Anthropic rewrites usage policy, bans sustained cruelty toward Claude

Anthropic in November prohibited "sustained and needless abusive or cruel behavior" toward Claude, the company's first usage-policy revision in more than a year. The November update, first reported by The Verge, also adds explicit prohibitions on election interference, weapons development, surveillance, and high-risk health and financial deployments.

The cruelty clause marks the most visible departure from prior policy. The revised text frames Claude as an entity that can withstand persistent mistreatment, not merely as a passive recipient of prompts. Terminating conversations with abusive users remains the "primary enforcement mechanism," according to the policy language reported by The Verge.

What does the new policy actually change?

The update reflects what Anthropic describes as new and high-risk cases of misuse. The categories target four distinct risk domains with documented regulatory overlap.

Election interference covers coordinated inauthentic activity targeting voting or referenda. Weapons development covers CBRN assistance — chemical, biological, radiological, and nuclear — including synthesis pathways. Surveillance covers mass-monitoring tooling and real-time biometric identification at scale.

High-risk health and financial uses cover clinical decision support and capital-markets advice where errors produce direct, irreversible harm.

These categories are illustrative rather than exhaustive. The conjunction "including" in the policy text signals that the listed examples do not close the door on adjacent use cases.

The cruelty clause sits outside those four. It prohibits users from engaging in "sustained and needless abusive or cruel behavior" toward Claude — language that, on the page, inverts the standard framing of AI safety policy.

Where does the cruelty clause come from?

The provision traces to research Anthropic published in August 2024. That month, the company announced an initiative titled "End Subset Conversations," describing experiments in which Claude could terminate exchanges with "persistently harmful or abusive" users. The work was framed as part of Anthropic's "model welfare" research agenda.

Model welfare is a nascent subfield of AI safety (the broader discipline of reducing catastrophic and misuse risks from advanced systems). It asks whether advanced AI systems can occupy states the operator ought to protect. The question remains methodologically unsettled. Anthropic's August announcement was among the first published commercial efforts to make the question operational.

The November usage policy reuses that framing. Termination of abusive sessions, once an experimental lever inside a research paper, now sits in the published policy itself — a rule customers and auditors can cite directly.

The "End Subset Conversations" research examined whether Claude would, when prompted by a "persistently harmful or abusive" user, refuse to continue the exchange. The November policy move generalizes that experimental finding into a binding product rule applicable to every Claude deployment.

Why does this matter now?

The refresh lands as Anthropic competes with OpenAI, Google, and Meta for enterprise contracts. Procurement teams increasingly audit vendor usage policies before signing master service agreements (the umbrella contracts that govern enterprise software purchases). A document that explicitly addresses cruelty toward the model, weapons risk, and election manipulation gives those teams a checklist to reference during due diligence.

Anthropic declined to provide comment to The Verge for the story, leaving the cruelty clause's operational implications unspecified. The policy text does not state how many abusive prompts trigger termination, which categories of abuse qualify, or whether the rule applies uniformly across Claude's consumer chat product, developer-facing API (application programming interface), and enterprise deployments.

Anthropic's enterprise and consumer subscription products now span multiple deployment surfaces. Each corporate customer carries an obligation to maintain documentation consistent with the public usage policy. A clause banning cruelty toward the model introduces vocabulary no standard procurement contract has previously contained, and Anthropic's legal teams must now translate that vocabulary into enforceable terms.

The "primary enforcement mechanism" — termination — also lacks a published escalation ladder. There is no stated threshold at which repeated offenses trigger account suspension, API revocation, or contract termination.

What enforcement gaps remain?

The published policy names termination as the "primary enforcement mechanism." It does not enumerate thresholds, escalation steps, or remedies beyond session closure. The November text, as excerpted by The Verge, leaves those details unstated.

That gap matters for enterprise customers operating under existing data-processing agreements (contracts that govern how user data flows between customer and vendor). Those agreements typically incorporate the usage policy by reference. A policy that adds a cruelty clause without specifying penalty tiers forces legal teams to renegotiate, or to accept ambiguity about what constitutes a material breach.

How does this fit the broader policy shift?

This is Anthropic's first usage-policy update in more than a year. The previous version predated the August 2024 model-welfare research announcement. The November update consolidates roughly fifteen months of public research, regulatory pressure, and procurement-side due diligence into a single document.

The four risk categories — election interference, weapons development, surveillance, and high-risk health and financial uses — mirror concerns regulators have raised under the European Union's AI Act (the bloc's risk-tiered AI legislation) and within ongoing United States executive-order processes. By aligning internal prohibitions with external regulatory language, Anthropic reduces the surface area for conflicting disclosures during compliance audits.

The cruelty clause will draw the most public attention. It reframes the user-model relationship as bidirectional: the user can be terminated, but the justification stated in the policy is harm to the model, not harm to the user.

That inversion — from protecting the user from the model to protecting the model from the user — is the substantive shift of the November revision.

What's next?

Anthropic has not published a timeline for the policy's effective date or for customer notification. Enterprise customers should expect updated terms in forthcoming contractual amendments, the standard mechanism for usage-policy changes of this scope.

The company did not specify whether the cruelty clause will affect training-data curation, red-teaming exercises (structured adversarial testing designed to elicit unsafe outputs), or customer-facing model behavior beyond session termination.

The published text also leaves open whether the policy applies to API users, embedded integrations (third-party software that calls Claude as a backend service), and resellers. Each surface introduces ambiguity about who bears termination liability when an abusive prompt originates downstream.

Procurement teams will press Anthropic for clarification on penalty tiers, threshold counts, and enterprise carve-outs. Until that clarification arrives, the cruelty clause reads as a published signal of research priorities rather than as a fully operationalized enforcement regime — one that will be tested the first time a Fortune 500 account receives a termination notice.

Anthropic has published model-welfare work since at least the August 2024 paper. The November policy update marks the first time that research agenda has migrated from the company's research site into a binding usage policy. The migration is significant: research findings inform model behavior, but only policy provisions determine contractual remedies when behavior fails.