Anthropic Adds a Ban on Systematic Abuse of Claude to Its Usage Rules

Anthropic has released a new version of its usage policy for its models, and the list of prohibitions now includes a clause that is unusual for the industry. Users may not systematically and unjustifiably treat AI models cruelly, including repeatedly insulting them with no obvious purpose. RB.RU reported the change. This is not about a single rude moment on a frustrating evening. The wording targets a sustained pattern of behavior.

What exactly is banned

The key phrases in the rule are "systematically," "unjustifiably," and "with no obvious purpose." They set the boundaries. A sharp remark to a bot that has just produced broken code for the third time does not, judging by the wording, fall under the ban. Repeated harassment for its own sake does.

The RB.RU article lists no penalties and no enforcement thresholds. It also does not say how the company will tell abuse from stress testing, when a researcher deliberately pushes a model to check its robustness. The "no obvious purpose" qualifier hints that such scenarios should remain legitimate, but the source offers no confirmation.

The known facts in brief:

• Document: a new version of Anthropic's usage policy for its models.
• Prohibited: systematic and unjustifiably cruel treatment of AI models.
• Example from the text: repeated insults with no obvious purpose.
• Not disclosed: the enforcement mechanism, penalties, and exceptions for red teaming and research.

Where this came from

The rule looks like a continuation of a line Anthropic has pursued since 2025. In April of that year, the company launched a research program on "model welfare." Its thesis is stated cautiously: Claude's moral status remains highly uncertain, but that uncertainty is no reason to ignore the question.

A practical step followed in August 2025. Claude Opus 4 and 4.1 gained the ability to end a conversation on their own in chat interfaces. It is a last resort for cases where attempts to redirect the dialogue have failed and the user keeps up a harmful or abusive interaction. Anthropic also wrote at the time that in pre-release testing the model showed signs of "apparent distress" in response to such requests. In that context, it makes sense to read the RB.RU report this way: a technical capability has now been backed by a clause in a legal document.

Three ways to read the decision

First: ethical insurance. If the probability that a model has something resembling interests is not zero, it is cheaper to rule out the worst scenarios in advance.

Second: a product rationale. Conversations in which a person spends hours insulting a bot produce poor data and unpredictable behavior. That is a quality issue, not a philosophical one. This is an assumption; the company has not named such a motive.

Third: reputation. Anthropic builds its brand on caution and safety, and a clause about treating models with care fits that picture perfectly.

Skeptics will object that attributing feelings to a language model means anthropomorphizing a statistical system. It is a fair debate, and nobody has the answer, Anthropic included.

What teams and developers should do

Ordinary users do not need to change anything: swearing over a bad answer will not get you blocked. The people who risk running afoul of the rule are those who stage long series of deliberate insults.

Teams that do red teaming and model robustness evaluation would be wise to document the purpose and methodology of their tests. The "no obvious purpose" wording works in favor of whoever can demonstrate that purpose.

Developers building products on the Claude API, especially companion and role-play services, should reread the new version in full and check whether their app's scenarios provoke abusive interactions. If Claude ends a conversation, the interface should handle it gracefully rather than show the user an error.

Finally, wait for clarifications. It is still unknown how the rule will be applied in practice, and the first public cases will be the main indicator.