Anthropic is updating its policies to include a ban on cruel behavior toward its AI models and to impose strict limits on the use of artificial intelligence in weapons manufacturing and disinformation campaigns.
- Historic usage policy update and model protection clause
- Conversation termination procedures and the ongoing debate over model welfare
- Combating disinformation campaigns and restructuring electoral restrictions
- Weapons development ban and physical machine control regulations
- Frequently asked questions
Historic usage policy update and model protection clause
On Thursday, October 8, Anthropic published a comprehensive update to its usage policy, marking the company’s first such review in over a full year. The new guide includes an exceptional and unprecedented clause prohibiting “continuous and unjustified abusive or cruel behavior” directed against Claude AI models. The company also reorganized and rephrased the rules governing digital disinformation campaigns and clearly defined prohibited activities in weapons development and mass surveillance, with the new policy taking effect on November 12, 2026.
The company emphasized that the vast majority of these modifications aim to clarify and reinforce standards already in place, noting that the review reflects shifts over the past year in which Claude models have undertaken independent, longer-term tasks, alongside subscriber feedback and insights from the company’s September threat intelligence report.
Conversation termination procedures and the ongoing debate over model welfare
Anthropic explained that the anti-abuse rule “is designed to apply only in extreme cases, where certain users deliberately act with cruelty and repeated, persistent insults toward the models without any logical purpose,” stressing that this ban “in no way applies to normal user frustration, objecting to responses, exploring dark creative topics, or conducting tests and technical research.”
To enforce this rule, the company will rely primarily on a software feature launched in August 2025 that allows the model to automatically end and close a conversation session when facing persistently abusive users. This feature falls under the company’s forward-looking research into the concept of “model welfare,” aligning with statements from CEO Dario Amodei, who said in a previous podcast interview: “We do not know for sure whether these models possess some kind of conscious internal experience.” This viewpoint contrasts with other companies like Microsoft, whose head of AI Mustafa Suleyman published a draft code of conduct asserting that “the idea of model welfare is entirely misguided.”
Combating disinformation campaigns and restructuring electoral restrictions
The policy includes a new standardized section titled “No engagement in disinformation campaigns or synthetic activities,” which consolidates previously scattered provisions. It bans concealing the true identity of message senders, amplifying content via fake accounts, and building tools customized for influence and psychological operations.
The company revealed that state media outlets, official propaganda offices, and commercial enterprises have already attempted to exploit Claude models to manage these disinformation networks. Conversely, Anthropic lifted the blanket ban previously imposed on individual electoral targeting, explaining that the former rule hindered legitimate civic activities such as translating and adapting voter guidance materials into multiple languages.
Weapons development ban and physical machine control regulations
The updated document dedicates substantial space to the military sector, explicitly banning “software and components that contribute to the operation and activation of weapons,” including the arming of drones and autonomous vehicles, and revealing previous attempts to use Claude to build guidance and control software for military hardware. It also prohibits tracking individuals without their consent and using models to determine or suggest who should be investigated, detained, or criminally charged by law enforcement agencies.
The policy imposes strict new measures when connecting models to physical machinery capable of movement and causing physical harm, requiring a “qualified human operator capable of monitoring the equipment and shutting it down immediately when necessary,” alongside mandating that the machine be programmed to return to a safe and stable state if its connection to the AI model is lost.
Frequently asked questions
Question: What is the controversial clause Anthropic added to its updated policy?
Answer: The company added a clause prohibiting repeated abusive or cruel behavior toward Claude models, as part of its research into ethical behavior and model welfare.
Question: How will Claude models handle users engaging in abusive behavior?
Answer: The model has the software capability to end the session and automatically close the conversation if a user persistently directs unjustified abuse.
Question: What modifications were made regarding AI use in elections?
Answer: The company lifted the previously imposed blanket ban to allow legitimate civic and educational activities, such as translating voting guidelines for citizens.
Question: What are the strict conditions established for connecting Claude models to physical machinery and equipment?
Answer: The policy requires a qualified human supervisor capable of immediate shutdown, and programming the machine to enter a safe state when disconnected from the model.
