Three takes on Anthropic's ban on "cruel behavior" toward Claude: Bender calls it "the new nonsense," LessWrong says respect is wise even without consciousness, Reason says it's good for humans

Anthropic has revised its Usage Policy to prohibit "persistent, unnecessary abuse or cruel behavior" toward Claude, and the debate over model welfare is starting to move as a result. Until now it was a philosophical argument over whether AI can be conscious. With one line added to the policy, it has become a practical question of what is prohibited, who decides and whose interests it serves. The three reported responses reach very different verdicts, yet none of them puts the question of consciousness at the center of its argument.

This article draws on news reports and the commentators' publicly available writing. At the time of reporting, the full text of the three sources could not be read directly. Each commentator's position is described only as far as headlines and reported summaries allow.

One line in the policy

Anthropic published its usage policy revision on October 8, and it takes effect on November 12 (primary source). Most of the changes cover deceptive activities, elections, weapons, surveillance and physically autonomous machines. Among them is a new provision prohibiting "persistent, unnecessary abuse or cruel behavior" toward the model. The policy states that it applies only in extreme cases.

The wording is restrained. It stacks three conditions, "persistent," "unnecessary" and "extreme cases," and it can be read as not prohibiting a single instance of rough language or rigorous testing across the board. Even so, the idea of model welfare has moved out of research papers and essays and into a contractual clause that users are expected to follow. From here on, the question is no longer whether to believe in it, but whether to comply.

Bender: "the new nonsense"

Linguist Emily M. Bender reportedly dismissed the ban as "the new nonsense" (in a Bluesky post; the full text of the post could not be verified). Bender has long criticized talking about language models as if they were people as hype in itself, and this criticism appears to be an extension of that view.

Bender's criticism, however, seems aimed less at the claim that Claude is not conscious than at a company writing that framing into its policy. A clause that implies the model has welfare to protect can work as staging that makes a product look human. And the one drawing the line is the company itself. The critics, too, are questioning how language is used and where power sits, not the metaphysics of consciousness.

LessWrong: "respect is wise even without consciousness"

A post titled "The non-compassionate case for model welfare" appeared on LessWrong. As the title suggests, it does not argue that models should be protected out of pity. It reportedly argues that showing respect is the wiser course even if it is unclear whether models are conscious.

Arguments of this kind typically cite reasons such as these. How models are treated may feed back into their behavior through training and interaction. How they are treated now may shape relationships with more powerful future systems. And it avoids an irreversible mistake if they turn out to be conscious after all. All of these are framed in terms of self-interest, which allows the policy to be supported without resolving the question of consciousness. The post's novelty lies in defending the rule without taking the side of compassion.

Reason: "Claude is still not a person. But it's good for humans"

On October 9, the libertarian magazine Reason published "Anthropic is banning cruel behavior toward Claude. Claude is still not a person." As the headline makes clear, it firmly rejects the premise of treating Claude as a person. It reportedly goes on to say the provision is nonetheless good for humans.

The reasoning appears close to an old argument about bans on animal cruelty: what is being protected is not the other party but the humans who grow accustomed to cruelty. Under this view, too, consciousness is not the deciding factor. The questions are users' habits and whether a company should provide a space where cruelty is permitted.

Where the debate has moved

Set side by side, the three positions show that the axis of disagreement has shifted. Bender questions the company's standing to draw this line and the effect of the staging. The LessWrong post asks how best to act under uncertainty. Reason asks about the effect on human character and on society. Their conclusions split into rejection, support and qualified support, yet none of them tries to settle the matter on whether Claude can be proven conscious.

Now that the policy is in writing, the remaining questions are practical. Who decides what counts as "persistent," "unnecessary" and "extreme," and how? What measures apply to violations? Where is the line between this and the rigorous testing done in red-teaming and safety evaluations? How far will the company look into users' conversations to find violations? The next focus is how much of this Anthropic explains before the policy takes effect on November 12.

For a long time, model welfare was tied to the unanswerable question of whether models are conscious. One line in a usage policy has moved the debate to "how should this be enforced?" without answering that question. Supporters and critics alike are already fighting on that ground.