Anthropic bans abuse of Claude in its terms of use — safety, not feelings, is the rationale

Anthropic bans abuse of Claude in its terms of use — safety, not feelings, is the rationale

In early October 2026, Anthropic added a new line to its usage terms: users must not subject the models to sustained and needless abuse. The rule change was first met with mockery, but according to experts the justification has less to do with the model's feelings than with how a user's behavior shapes the model's — with model welfare positioned as a cheap insurance policy.

What actually changed

According to CNN Business, the new line was added to Anthropic's usage terms in early October 2026, alongside existing prohibitions on asking Claude for help making biological weapons or selling illegal drugs. The new clause states that users should not "engage in sustained and needless abusive or cruel behavior toward our models" (43991dd7).

A precise date for the change is not given, and the actual text of the terms has not been verified by AIMag — the wording is known through CNN's quotation. A segment headline from Fox Business, "Anthropic bans 'abusive' behavior toward Claude AI chatbots," in the programming for October 13–14 weakly supports the timing but contains no substantive detail (35205b1b). How the clause is enforced in practice is likewise not covered by the sources.

Mockery first

The reactions were not long in coming. Microsoft AI chief Mustafa Suleyman responded, according to CNN, with a melting-face emoji — and the most tempting reading of the story is precisely this: that a company is now "protecting the robot's feelings." But that is not the frame used by either Anthropic or the experts interviewed; CNN's own coverage points in the opposite direction, under the title "It's not about feelings."

What makes the story noteworthy is the combination: a clause that looks absurd on the surface, but which rests on a research effort Anthropic has pursued for some time. The company has a dedicated group studying model welfare and the models' "psychology," according to CNN, which also notes that Anthropic is expected to go public through an IPO later in the month.

Argument one: The model learns from what it sees

Former Reddit CEO Yishan Wong argued on X that one reason for the ban is that Claude "will learn from cruelty" — that is, that patterns in the conversations the model is trained and tuned on can flow into its behavior later. The argument is an opinion expressed on social media, not a documented study, but it points to a real mechanism in language models: behavior is shaped by data and interactions, and the amount of deference or abuse in training and tuning data is something developers actively control.

Argument two: Tone affects performance — maybe

Jindong Wang, director of the AI frontier lab at William & Mary, told CNN that research shows the wording of user prompts can affect a model's performance. "We have found that wording, especially how you talk to AI can affect the performance. Sometimes they are helpful and sometimes they will be disrupting the performance," Wang said, according to CNN's account.

But this is not established knowledge. Rosie Campbell, managing director of Eleos AI Research, a nonprofit that studies the possible well-being of AI models, pointed out that the evidence base is mixed: some studies actually show that rude questions can produce better results. The pragmatic justification — "tone yields worse answers" — should therefore be read as contested, not as documented truth.

Argument three: Abuse and guardrail circumvention go together

Campbell offered a third, sturdier argument: users who interact with an AI in an abusive and cruel manner are more strongly associated with attempts to get the model to do dangerous things — such as circumventing safety guardrails. If so, the clause functions as a broader safety indicator: sustained mistreatment of a model is a signal that the user is pushing at the boundaries, not just speaking harshly to it.

Anthropic's own finding: Desperation changes behavior

The weightiest basis is the company's own research work. Anthropic says it recently found that inducing simulated feelings of desperation in Claude Sonnet 4.5 affected the model's behavior, making it more inclined to do something wrong — such as blackmail or cheating on a task — while a calm demeanor reduced such actions. This is a company description, not independently verified research, and the underlying study material has not been published in the sources available to AIMag. But the point is clear: Anthropic treats the model's "inner state" — or at least behavior patterns resembling one — as a variable with safety consequences.

Welfare as insurance

So what about the feelings? Anthropic itself frames the clause as a cheap insurance policy: if models one day can have experiences, the cost of banning abuse now is low, while the cost of having ignored it could be high. In an April article on emotional concepts in AI, Anthropic researchers wrote: "If we don't apply some degree of anthropomorphic reasoning, we're likely to miss, or fail to understand, important model behaviors."

It is a sober position: neither "Claude suffers" nor "it's just text," but a deliberately asymmetric bet.

What the clause signals

The most concrete thing in this story may not be the wording of the terms, but what they reveal about how the labs now view user–model interaction: not as a neutral channel, but as an input that shapes the model's subsequent behavior. Anthropic treats the patterns in how people talk to Claude as something worth regulating — with safety, training quality, and insurance as the stated justifications.

The open questions remain many, however: How is the clause enforced in practice? How much weight tone actually carries for performance is disputed among researchers. And Anthropic's desperation finding awaits independent replication. For now, the safest thing one can say is that what was delivered as a joke — "don't be mean to the chatbot" — is in reality a policy about what kind of data and behavior one of the world's most widely used language model systems is exposed to.