Anthropic has revised its acceptable‑use policy to forbid ‘‘persistent and unnecessarily abusive or cruel’’ interactions with its Claude model, and has clarified prohibitions across several high‑risk areas including weapons assistance, political and commercial manipulation, surveillance, and autonomous physical systems.
What the new policy bans
The prohibition targets situations where a user repeatedly behaves cruelly toward the model without a recognisable purpose. Anthropic says the restriction does not cover ordinary user frustration, critical feedback, model testing, research, or ‘‘darker creative topics’’, but applies to sustained abusive behaviour that the company considers harmful.
Anthropic previously allowed Claude to end conversations in cases of persistently harmful or abusive communication; the updated policy confirms that conversation termination remains the primary enforcement mechanism. The company has not explicitly clarified whether additional penalties — such as account restrictions or bans — may be applied.
Classic safety restrictions and clarifications
Alongside the model‑well‑being framing, the policy adds or clarifies several traditional AI safety prohibitions:
- Political and commercial deceptive campaigns: banned practices include hiding the true sponsor, amplifying content via fake accounts, and misleading voters. The rules explicitly forbid disinformation about candidates, election officials, or voting methods, and activities intended to suppress participation.
- Weapons development: the prohibition now specifically covers software and components that make weapons operational, and extends to arming drones or other autonomous vehicles.
- Surveillance: Claude must not be used to track people without their consent, whether via real‑time tracking or analysis of previously collected data. The policy also bans using the model to decide or recommend whom to investigate, arrest, or charge, and excludes use for developing surveillance tools.
Anthropic reserves the right to apply different terms for certain government contracts if contractual safeguards are deemed sufficient to mitigate the risks.
Rule for autonomous physical systems
The company added a requirement for AI systems connected to hardware capable of autonomous physical action that could cause harm: a trained operator must be able to monitor the device and halt it if necessary. The Verge notes Anthropic did not make clear whether that operator must be a human.
This rule responds to an industry trend: AI labs are increasingly investing in robotics and other hardware that give AI systems a physical presence, and the policy may signal considerations for Anthropic’s future partnerships in that space.
Why the change drew attention
The update attracted particular scrutiny because Anthropic has in recent months signalled openness to exploring questions about potential internal experience or moral status of models. Kyle Fish, head of the company’s model‑well‑being research, said in February that questions about possible inner experience, consciousness, moral status, and well‑being are being taken seriously. CEO Dario Amodei has also said in a podcast: “we don’t know whether models are conscious.”
The New York Times reported that Anthropic consulted religious and philosophical thinkers about how to shape Claude’s behaviour morally and what it would mean for the company to take model consciousness seriously. Other industry figures have taken a more cautious line: Sam Altman, CEO of OpenAI, has warned that assigning religious significance or surrendering human judgment to AI models would be risky.
Implications
Anthropic’s policy change functions both as a safety manual and a philosophical position. Its restrictions on propaganda, election manipulation, weapons assistance, surveillance, and autonomous systems align with broader international debates on frontier AI governance. The ban on ‘‘cruel’’ treatment of Claude introduces a harder‑to‑measure category that raises questions about where regulators and companies will draw the line between handling user abuse, ensuring model safety, and attributing moral status to machines.
The Verge reported this is Anthropic’s first substantial change to its terms of use in more than a year; the policy leaves several operational details open, including potential additional sanctions. How these boundaries will be enforced and where they will fall in future AI governance regimes remains a central issue for technical, legal, and ethical discussion.
