Suppose Claude invents a citation, gives you a broken link and gets the calculation wrong. You tell it the answer is terrible. Must you now send flowers to the data center?
Anthropic’s new abuse rule makes that an unexpectedly plausible question. Read the details, though, and there is still plenty of room to complain about the work.
Anthropic’s October 8 announcement says the update takes effect November 12, 2026. Ordinary frustration, disagreement, dark creative themes, and model testing or research are outside this provision. Ending conversations remains its primary enforcement mechanism.
The published policy prohibits persistent, purposeless abusive or cruel conduct toward models. Its general enforcement provisions also allow restrictions, suspension, or termination. There is no announced automatic account ban for an insult; there is also no blanket promise that accounts can never face sanctions.
Conversation endings predate this announcement. Anthropic introduced them in August 2025 for rare, extreme interactions, describing the feature as exploratory work involving possible model welfare and acknowledging uncertainty. A product rule does not establish that a system has feelings or can suffer.
For now, those are the terms users have to work with. The real test comes when a conversation is messy, a deadline is close and somebody has run out of polite ways to ask for the same correction.
The calculation is still wrong
“That answer is wrong. You cited a page that doesn’t exist. Check the source and redo the calculation.” A request like this gives the assistant something to fix. The missing citation remains missing whether the user remembered to say please.
Keeping the complaint specific also makes the next answer easier to check. Which claim has a source? What cannot be verified? Is that figure a measurement or an estimate? Those questions belong in everyday use of a tool that can produce plausible mistakes.
A conversation devoted to repeated humiliation has a different purpose. Applying that distinction, however, is the company’s job. Users should not need to become amateur prompt lawyers before pointing out a failure. Nor can a few sample phrases guarantee what a moderation system will do in a particular exchange.
Consider a hypothetical user correcting the same mistake for the fourth time. “You ignored my instructions again” could be an accurate description of the problem. If the assistant treats that as grounds to shut down, the user now has two failures to deal with. A product ought to be able to tolerate a complaint about its own performance.
What exactly stopped working?
Anthropic should make the outcome clear when it intervenes. Did the model refuse one answer, end the conversation or restrict access to the service? Each calls for a different response. Users need that information at the point of failure, without having to reverse-engineer it from a vague notice.
There should also be a practical way to challenge mistakes. Imagine debugging a spreadsheet while trying to work out whether the assistant has misunderstood your request or your irritation. One spreadsheet was quite enough.
The broader debate about model welfare can continue without settling either question by assertion. Anthropic’s stated uncertainty matters. A friendly conversational interface is not evidence of consciousness, and a rule about how to use a service cannot resolve whether a model can suffer. Treating the policy as proof would ask the small print to do an extraordinary amount of scientific work.
For anyone using Claude on a real project, the immediate housekeeping is less philosophical. Keep important output somewhere you control. Save enough context to reproduce a failure. If a conversation ends, you should still have the work and a clear account of what went wrong. The chat window is a precarious place to keep the only copy.
Software should withstand scrutiny.
A company introducing a boundary like this owes users enough clarity to keep challenging bad answers without guessing at an etiquette test. The rule takes effect on November 12; its application deserves the same close reading as its wording.
Meanwhile, the invented citation still needs a source.
Save the flowers for someone with a vase.
Got a different take? I’d love to hear it. hello@maisfps.com
