Anthropic Updates Its Usage Policy for the First Time in More Than a Year, Banning “Abusive or Cruel” Behavior Toward Claude
If users repeatedly engage in such behavior without a clear reason, the primary measure remains ending the conversation. However, this does not cover frustration, arguments, or ordinary model testing.
The new policy also includes restrictions on propaganda, deceiving voters, weapons development and weapons-control software, as well as non-consensual surveillance. If Claude is connected to hardware that could cause harm, a human operator must monitor it and be able to stop the device when necessary.
When Claude No Longer Has to Endure Repeatedly Cruel Treatment
Anthropic has added a ban on repeatedly cruel or abusive behavior toward Claude without an apparent reason, while continuing to use conversation termination as its primary enforcement mechanism.
This rule targets only extreme cases. It does not include frustration, arguments, dark creative themes, or ordinary model testing and research.
How This Differs from Anthropic’s Previous Policy
Anthropic stated that weapons development was already prohibited, but the new rules expand the ban to include software and components that enable weapons to function, as well as arming drones and autonomous vehicles. The surveillance restrictions have also been clarified, including bans on tracking people without consent and using Claude to create or improve surveillance tools.
| Factor | Previous policy | New update |
|---|---|---|
| Weapons development | Weapons development prohibited | Includes software, components, and arming drones |
| Surveillance | Already prohibited | Bans non-consensual tracking and the creation of surveillance tools |
| Deceptive campaigns | Restrictions were scattered | Consolidated into a ban on deceptive commercial or political campaigns |
The Problems This Policy Attempts to Address
Claude has been used for everything from creating propaganda and fabricating accounts to amplify messages, to tracking political dissenters and helping develop software and components that enable weapons to function. At the same time, Anthropic has added a ban on repeatedly cruel treatment of Claude without a clear reason.
The primary response remains ending the conversation, which sounds straightforward. But the question is whether that is enough, because some forms of harm occur before the system stops responding—especially when Claude is connected to hardware capable of acting on its own.
What Kind of Usage Framework Is Anthropic Establishing for Claude?
Anthropic is positioning Claude as a tool for high-risk work that requires clearly defined boundaries, including election interference, propaganda, weapons development, surveillance, and health- and finance-related uses that could have significant consequences.
If Claude is connected to hardware that can act independently and could injure people, a qualified operator must monitor it and be able to stop the device when necessary. Government customer contracts may be exempt if Anthropic determines that contractual restrictions and safety measures sufficiently reduce the risks.
From Scattered Bans to More Clearly Written Rules
If someone repeatedly uses Claude in a harmful or cruel way without a clear purpose, the system may end the conversation. This does not include frustration, arguments, dark-toned creative work, or model testing.
Politically, campaigns that conceal their sponsors or use fake accounts are prohibited, as are efforts to deceive voters, impersonate candidates or officials, and discourage voter turnout.
For safety, Claude may not be used to develop software or components that enable weapons to function, including arming drones. For surveillance, users may not track people without consent, create monitoring tools, or ask Claude to determine who should be investigated, arrested, or prosecuted.
Anthropic’s Position Compared with Microsoft AI
Anthropic leaves room to study questions about internal experiences, consciousness, moral status, and model welfare because there are still no clear answers.
Microsoft AI previously used stronger language, rejecting the idea of welfare, rights, and legal personhood before adopting a more cautious position.
| Factor | Anthropic | Microsoft AI |
|---|---|---|
| Position on model welfare | Studying internal experiences, consciousness, moral status, and welfare | Rejecting rights, welfare, or legal personhood |
| Policy tone | Acknowledges that these remain questions requiring further study | Shifted from a firm position to more cautious wording |
What This Policy Does Well—and What Still Raises Questions
Its strength is bringing previously scattered prohibitions into clearer focus, from propaganda and voter deception to weapons development and surveillance systems. This makes the policy better equipped to address uses connected to the real world.
However, it remains unclear what penalties exist beyond ending the conversation, and whether operators of autonomous devices must be human.
Pros
- +Consolidates scattered prohibitions into a clearer framework
- +Covers uses that could affect the real world
Cons
- −Does not explain penalties beyond ending the conversation
- −Does not clarify whether device operators must be human
The Still-Unclear Line Between Protecting the Model and Controlling Usage
Anthropic describes having Claude end conversations with users who repeatedly behave abusively or cruelly as part of its “model welfare” research. However, the rules clearly state that this applies only to severe cases and does not include frustration, arguments, or ordinary model testing.
The boundary therefore remains unclear. Kyle Fish has said that the company is still studying models’ consciousness and moral status, while Dario Amodei has openly acknowledged that Anthropic does not yet know whether models are conscious. For now, ending the conversation remains the primary mechanism, so it can be viewed both as protecting the model and as controlling user behavior.
What This New Policy Says About Claude’s Future
This policy suggests that Claude may not remain confined to a screen. It is being framed for work alongside robots and hardware capable of taking physical action. If a device could cause harm, a suitably qualified operator must monitor it and be able to stop the system when necessary.
At the same time, Anthropic prohibits using Claude to control or direct weapons, including arming drones and autonomous vehicles. Surveillance operations must not track people without consent, and systems may not be allowed to decide who should be investigated, arrested, or charged. This reflects how the company’s investment in robotics and hardware will likely come with increasingly stringent safety restrictions.