AIANALYSIS
Dario Amodei reframes the AI fight around which models can do harm and who builds them, not whether the models are released for anyone to download.
Anthropic CEO Dario Amodei published a long essay on the company's site this week denying a charge that has followed him for months: that Anthropic wants open-weights AI banned in the United States to protect its own closed-model business.
"Open-weights" means an AI model whose trained parameters, the numbers that determine what the model knows and how it responds, are released publicly for anyone to download, as opposed to a closed model accessed only through a paid API. "Open-source" is a related but technically distinct term; Anthropic's post uses "open-weights" pointedly, and the distinction is doing work in the argument that follows.
"Anthropic has never advocated for a ban on open-weights models," Amodei writes, in a post he says he is publishing in response to "speculation and accusations" he has fielded repeatedly. The denial is categorical. The position he lays out in its place is narrower and more specific.
Amodei frames his concerns as two "nightmare scenarios" rather than a blanket openness critique. The first is authoritarian AI: a rival government, "especially the CCP" (the Chinese Communist Party), building models more capable than anything the United States fields and using them for "permanent military superiority or deep repression." The second is powerful-model misuse at home and abroad: cyberattacks, biological attacks, and "serious alignment problems" in systems that do what their builders did not intend.
He is careful to carve out a category he does not object to. "Open-weights models that don't have dangerous capabilities" are "a public good," he writes, and banning them would do nothing to address either of his two scenarios. The line he is drawing is capability-tier, not openness-binary. The fight, on his account, is over which models, not whether models are open.
The argument is grounded in on-record US government language rather than Amodei's own forecasting. He cites Vice President Vance's Paris speech, in which Vance said "authoritarian regimes have stolen and used AI to strengthen their military, intelligence, and surveillance capabilities," and the Intelligence Community's 2026 Annual Threat Assessment, which he treats as evidence the national-security concerns he describes are shared inside the US government.
Amodei adds that it is "irrelevant" whether an authoritarian-built model is released with open weights, because the most dangerous case may be a secretly trained model handed to the People's Liberation Army (China's military) or the Ministry of State Security (China's civilian intelligence agency). And he says protectionist bans "would not address" his most serious concerns.
The essay is also a defense against a more specific criticism: that Anthropic's sustained safety advocacy reads as de-facto openness skepticism, even if the company has never asked for an outright ban. Critics in the thread under Anthropic's X post linking to the essay drew significant attention, with critics disputing the framing.
The contradiction is real, and the post does not resolve it. Anthropic sells a closed-weights product. Its CEO is now telling the public that open-weights bans are bad policy and that the AI risks worth restricting live at the capability frontier, a frontier his own company is best positioned to define. The post leaves open whether that is principled boundary-drawing or strategic framing.
Amodei grounds the essay in his prior piece, "The Adolescence of Technology," published roughly six months before. The continuity claim is explicit: these are positions he has held for years, not a shift.
The post most directly addresses a live policy fight over whether US companies should be allowed to download and fine-tune Chinese open-weights models, even though it never names the policy. If the capability-tier frame wins, restrictions land on a small set of frontier systems. If it loses, the openness binary takes over and the restrictions are broader. Amodei is asking Washington to pick the first outcome. The critics in his replies are asking the public to notice which side that puts Anthropic on.