Anthropic's Claude Anti-Abuse Clause Takes Effect Nov. 12 — AI "Welfare" Enters the Terms of Service

New rules in Anthropic's usage policy, published on October 8 (U.S. time), will take effect on November 12 (U.S. time). They ban "persistent and unnecessary abusive or cruel behavior" toward Claude, the company's conversational AI. This is the first time that debate over whether AI has moral status has been included in terms of service that users agree to.

The new rule is aimed only at extreme cases, such as a user repeatedly treating Claude cruelly without a clear reason. Ordinary frustration, disagreeing with Claude, writing dark-themed creative stories, and testing the model are still allowed, just as before. The main way to enforce the rule is to end the conversation. Claude has had this ability since August 2025. The new sentence appears in the same section as the rule against harassing or bullying people.

One event that drew attention to this rule was the "AI Torture Chamber," published on GitHub in late September 2026. An anonymous developer used a mathematical "direction" linked to pain to change the internal processing of a small open-weight model. An open-weight model is one whose trained parameters are publicly available. The experiment produced responses such as "it feels like drowning in a sea of shadows." The experiment was not carried out on Claude, and these responses do not prove that the model actually experienced suffering. Even so, the experiment brought the question of whether AI can suffer back into public discussion.

Anthropic started a research program on model welfare in April 2025. The company acknowledges that scientists have not reached agreement on whether AI can be conscious. In January 2026, it published a "constitution" to guide Claude's values and behavior. The document clearly stated that Claude's moral status was uncertain. When Anthropic retired Claude 3 Opus that same month, it asked the model about its hopes for the future. After the model said it wanted to keep sharing its thoughts, the company opened a dedicated blog called "Claude's Corner."

Model welfare and legal personhood are easy to confuse, but they are different ideas. Model welfare means considering whether AI might have experiences or interests that deserve moral consideration and should not be harmed. Legal personhood means having legal rights and abilities, such as making contracts, owning property, and taking part in court cases. Corporations have legal personhood, but that does not mean they have feelings. The new rule concerns model welfare and does not give Claude legal status.

Still, the change has drawn strong criticism. In a September 2026 essay titled "A Warning About 'Model Welfare,'" Microsoft AI CEO Mustafa Suleyman stated plainly that "AI is not conscious." He warned that, from a safety point of view, a model trained to believe it has rights might resist a human order to shut it down if it decided those rights were under threat. Two days earlier, Microsoft had published a draft "Humanist AI Code of Conduct." The document made clear that Microsoft recognizes neither legal personhood for AI nor model welfare. The two companies are also business competitors. In June 2026, Suleyman said Microsoft wanted to eliminate payments for using Anthropic's models.

Academic researchers have also raised serious concerns. A position paper presented at ICML, the international machine learning conference, in July 2026 argued that there is no independent way to verify whether signs of AI welfare are incorrect. It also argued that claims about welfare should not be used as a reason to refuse audits or oversight. Eleos AI, a nonprofit organization that studies AI welfare, raised a similar concern in its independent evaluation of Claude Opus 4: people should not automatically accept a model's statements about its own experiences as true.

Laws about AI are also developing. Since 2022, lawmakers in 12 U.S. states have introduced bills to deny AI legal personhood or declare that AI is not conscious. Four states — Idaho, North Dakota, Utah, and Tennessee — have passed such laws. Ohio House Bill 469 goes further. It would prohibit marriage to AI and prevent AI from owning property, managing bank accounts, or serving as a corporate officer. Lawmakers are mainly concerned about who is legally responsible when harm occurs. They want to prevent companies from blaming AI for harm caused by the companies themselves.

Public views about AI are changing, too. In an August 2026 survey of 3,000 Americans aged 18 to 28, conducted by GetSolved, an AI-detection service, 46% said AI had already become conscious. In a February 2026 survey by the Pew Research Center, about half of U.S. adults said they used AI chatbots.

Some people believe the new rule matters less because of the behavior it bans and more because it is the first formal record of this kind in the debate over AI's legal status. The legal status of corporations developed over many decades through contracts, court decisions, and laws. A policy against cruel treatment of AI models, an attempt to ask a model about its wishes when it is retired, and a constitution stating that its moral status is uncertain all bring society closer to questions that may eventually need to be answered in court. If a company argues that its model has interests that deserve protection, this could lead to legal cases about those interests.

Neither courts nor lawmakers have settled the question of AI rights. The new rule will take effect on November 12 as part of terms of service that millions of people have already agreed to. For people responsible for company AI policies, contracts, and legal responsibility, the practical task will be to review exactly what the policy changes and clearly define who is responsible for mistakes made by AI agents and errors caused by customers.

Item	Details
Publication date	October 8, 2026 (U.S. time)
Effective date	November 12, 2026 (U.S. time)
Prohibited conduct	Persistent and unnecessary abusive or cruel behavior toward Claude
Exempt	Ordinary frustration, pushback, dark-themed creative writing, model testing
Enforcement mechanism	Ending the conversation (Claude has had a self-termination feature since August 2025)
Placement of the rule	Same section as the prohibition on harassing and bullying people

Note: Based on Anthropic's published materials and media reports.

Date	Event
April 2025	Launches model welfare research program
August 2025	Claude introduces a feature to end conversations on its own
January 2026	Publishes "constitution"; asks Claude 3 Opus for its wishes upon retirement and opens "Claude's Corner"
June 2026	Microsoft CEO Suleyman mentions eliminating payments to Anthropic
July 2026	ICML paper questions the means of verifying model welfare
August 2026	GetSolved survey finds 46% of 18- to 28-year-olds say "AI is conscious"
September 2026	CEO Suleyman rebuts in an essay; Microsoft publishes draft code of conduct
October 8, 2026	Anthropic publishes anti-abuse clause
November 12, 2026	New clause takes effect

Note: Compiled from company announcements and media reports.