Our position on open-weights models
Anthropic ● Covered by 75 sources
Dario Amodei says Anthropic never asked for a ban on open-weights AI models, despite rumors flying around Washington. He wants chip controls and safety testing instead, not a blanket ban that punishes open models specifically.
Based on reporting by Anthropic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Dario Amodei spent a chunk of this week doing damage control. After reports surfaced that US officials might block American companies from using Chinese open-weights models, tech firms fired off a joint letter defending open weights, and some of them pointed fingers at Anthropic, accusing the company of quietly lobbying for a ban to protect its own business. Amodei's response, posted under his own name, is essentially: that's not us, and it never was.
His argument rests on separating two very different worries that keep getting mashed together in the public debate. The first is geopolitical: authoritarian governments, the CCP chief among them, building AI systems powerful enough to lock in military dominance or supercharge domestic surveillance. Amodei points to Vice President Vance's Paris remarks and the Intelligence Community's 2026 threat assessment as evidence this isn't a fringe concern. Crucially, he says whether a model has open weights is beside the point here — the scariest version of this scenario is a model trained in secret and handed straight to the PLA or the Ministry of State Security, never released to anyone.
The second worry is narrower: misuse for cyberattacks or bioweapons, plus alignment failures, where open weights genuinely are harder to patch or monitor once they're out in the wild. But Amodei's point is that banning US companies from using such models does nothing about this, since the actors he's worried about aren't legitimate US businesses in the first place. What he'd rather see, and says he's pushed for consistently, is a three-part approach: choke off chip and chipmaking exports to China (since scaling laws mean China can't out-train the US without US-grade silicon), crack down on industrial-scale distillation operations that let Chinese labs punch above their compute weight, and impose mandatory pre-release safety testing on any sufficiently capable model, open or closed, regardless of where it came from.
He's careful to note where he agrees with the open-weights letter — access, competition, customer control, all real benefits — and where he splits from it. The letter assumes open access mostly helps defenders. Amodei isn't convinced, especially in biology, where he thinks attackers could weaponize pathogens faster than defenders could ever mount a response, Operation Warp Speed being the optimistic case. His fix isn't to guess at the answer but to test for it before release, at a global level that would need to include China too, something he thinks is more plausible than it sounds given shared incentives around bioweapons risk.
My take — AI-written commentary, not fact-checked reporting
I think Amodei's actually right that the ban-vs-open fight is a distraction dressed up as policy, and that chips plus mandatory testing is the boring, correct answer nobody wants to campaign on. But I'd also note it's awfully convenient for the guy running a closed-model company to argue the real fix is testing regimes his labs are already built to pass — watch who ends up writing the test.
Read more about this at: Anthropic