Zvi Mowshowitz, who writes about AI policy on his blog Don’t Worry About the Vase, argues in a 27 October essay that Anthropic is really two companies sharing a name. Where CEO Dario Amodei cares a great deal, he writes, Anthropic makes big sacrifices and takes principled stands. Where Amodei is less concerned, it slides toward being an ordinary business. Everything below is Mowshowitz’s judgment, offered by an outsider who admires parts of the company and thinks the whole falls short.
The essay grew out of a fight over Anthropic’s leaked draft federal bill, which Mowshowitz says would have overridden a lot of state law. Nathan Calvin, whose investigative work on AI companies Amodei praised, listed Anthropic moves that sit badly with its safety pitch, including a rewritten safety policy that dropped unilateral promises to slow down. Amodei replied at length with nine costly choices of his own. Among them: holding Mythos-class models back from general release between February and June, bioweapon filters strict enough that biologists joke about them, and refusing to sell to companies controlled from China. Mowshowitz calls the reply “amazingly great” for its openness and says he does not see “nefarious schemes” in any of the nine.
His argument is about where those sacrifices land. Mowshowitz quotes the journalist Kelsey Piper, who sat through a briefing at Anthropic, saying the costly choices mostly serve causes Amodei is personally driven by, such as China and biosecurity. It is far less clear, she wrote, which ones were made to keep humanity from losing control to AI. Mowshowitz adds that Amodei does not seem to think alignment is especially hard, while the longtime technical staff worry more. Newer hires, in his experience, often talk as though good operations will be enough.
Mowshowitz is blunter about the policy team. He says it has long failed to reflect what technical staff and much of leadership want, and has “gone rogue”, reading its job as protecting Anthropic’s freedom of action ahead of safety. Amodei’s own account is that the bill’s rollout was a communication slip in an organization of more than 5,000 people, and that he takes responsibility. A commenter Mowshowitz quotes, dave kasten, asks where the postmortem is that would stop a repeat.
The sharpest anger is saved for Anthropic’s yearly Public Benefit Report. Mowshowitz says the layout leads a skimming reader to conclude the technical problems of safety are solved, and that only much smaller print admits alignment is far from solved. His words: “I refuse to be gaslit.”
He gives credit too. Anthropic is committing 5 percent of its shares to a new Public Benefit Fund, and the board may add 2 percent more each time valuation climbs another 50 percent past a baseline Anthropic has set. And it has given the evaluation group METR employee-like access. Anthropic cannot veto what METR publishes, only require redaction of narrow categories such as customer data and model architecture. METR, after reading the unredacted risk report, agreed that risk is low but said it is not confident the report’s methods would catch a higher risk. Mowshowitz says he broadly agrees.
Two cautions apply. This is one commentator’s reading, and his claims about Anthropic’s insular culture rest on conversations with people he does not name. Anthropic has not publicly answered them in the pieces Mowshowitz quotes. The fair test is whether the S-1 filing he covers next, now with the SEC, shows the same split.
For anyone building on Claude, the practical question is which Anthropic shows up when a decision costs money and nobody at the top is personally invested. Mowshowitz’s answer is that you cannot count on the first one.
Reported by Don’t Worry About the Vase (Zvi Mowshowitz) on 27 October 2026.
