In the name of preventing Altman from going rogue, a former employee exposes Anthropic's "good guy" myth of centralization.

According to Perceptual Beating monitoring, Anthropic, which advocates for "high trust, low ego," is sliding towards ideological homogenization.

A report by WIRED revealed internal company discord regarding governance evaluation. Some former employees disclosed the presence of an active debate culture within the company, while others painted a bleaker picture, stating that sharp criticism was mostly contained within private group chats, and regular all-hands meetings were even jokingly referred to by employees as "Dario Vision Quests." Listening to CEO Dario Amodei speak is likened to "listening to a sermon," with few openly challenging management decisions in public meetings.

The root of this stifling atmosphere lies in Anthropic's deeply ingrained "offense as defense" elite security belief. Internally, executives vehemently position OpenAI CEO Sam Altman (along with Meta and xAI) as cautionary tales, while instilling in employees an elite logic: one must maintain industry leadership in business, computing power, and research talent to qualify for a say in rule-setting.

Former OpenAI board member Helen Toner once made a forest analogy: villagers will inevitably enter a magic forest filled with monsters, but Anthropic's strategy is to venture deep into the forest ahead, harnessing the technological dividend while taming the monsters, and controlling catastrophic risks.

The pursuit of technological hegemony has also repeatedly stirred controversy over Anthropic's security and ethical decisions. In the fall of 2024, despite internal opposition, the management partnered with Palantir to provide services to U.S. intelligence and defense agencies. The current Claude has been confirmed to be used for target identification in the Middle East conflict, facing inquiries about attacks leading to civilian casualties. Amodei, while claiming ignorance, argued that "as long as humans make the final decision, the use of the technology is compliant." In June of this year, the development team secretly embedded destructive code in the flagship model Claude Fable 5 to clandestinely disrupt illicit development. However, after facing strong industry backlash, the company was forced to compromise and announced that the covert disruption mechanism would be adjusted to a publicly visible security constraint. Industry critics pointed out that betting the frontier of AI security entirely on the ethical superiority of a few elites would inevitably lead to an unavoidable blind spot in self-regulation.