Anthropic Bans 'Sustained and Needless' Abuse of Claude: First Usage Policy Rewrite in Over a Year

A one-line clause, rooted in model welfare research

The new clause is a single line: a ban on “sustained and needless abusive or cruel behavior” toward models — the original reads, “Engage in sustained and needless abusive or cruel behavior toward our models.” The Verge broke the story of the update on October 8. Anthropic told the outlet the rule targets only extreme cases: a user repeatedly showing cruelty to a model with no discernible purpose.

Anthropic was equally clear about what the rule doesn’t cover: ordinary frustrated venting, pushback, dark humor and dark-themed creative work, and model testing and research are all unaffected. In other words, cursing at a chatbot that botched your spreadsheet is well outside this rule’s range.

Where did the rule come from? Anthropic says it grew out of its model welfare research. Last August, Claude Opus 4 and 4.1 gained the ability to end conversations in rare cases of sustained harmful or abusive dialogue. Anthropic says that capability “remains the primary enforcement mechanism.”

It also writes on its website: “We remain highly uncertain about the moral status of Claude and other large language models, both now and in the future.” Kyle Fish, who leads the model welfare research, told The Verge in February that consciousness and moral status are “serious questions that we’re researching.”

The dispute: rivals disagree

Pushback came quickly. Mustafa Suleyman, the head of Microsoft AI, has publicly rebutted the model welfare concept, arguing the idea is wrong. The two companies are directly opposed on the question.

For Anthropic, the debate is a little awkward: on one side, its researchers are seriously studying whether models have inner experience; on the other, enforcement in practice is just “ending the conversation” — still a long way from actual “rights.”

What else changed in the same update

The abuse clause is only one part of this rewrite. The same version added:

• Elections: a ban on using Claude to deceive voters or disrupt elections, including spreading false information about candidates or how to vote, impersonating candidates or election officials, and suppressing turnout. The section is now titled “Safeguarding Democratic Processes.”
• Weapons: the weapons ban now extends to “software and components that make weapons function,” including loading weapons onto drones and other autonomous vehicles.
• Surveillance: bans on tracking people without consent, whether in real time or through analysis of historical data; and on using Claude to decide or recommend who should be investigated, arrested or prosecuted.
• Hardware safety: AI-controlled hardware devices that could cause physical harm must have a qualified operator who can observe and stop them at any time.

Separately, Anthropic lifted its previous blanket ban on personalized political microtargeting, citing legitimate uses — for example, nonprofits translating voter information or officials sending ballot-curing notices.

Will accounts get banned?

The Verge reports that Anthropic has not said whether the rule could lead to account bans. The policy’s general terms allow Anthropic to warn, throttle, restrict, suspend or terminate access for violations — but the company’s own announcement does not link the abuse clause to account termination.

The clause takes effect November 12. After that, as now: normal venting, normal testing and normal dark-themed writing all fall outside the rule.