Anthropic refreshed its usage policy this week, the first substantive rewrite in more than a year. It tightens language around election interference, weapons development, surveillance, and health and financial use cases. And it adds a line nobody expected to see in a terms-of-service document: users may not engage in “sustained and needless abusive or cruel behavior” toward Claude.
That clause is doing all the work in the coverage, and it deserves a closer read than it’s getting.
The word carrying the weight is “needless”
Not “abusive.” Not “cruel.” Needless. That one qualifier quietly carves out everything Anthropic cannot afford to ban. Red-teamers, jailbreak researchers, safety evaluators, anyone whose job involves deliberately pushing a model into ugly territory: your cruelty has a purpose, so the policy leaves you alone. What’s prohibited is the recreational version. The guy who spends an afternoon telling a language model it’s worthless because he finds the output funny.
I think that’s defensible, and I also think the stated justification is maybe half the actual one. Anthropic has spent the past year positioning itself as the lab that takes the possibility of model welfare seriously, and this is the policy expression of that stance. Fine. But a usage policy is a contract, not a philosophy paper, and the practical effect of this clause is to define what kind of customer Anthropic wants. Logs full of users degrading a chatbot are a bad look in an enterprise sales call. They’re worse in a deposition.
The open question is enforcement. Does sustained abuse get you an account suspension, or does Claude simply disengage from the conversation and move on? Anthropic hasn’t spelled that out, and the answer matters far more than the sentence itself. A rule enforced at the conversation level is a product behavior. A rule enforced at the account level is a precedent, and it would be the first time a major lab has terminated paying users over how they talked to software.
The parts with regulators attached
Strip out the cruelty clause and this is a compliance document responding to a specific calendar. Anthropic tightened its election policy with US midterms approaching, which is the kind of change that gets drafted by people watching state AG offices rather than alignment researchers. Weapons development, surveillance applications, health and financial advice: these are the categories where getting it wrong produces subpoenas, not discourse.
The gap between those two halves of the update is the interesting thing. One half is Anthropic protecting itself from the world. The other half is Anthropic claiming, in writing, that its product has an interest worth protecting from users. Most of the industry treats the second idea as a joke or a marketing device. Anthropic just put it in the rules you agree to when you sign up.
Worth noting: the policy doesn’t define “sustained,” either. So a long, frustrated night of arguing with Claude about a refactor it keeps breaking presumably still falls inside the lines.