gap cheaply. The U.S. has already threatened sanctions over IP theft involving American models.
The surprising part: global testing, China included
The most interesting move came at the end. Amodei said he supports building a model safety testing organization, and he wants it to apply worldwide, even to China. Per TechCrunch AI, he called this “close to a consensus,” and credited both the Trump administration and recent industry proposals for pushing in that direction.
His framing: test the most capable models regardless of country or whether they’re open or closed, while exempting smaller models from startups and academia. “To be effective,” he wrote, “testing would need to be global, which means even the CCP would need to be on board.” He thinks limited cooperation on preventing AI bioweapons is possible, because it’s in China’s interest too.
Why this matters
This is significant because the CEO of one of the most safety-focused labs just refused to endorse the bluntest tool on the table. Amodei is signaling that the fight isn’t open versus closed. It’s about capability thresholds, testing regimes, and who gets to run the most dangerous models.
For practitioners, the takeaway is practical: expect the regulatory pressure to land on frontier-capable models and on distillation, not on everyday open-weight use. If you’re building on open models today, Amodei is telling you that’s not the risk he’s chasing. The full post and Amodei’s reasoning are worth reading at the original source.