Anthropic updates its rules for the first time in a year: no cruelty to Claude

For the first time in more than a year, Anthropic has rewritten the acceptable use policy for its models. Most of the changes target new and dangerous abuse scenarios: election interference, weapons development, surveillance, and the use of AI in medicine and finance. One clause stands out more than the rest. It bans “systematic and needless cruel or abusive behavior” toward Claude itself. Documents like this usually protect people from the consequences of AI. Here, part of the protection points toward the model.

What exactly changed

According to The Verge, the update reflects new, high-risk cases of misuse. The key points:

• the first change to the usage policy in more than a year;
• new or clarified risk areas: election interference, weapons development, surveillance, and medical and financial scenarios;
• a new ban on “sustained and needless abusive or cruel behavior” directed at Claude;
• the main enforcement mechanism: the model itself ending the conversation.

The available portion of the article does not say what else, besides closing the conversation, a violation might bring. It also does not say how the wording on elections, health and finance differs from the previous version. You will have to read the document itself to find out.

Where the clause on cruelty to the model came from

Last August, Anthropic announced that Claude would be able to end conversations with users who “persistently harm” it or behave abusively. This was part of its “model welfare” research. At the time it looked like a product feature and an experiment. Now the behavior is written into the policy, which gives the company a formal basis to respond to such cases.

The company treats the topic as a subject of research. The Verge's account offers no evidence that the model has experience or suffering, and Anthropic does not claim that it does. Still, the rule shows how the lab approaches its relationship with its product. It prefers to err on the side of caution and work through the question in advance, while there is no scientific answer.

The word “needless” matters. The ban targets sustained, pointless abuse, not every harsh message. Rigorous testing, red teaming and jailbreak-resistance checks all have a purpose. The policy text draws the line between them and abuse for abuse's sake in words, and those words are where the disputes will center.

What teams working with Claude should do

First, read the new version in full rather than a summary. If you build a product in medicine, finance, analytics or monitoring, check your use cases against the updated sections. Anthropic has explicitly named these areas high-risk, and checks there may become stricter.

Teams that do red teaming should prepare in advance. Document the goal, scope and methodology of your tests. That makes it easy to tell aggressive prompts with a clear purpose apart from “pointless cruelty.” Cap the duration of automated runs: sessions that spend hours pushing the model through abusive templates with no research value risk falling under the new rule.

Finally, build handling of model-ended conversations into your interface. If Claude closes a conversation, the user should get a clear explanation and the option to start a new chat. Otherwise, support will get a flood of tickets about a “broken” assistant.