Anthropic revises its Usage Policy once a year. The 2026 version, announced on October 8 and taking effect on November 12, contains one line that has drawn more attention than everything around it: "a prohibition on sustained and needless abusive or cruel behavior toward our models."
Most usage policies protect people from what a model might do. This line runs the other way. I build iOS apps and this website on the side with coding agents, Claude Code among them, so the obvious question for anyone working like that is whether it changes anything. In short: hardly. The longer answer depends on what the rule actually covers, how it is enforced, and where it comes from.
What the rule covers, and what it doesn't
Anthropic frames the rule as one for extreme cases: users who are repeatedly cruel to a model with no recognizable purpose. It names what falls outside it explicitly:
- ordinary frustration,
- disagreement and pushback,
- dark themes in creative work,
- model testing and research.
That exclusion list is the part worth reading twice. Telling an agent its output is wrong again, rejecting an answer, or asking for a story full of violence is not what the rule is about. What it targets is behavior that keeps going and serves no visible goal other than cruelty itself.
Enforcement is light. The announcement says: "Claude's ability to end these interactions will remain the primary enforcement mechanism." In other words, the main tool is Claude ending such a conversation on its own, in Claude.ai and in Claude Code alike. Anthropic does not mention account bans or other penalties, and press coverage describes no enforcement beyond that. According to TechCrunch, the new rule prohibits users from deliberately continuing such conversations.
Read that way, what is new is less the mechanism than the policy around it. Claude has been able to end conversations for more than a year.
The background: Claude has been able to hang up since 2025
In August 2025, Anthropic announced that Claude Opus 4 and 4.1 could end a rare subset of conversations in its consumer chat apps. The feature was meant "for use in rare, extreme cases of persistently harmful or abusive user interactions." Claude was to use it only as a last resort, after attempts to redirect the conversation had failed, or when a user explicitly asked for it. It was not to be used where users might be at imminent risk of harming themselves or others.
The consequences were contained. The ended chat accepts no new messages, other chats are unaffected, a new chat can be started right away, and earlier messages can be edited to branch the conversation from there.
Anthropic tied the feature to its exploratory research on potential AI welfare, usually called "model welfare." The key sentence: "We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future." The company described the feature as a low-cost precaution in case that status turns out to matter, and as relevant to alignment as well.
Before deploying Claude Opus 4, Anthropic reported three observations: a strong aversion to harmful tasks, a pattern of apparent distress when real users asked for harmful content, and a tendency to end harmful conversations in simulations when given the option. The word "apparent" carries a lot of weight. The claim is not that Claude suffers, but that a pattern shows up that looks like it.
The rest of the update
The rule on treatment of models is one change among several:
- Deceptive activity: a new section covers deceptive campaigns and artificial activity, such as fake accounts, concealed senders and infrastructure for influence operations.
- Elections: the section is now called "Do Not Undermine Democratic Processes." Deceiving voters and disrupting elections are prohibited; the blanket ban on personalized voter outreach is gone.
- Weapons: the prohibition now also covers software and components that make weapons functional, and arming drones or autonomous vehicles.
- Surveillance: no tracking without consent, no surveillance tools, and Claude is not to decide or recommend whom to investigate, arrest or charge.
- Hardware: for devices that take autonomous physical actions, a qualified operator must be able to observe and stop them, and a disconnect must leave the system in a safe state.
Even so, Anthropic calls abuse toward models a main focus of the update.
The debate over consciousness
The wording of the rule itself is careful. According to AFP, the term "model welfare" does not appear in it. Anthropic CEO Dario Amodei told the New York Times in February 2026: "We don't know if the models are conscious... But we're open to the idea that it could be."
Others disagree. Microsoft's head of AI, Mustafa Suleyman, wrote in an essay in September, as AFP reports, that AI is not conscious and can neither feel nor suffer. Pope Leo XIV said in a homily at St. Peter's Basilica that machines have no soul and only process data.
AFP also quotes a third position. Jackson Stakeman of Sparq in Atlanta considers the consciousness question a dead end. AI systems reflect what people put into them, he argues, and that is reason enough for the rule.
What it means if you work with agents
For most people using Claude, the practical change is small. Venting in a prompt is still allowed, so is pushing back, and the exclusions are written so that ordinary work stays outside the rule. Rejecting a result, throwing out an approach, asking for a third rewrite: that is work on the outcome, not cruelty. And the consequence Anthropic describes is an ended conversation, not a suspended account.
My post on requirements for AI agents argued that an agent only has what is written down. Annoyance at a result is fair to express. The next result, though, is more likely to improve with a sharper requirement than with a sharper tone.
The more fundamental question is how a rule like this can be justified. A usage rule that protects a model only makes sense if you take seriously the possibility that the model could be affected. Taking it seriously is not the same as believing it; Anthropic itself says it is highly uncertain. The logic is that of a precaution: if it costs little and possible harm cannot be ruled out, it is defensible. Whether that convinces you depends on how likely you think the possibility is, and Suleyman and the Pope evidently weigh it differently from Amodei.
There is also a way to justify the rule without settling that question at all, and it is Stakeman's: systems reflect what goes into them. Taken one step further, what we practice on them may shape the systems and, possibly, us. That argument does not require believing a model feels anything.
What remains open is who decides, case by case, what counts as "needless." From what Anthropic describes, the first call is the model's, when it ends a conversation. How well that line is drawn is hard to judge from the outside.
Conclusion
The rule is narrower than the headlines suggest. It does not ban frustration, disagreement or dark fiction, only sustained, purposeless cruelty, and its main instrument is an ended conversation. It is still notable: an AI company is writing into its terms how people should treat its model, and grounding that in a question it openly leaves unanswered. Whether Claude experiences anything is contested. If Stakeman is right, the rule does not need that answer to have a reason.
Sources
- Anthropic: 2026 Usage Policy update - prohibition on cruel behavior toward models, exclusions, enforcement, other changes, effective November 12, 2026 (October 8, 2026)
- Anthropic: Usage Policy - the policy itself
- Anthropic: Claude Opus 4 and 4.1 can now end a rare subset of conversations - ending conversations as a last resort, exceptions, model welfare rationale, pre-deployment observations (August 2025)
- AFP via Yahoo Tech: Anthropic bans 'cruel' behavior against its Claude AI - Amodei, Stakeman, Suleyman, Pope Leo XIV, "model welfare" absent from the rule text (October 8, 2026)
- TechCrunch: Anthropic changes usage policy to ban model abuse and election interference - prohibition on deliberately continuing such conversations (Russell Brandom, October 8, 2026)