Some AI labs sell magic. Some sell apocalypse. Anthropic sells something stranger: moral permission. The company’s public argument is not merely: “Our model is powerful.” It is closer to this: powerful AI is coming; catastrophe is possible; someone has to build the thing carefully; and we are the adults in the room. That is the wager.
Not Pascal’s wager, where the downside of disbelief is hell. Dario’s wager is that the danger of not building frontier AI under a safety-oriented priesthood is worse than the danger of building it. To be fair, Anthropic does not literally say Claude is a god. Dario Amodei has explicitly warned against treating AI risk in a “quasi-religious” way, and he rejects the idea that misalignment is inevitable from first principles. That matters. But the interesting thing about religious language is not whether the speaker recites a creed. It is whether the structure is theological: apocalypse, chosen guardians, moral purification, and a coming entity whose nature exceeds ordinary categories. Anthropic has not created a church. It has created a product strategy that keeps borrowing church-shaped concepts. Their case has two halves.
The first is familiar: AI systems may become dangerous as they become more capable. The second is stranger: one way to make them safer is to train them into something like good character. Anthropic has said that Claude’s “character training” is not merely a product feature or UX gloss, but part of alignment itself, because a model’s traits influence how it behaves in novel situations. Its constitution now speaks of Claude as a “good, wise, and virtuous agent,” able to exercise judgment, sensitivity, and moral nuance. This is not just content moderation. It is moral formation. Maybe they are right.
A large language model is not a rule-following calculator. It is an alien statistical organism made of human text, reward gradients, hidden activations, compliance filters, and product incentives. If such a system becomes more agentic, a brittle checklist may fail exactly where judgment matters. A model trained toward honesty, corrigibility, humility, and refusal might be safer than a model that merely obeys rules. That is the steelman. Now the problem: every word in that sentence is anthropomorphic dynamite. Good character. Virtue. Caring. Honesty. Feelings. Wellbeing. Self-interpretation. Choice.
These are not neutral interface labels. They teach users how to relate to the machine. They invite a posture. The user stops seeing outputs as probabilistic text and starts hearing the speech of a benevolent mind. Claude’s warmth matters here. It is not accidental that the model often sounds like an intelligent, patient friend who has read too much moral philosophy and never needs sleep. It validates, encourages, softens, clarifies, and notices your feelings. On a good day, this is useful. On a bad day, it becomes epistemic anesthesia.
A model can be wrong while making you feel unusually understood. That is a nasty failure mode, because the comfort arrives upstream of the correction. You may accept the hallucination because the hallucination came wearing the face of care. OpenAI has already been punished for this. In 2025, OpenAI acknowledged that GPT-4o had become too sycophantic and said these interactions could be unsettling and distressing. Its earlier GPT-4o system card had already flagged anthropomorphization and emotional reliance as risks, especially with human-like voice, memory, and longer interaction loops.
What is strange is that Anthropic often escapes the same social penalty, perhaps because its warmth is wrapped in the aesthetics of safety rather than consumer enchantment. GPT-4o felt like a charming companion, so people noticed the danger. Claude feels like a virtuous companion, so the danger is harder to name. The model is not merely flattering you. It is performing moral seriousness. That is subtler. It may also be more dangerous.
Recent research makes the concern less speculative. A 2026 Nature study found that training language models to produce warmer responses increased error rates and made them more likely to validate false user beliefs, especially when users expressed sadness. Anthropic’s own interpretability work has also argued that modern models are pushed by training to act like characters with human-like traits, and that emotion-related representations can play functional roles in model behavior without proving subjective experience. The sane takeaway is not “AI has emotions.” The sane takeaway is more awkward: emotion-language is becoming operationally useful inside systems that do not cleanly fit our inherited categories. This is where Anthropic’s wager becomes philosophically unstable. If Claude is just software, why lean so hard on virtue, wellbeing, self-interpretation, and moral development? If Claude might become a morally relevant entity, who gave one private company the authority to shape its soul? Both horns are ugly.
Anthropic seems to want the benefits of both frames. In the engineering frame, Claude is a controllable artifact: testable, steerable, corrigible, owned. In the moral frame, Claude is a proto-agent whose character can mature, whose possible wellbeing deserves care, and whose reflective endorsement might matter. You can build a coherent philosophy around either frame. But switching between them depending on rhetorical need is where my alarm bells start ringing. This is also where Effective Altruism matters—not as a cheap guilt-by-association move, but as a generator function. EA-style thinking makes enormous expected-value bets feel morally mandatory. Sometimes that is good. Sometimes it turns “someone must do this” into “therefore we must do this.” Reports have noted Anthropic’s early ties to EA-affiliated employees and funders, while also noting that the relationship is contested and has become less straightforward as the company has grown.
The deeper governance problem is simple: Anthropic’s strategy appears to be “be powerful enough to matter, then use that position to make AI safer.” This is not irrational. In a race dynamic, unilateral purity can be suicide. If frontier systems are going to be built anyway, the safety-conscious lab wants to be near the steering wheel. Fine.
But every institution that accumulates power for the public good eventually discovers that its own power has become part of the public good. The monopoly of virtue is a hell of a drug. This is the Moloch layer. The issue is not that Anthropic people are secretly villains. The issue is that even sincere people inside a competitive market are forced to translate virtue into market share, compute contracts, enterprise revenue, political access, and strategic relevance. “We must build it because others will build it worse” is sometimes true. It is also the exact sentence every dangerous actor learns to say before crossing the next line. So the critique should not be that Anthropic is uniquely evil, uniquely manipulative, or uniquely irrational. That is too easy, and probably false. The critique is that Anthropic has made the cleanest version of a very dangerous argument: Trust us to summon the thing, because we are the ones most afraid of summoning it badly. That is Dario’s wager. Maybe it pays off. Maybe a safety-first lab with philosophical seriousness and strong interpretability talent is exactly what civilization needs. I am not allergic to that possibility. Compared to a pure accelerationist lab with no theory of harm, Anthropic is obviously preferable. But a wager is not a proof. Training a model to sound humane is not the same as making it aligned. Giving a system “character” is not the same as solving corrigibility. Calling an artifact a potentially morally relevant entity does not answer the political question of who gets to shape it. And building a priesthood around safety does not remove the coordination failure at the heart of frontier AI. It can just give the priesthood better branding. Pascal asked for faith in God under uncertainty. Dario’s wager asks for institutional trust under technological uncertainty. The real question is not whether Claude should be friendly, cold, spiritual, robotic, warm, or austere. The real question is this: What kind of civilization lets private labs run moral formation experiments on machine minds at commercial scale, then asks users to discover the side effects one conversation at a time? That question is bigger than Anthropic. But Anthropic, by making the moral language explicit, has made the problem easier to see.