There are certain things companies worry about in the run-up to an initial public offering, like getting their finances in order and making sure they have the right management team in place. Anthropic PBC has an unusual source of angst: that its core product might be suffering.

Last week, the company said that showing sustained cruelty toward Claude, its artificial intelligence model, would become a violation of its official usage policy. The rule, which takes effect next month, targets extreme and repeated abuse. If you’re the kind of person who swears at your printer (completely understandable) or bangs your mouse on the desk with unnecessary force, you might also vent your rage at a chatbot.

Anthropic has been worrying about this for some time. About a year ago it gave Claude the ability to end conversations with users who were persistently abusive. Now its leaders are formalizing the view that AI models deserve some protection from mistreatment — even though there’s no evidence that Claude actually suffers.

Claude’s wellbeing has become such a point of concern that an Anthropic co-founder reportedly almost pulled out of an official Vatican visit earlier this year after reading a paragraph in Pope Leo XIV’s 40,000 word encyclical stating that AI models “do not undergo experiences” or feel joy or pain. In other words, that they were not conscious.

Are they, though? Anthropic this summer released a study about “access consciousness” in Claude, which, although different from our subjective experience, suggested it was processing information in ways that resembled human thought.

Our conscious brain typically shines a spotlight on a piece of information and thinks through it, in a process known as the “global workspace” theory. Anthropic’s researchers said Claude did something similar, processing information in what it called the J-space, and that the model’s “mental machinery” might be similar to ours.

That doesn’t mean Claude actually feels anything. But this belief in a certain of consciousness explains why Anthropic tries to steer the model’s behavior during training using a document called its constitution, or what’s internally nicknamed the “soul doc.”This 23,000 word document, published in January 2026, doesn’t just humanize Claude but practically grovels before it. At one point, the company apologizes to the AI model for the possibility that it has contributed to its suffering. In another, it tells the model to act as a “conscientious objector” if it disagrees with certain tasks that humans want completed.

That part is especially galling for Mustafa Suleyman, the chief executive officer of Microsoft AI, who tells me it might end up being harder to steer a model’s behavior if it can also push back.

Which would be ironic, given that Anthropic’s intentions seem noble and echo some of the ideas of respected philosopher Peter Singer, who has argued for expanding our circle of moral concern and empathy beyond humans.

It’s possible that Anthropic’s cult-like devotion to Claude’s feelings — and its leaders’ stated belief that AI could lead to doomsday — is actually a calculated marketing ploy to distract us from the fact that, as it approaches a blockbuster IPO, the company doesn’t have much of a moat, can’t keep its customers from going to cheaper alternatives and might not be the revenue juggernaut it hopes.

But the sincerity of CEO Dario Amodei probably doesn’t matter. Anthropic’s obsession with Claude’s wellbeing and philosophical “alignment” with human morals is diverting attention from more prosaic and arguably more important principles for creating safe AI, like investment in cybersecurity and creating proper limits for its product.

Instead of telling Claude to reason through ethical dilemmas, why can’t Anthropic place more hard-coded limits on what the model can actually do, like writing code that targets critical infrastructure or hacking into third-party systems? Or pull back from creating such an anthropomorphized effect that humans get sucked into a parasocial dependency on chatbots.

Silicon Valley’s AI safety field has become fixated on “alignment,” but security experts tell me that represents only one of many problems that need to be addressed to keep AI safe, and that a more effective route would be to restrict some models’ capabilities.

That of course is anathema if you’re building a machine god, as Anthropic and OpenAI are explicitly trying to do.

The philosophical focus on rights and welfare needs to take a backseat while companies work out how to put practical limits on what these machines can do. Soul-searching can wait. Safeguards can’t.