PolicyLaw, regulation & governance
Anthropic's new usage policy bans needless cruelty toward Claude and lifts its blanket ban on campaign targeting
Anthropic's revised usage policy, effective November 12, 2026, bans sustained, purposeless abuse of Claude and enforces it mainly by ending chats. It also groups its influence-operation rules together, sharpens its weapons and surveillance bans, adds hardware safety requirements and allows non-deceptive personalized campaign outreach.
AT Pulse research deskOctober 9, 2026 · 5 min read · AI-assisted analysisSustained, needless cruelty toward Claude now violates the policy. Anthropic says it will enforce the rule mainly by having Claude end the conversation, not by suspending accounts. Model testing and research are explicitly exempt. The rule puts in writing a conversation-ending ability Claude has had since August 2025, which Anthropic described as a low-cost precaution given its uncertainty about model welfare. [3] [1] [4] [6] [2]
The election changes go in two directions. The rules against deceptive influence campaigns are now in one section, but the blanket ban on personalized campaign and voter targeting is gone. Anthropic says the old ban blocked legitimate civic work, such as multilingual voter information. Deceptive targeting, misuse of voter data, impersonating officials and suppressing turnout are still banned. Anthropic cites cases of fake-account networks and fabricated news sites built with Claude, but these are its own reports and the reviewed sources do not corroborate them. [3] [4] [7]
When Claude controls hardware that acts on its own and could injure someone, a qualified operator must be able to monitor and stop it, and the system must fall back to a safe state if Claude disconnects. The surveillance rules now bar Claude from deciding whom to investigate or arrest, and from tracking people without their consent. Fraud monitoring, content moderation, journalism and legal research remain allowed. [3]
The policy text shows what changed, but how vague terms like 'needless' will be enforced is still unknown. Campaigns and civic groups get room for personalized outreach.
Read the full assessment
Hardware and surveillance-adjacent builders must review their designs before November 12.
Executive brief
Anthropic's new usage policy makes "sustained and needless" cruelty toward Claude a policy violation. The company post says the rule covers only extreme, purposeless, repeated abuse, and it is enforced mainly by Claude ending the conversation. The same update, which takes effect November 12, 2026, groups the rules against influence operations into one section, makes the weapons and surveillance bans more explicit, and adds safety requirements for Claude connected to physical hardware. It also drops the blanket ban on personalized campaign and voter targeting, which got less coverage than the cruelty rule (Superpower Daily).
What changed and event timeline
Claude gets an "end conversation" ability
Claude Opus 4 and 4.1 could close chats in rare cases of persistent abuse. Anthropic described this as exploratory model-welfare work ().
Previous policy overhaul
Anthropic added rules against compromising computers and networks, written with agentic tools in mind. It also replaced its broad ban on lobbying and campaign content with narrower limits aimed at deception. The changes took effect September 15, 2025 ().
New policy published
It bans cruelty toward the models, election deception and deceptive influence campaigns, and spells out rules on weapons and surveillance (;).
Capabilities and access
- Claude Opus 4 and 4.1 got the conversation-ending ability in Anthropic's consumer chat apps (Anthropic research).
- When a chat is ended, no further messages can be sent in it, but the user's other chats are unaffected (MacRumors).
- Region restrictions now explicitly cover entities that are majority-owned or controlled by people or companies in unsupported regions (Anthropic).
Read the full section
- Claude Opus 4 and 4.1 got the conversation-ending ability in Anthropic's consumer chat apps (Anthropic research). The reviewed sources do not say which current models or surfaces have it.
- When a chat is ended, no further messages can be sent in it, but the user's other chats are unaffected (MacRumors). Users can start a new chat or edit an earlier message to branch the old one (Anthropic research).
- Region restrictions now explicitly cover entities that are majority-owned or controlled by people or companies in unsupported regions (Anthropic).
Technical analysis for researchers and developers
- Hardware integrations: These requirements apply when Claude drives hardware that acts on its own and could injure someone.
- High-risk uses (unchanged in substance): A qualified human reviewer with real authority must stay in the loop, and affected people must be told AI is involved.
- Evaluations and testing: The cruelty rule explicitly excludes model testing and research (TechCrunch), so red-teaming remains allowed.
Read the full section
- Hardware integrations: These requirements apply when Claude drives hardware that acts on its own and could injure someone. A qualified operator must be able to watch and stop the equipment, and the system must fall back to a safe state if Claude disconnects (Anthropic). Teams should design for losing the model connection and provide a manual stop.
- High-risk uses (unchanged in substance): A qualified human reviewer with real authority must stay in the loop, and affected people must be told AI is involved. The section was rewritten to state which recommendation types need these safeguards.
- Evaluations and testing: The cruelty rule explicitly excludes model testing and research (TechCrunch), so red-teaming remains allowed.
Claims and evidence
- Why Claude ends abusive chats: In Anthropic's own testing, Opus 4 strongly preferred not to engage with harmful tasks, showed what looked like distress when users kept pushing harmful requests, and ended such conversations when allowed to (Anthropic research).
- Misuse behind the new campaign rules: Anthropic says state media and commercial firms used Claude to run fake-account networks and fabricated news sites (Anthropic; MacRumors).
- Why the policy changed: Anthropic cites longer autonomous work by its models, observed threat patterns, customer feedback, and matching what it already enforces (Anthropic).
Read the full section
- Why Claude ends abusive chats: In Anthropic's own testing, Opus 4 strongly preferred not to engage with harmful tasks, showed what looked like distress when users kept pushing harmful requests, and ended such conversations when allowed to (Anthropic research). This is the company's interpretation of model behavior. No independent replication was found in the reviewed sources.
- Misuse behind the new campaign rules: Anthropic says state media and commercial firms used Claude to run fake-account networks and fabricated news sites (Anthropic; MacRumors). The reviewed sources give no independent corroboration of these cases.
- Why the policy changed: Anthropic cites longer autonomous work by its models, observed threat patterns, customer feedback, and matching what it already enforces (Anthropic).
Context and prior work
- The cruelty rule puts in writing a behavior Claude already had: The Verge calls the update the first policy change in over a year, and Claude has been able to end abusive chats since August 2025.
- Anthropic has said it is highly uncertain whether Claude has moral status, and describes conversation-ending as a low-cost precaution in case model welfare matters (Anthropic research).
- TechCrunch notes that Anthropic has worked with religious scholars, in some cases arguing that Claude could be conscious (TechCrunch).
Read the full section
- The cruelty rule puts in writing a behavior Claude already had: The Verge calls the update the first policy change in over a year, and Claude has been able to end abusive chats since August 2025.
- Anthropic has said it is highly uncertain whether Claude has moral status, and describes conversation-ending as a low-cost precaution in case model welfare matters (Anthropic research).
- TechCrunch notes that Anthropic has worked with religious scholars, in some cases arguing that Claude could be conscious (TechCrunch).
Limitations, safety and contested findings
- Vague threshold: "Sustained," "needless" and "no discernible purpose" are not defined further in the reviewed sources, so how the rule is enforced will decide what it means.
- Enforcement method: Coverage says the rule is enforced by ending conversations, not suspending accounts (Superpower Daily).
- Election targeting: Headlines say Anthropic is tightening election rules, but the update also loosens them by lifting the blanket ban on personalized campaign and voter targeting.
Read the full section
- Vague threshold: "Sustained," "needless" and "no discernible purpose" are not defined further in the reviewed sources, so how the rule is enforced will decide what it means.
- Enforcement method: Coverage says the rule is enforced by ending conversations, not suspending accounts (Superpower Daily). The policy itself says ending chats is the "primary" mechanism, which leaves room for other measures (Anthropic).
- Election targeting: Headlines say Anthropic is tightening election rules, but the update also loosens them by lifting the blanket ban on personalized campaign and voter targeting. Anthropic says the old rule blocked legitimate civic work such as multilingual voter information and ballot-correction notices.
- Self-harm exception: Claude is not supposed to end a chat when the user appears at imminent risk of harming themselves or others (Anthropic research).
Business and practitioner implications
- Campaigns, nonprofits and civic-tech groups: Personalized outreach becomes allowed from November 12.
- Robotics and hardware teams: Add an operator stop and a safe fallback state before connecting Claude to equipment.
- Surveillance and law-enforcement tools: Builders should review their products. Claude may not decide whom to investigate or arrest, or track people without consent.
Read the full section
- Campaigns, nonprofits and civic-tech groups: Personalized outreach becomes allowed from November 12. Targeting that relies on deception, misused voter data, impersonating officials or suppressing turnout stays banned.
- Robotics and hardware teams: Add an operator stop and a safe fallback state before connecting Claude to equipment.
- Surveillance and law-enforcement tools: Builders should review their products. Claude may not decide whom to investigate or arrest, or track people without consent. Fraud monitoring, content moderation, journalism and legal research remain allowed (Anthropic).
- Support and consumer apps: Normal user frustration is explicitly unaffected.