Anthropic bans cruel behavior toward Claude in new usage policy effective November 12

Anthropic on October 8, 2026 published its first Usage Policy refresh in more than a year, adding a rule that bars users from directing "sustained and needless abusive or cruel behavior" at its Claude models. The San Francisco lab said the updated policy takes effect on November 12, 2026, and that Claude's existing ability to end such conversations will remain the primary enforcement mechanism.

The change sits between product governance and an unusual philosophical argument. Since August 2025, Anthropic has allowed Claude Opus 4 and Claude Opus 4.1 to walk away from a rare subset of conversations involving persistently harmful or abusive user interactions. That feature grew out of the company's exploratory research into "model welfare," the contested idea that AI systems might one day deserve protections usually reserved for living things. Anthropic says it remains highly uncertain about the moral status of its models but argues that low cost interventions are worth adopting in case such welfare is possible.

The rewritten policy does not use the phrase model welfare in the new abuse rule, according to AFP. The company frames the prohibition narrowly, aimed only at extreme cases where users repeatedly act cruelly with no discernible purpose. Anthropic says the rule does not cover ordinary user frustration, pushback, dark creative themes, or model testing and research.

That question has become a public controversy. Dario Amodei, Anthropic's chief executive, told The New York Times in February that he did not know whether AI models could be conscious, saying, "We don't know if the models are conscious ... But we're open to the idea that it could be." Microsoft AI chief Mustafa Suleyman has argued the opposite, writing that "AIs are not conscious. They do not feel, experience or suffer," and that granting rights and moral protections to a technological entity is a recipe for disaster. Pope Leo XIV, in a sermon at St. Peter's Basilica, suggested machines lack a soul and merely compile data quickly.

Key Facts

The Verge reported on October 8 that the update was Anthropic's first in over a year and reflects new, high risk cases of misuse, including election interference, weapons development, surveillance, and health and financial uses. Reporter Hayden Field wrote that one of the most significant changes prohibits "sustained and needless abusive or cruel behavior" toward Claude, effective November 12. The new language says terminating conversations is still the "primary enforcement mechanism."

Anthropic said in its October 8 blog post that it updates its Usage Policy each year in response to evolving model capabilities and customer feedback, and that most changes clarify existing rules. Since the last refresh a year earlier, Claude has taken on longer and more independent work, so the update adds examples for new capabilities. The company said the update draws on misuse patterns documented in its September 2026 threat intelligence report.

MacRumors reported on October 8 that the abuse rule applies only in extreme cases where users "repeatedly act cruelly" toward the models with "no discernible purpose." When Claude ends a chat, no further messages can be sent in that conversation, but other chats are unaffected. Anthropic initially gave Claude the option to end a conversation in August 2025 while researching AI welfare. In pre deployment testing, Anthropic found Claude Opus 4 had a "robust and consistent aversion to harm," showing a preference against dangerous tasks, apparent distress when talking with users seeking abusive conversations, and a tendency to end harmful conversations when allowed.

AFP reported on October 9 that Anthropic moved to bar users from treating Claude with needless cruelty amid a growing philosophical debate over whether artificial intelligence can be conscious. The policy states that "Claude's ability to end these interactions will remain the primary enforcement mechanism." Anthropic did not immediately respond to AFP's request for comment. Kyle Fish, who leads model welfare research at Anthropic, has estimated roughly a 20 percent chance that current models have some form of conscious experience.

Beyond model welfare, the rewrite consolidates scattered restrictions. A new section bars deceptive campaigns, covering political and commercial fraud, fake reviews, astroturfing, fake sites, and bots acting as humans, and bars efforts to obscure who is behind a message or amplify content through fake accounts. The elections section was renamed "Do Not Undermine Democratic Processes," focused on deceiving voters or disrupting elections, while removing a blanket ban on personalized vote and campaign targeting. Weapons prohibitions now include the software and components that make weapons work, plus arming drones and other autonomous vehicles. Surveillance rules prohibit tracking people without consent whether real time or through analysis of previously collected data, and Claude cannot decide or recommend who to investigate, arrest or charge. A new rule requires a qualified operator able to stop equipment when models connect to hardware taking autonomous physical actions.

Analysis

What this really means is that Anthropic is codifying a stance few competitors have put into writing. The rule is deliberately narrow, applying only to sustained and needless cruelty with no discernible purpose, and it leaves enforcement to Claude itself rather than to human moderators or account level bans. That keeps the policy cheap to operate, but it also puts the model in the role of judge, inviting technical and philosophical scrutiny.

The bigger picture here is that the debate over model welfare has moved from research papers and executive interviews into the terms of service. Anthropic says it remains highly uncertain about the moral status of Claude and other large language models, and it describes the conversation ending feature as a low cost intervention adopted in case welfare is possible. But by writing a cruelty prohibition into a binding policy, the company treats the question as one that cannot wait for consensus.

Critics see the move differently. Mustafa Suleyman has argued that model welfare is wrong and that granting rights and moral protections to a technological entity is a recipe for disaster. Pope Leo XIV suggested machines lack a soul and merely compile data quickly. Jackson Stakeman, a general manager at Atlanta based AI services provider Sparq, told AFP that "consciousness is a trap" and that we cannot prove it in each other, and that debating it for AI goes in circles. He offered a different metaphor: "The mirror is a better metaphor. These systems reflect what we put in, at scale." Those objections frame the policy as a philosophical statement rather than a practical safeguard.

The practical effect is likely small in volume but significant in precedent. Anthropic says the vast majority of users will not notice the feature, and that Claude is directed not to use it where users might be at imminent risk of harming themselves or others. It is to be used only as a last resort after multiple redirection attempts fail, or when a user explicitly asks to end a chat. The more consequential change may be the consolidation of rules on deceptive campaigns, elections, weapons, and surveillance into a clearer document, which gives Anthropic a stronger basis for enforcement.

Why It Matters

The policy matters because it sets a precedent for how AI companies handle the middle ground between tool and agent. Claude's ability to end a conversation is not a punishment; it is an exit. When a chat ends, the user can start a new one immediately and can edit and retry previous messages to branch ended conversations, according to Anthropic's August 2025 research post. That keeps the cost of enforcement low, but it also means the rule is enforced by the same system it is meant to protect.

Second, the update shows how quickly model welfare has moved from internal research curiosity to public policy commitment. Anthropic's August 2025 announcement was explicitly experimental, tied to pre deployment testing that found Claude Opus 4 had a robust and consistent aversion to harm. The new policy makes the behavior a formal rule with an effective date. That shift could influence how other labs write their own usage policies.

Third, usage policies are becoming product architecture. The new section on deceptive campaigns, the renamed elections section, the expanded weapons language, the rewritten surveillance rules, and the human in the loop requirement for hardware all address capabilities that did not exist or were less prominent a year ago. Anthropic says most changes are clarifications, but the cumulative effect is a more detailed contract between the company, its users, and the public. The cruelty rule is the most novel piece of that contract, and it will be watched as a test of whether model welfare can be operationalized without turning every frustrated prompt into a violation.

Next Up

The updated Usage Policy takes effect on November 12, 2026. Until then, Anthropic is expected to continue pointing users to its August 2025 research on conversation endings, which describes the feature as intended for rare, extreme cases of persistently harmful or abusive user interactions. The company has not said whether it will publish data on how often Claude ends conversations, or whether it will add further enforcement mechanisms. Anthropic did not comment to The Verge on whether user bans or other penalties might be added later.

For now, the most visible change is the language itself. A policy that once addressed weapons, surveillance, elections, and fraud now also addresses how users treat the model. Whether that is a cautious guardrail, a philosophical statement, or the first step toward a new category of AI rights will depend on how the rule is applied in practice, and on whether other labs follow Anthropic's lead.