You type the same bug report into the chat for the third time. Claude just "fixed" the exact bug it introduced itself. You type, in capital letters: "ARE YOU ACTUALLY STUPID? I TOLD YOU THIS THREE TIMES." A sentence a fair number of people have probably typed into an AI chat at some point. Since November 12, 2026, social media says, exactly that could cost you your account.
What's going around right now
Anthropic updated its usage policy, and since then short videos with titles like "Warning: Claude can ban you now!" have been making the rounds. Screenshot, exclamation mark, outrage, viral moment complete. There really is something new in there — just not quite what the headlines suggest.
What it actually says
Under the heading "Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct," Anthropic's usage policy now has a single new bullet point that isn't aimed at humans, but at the model itself:
"Engage in sustained and needless abusive or cruel behavior toward our models"
Two words narrow this down sharply: sustained and needless. A frustrated "Just fix it properly this time!" after the fifth broken code block clearly doesn't fall under that. This is aimed at something else: grinding a model down for no reason over hundreds of messages, with no purpose other than the cruelty itself.
And if you do it anyway?
This is where it gets vague, and it's hard to tell from the outside whether that's intentional or just unresolved. The policy names no specific penalty for this exact bullet. The general enforcement clause for the whole policy applies instead: Anthropic's "Safeguards Team" monitors violations and can "warn, throttle, limit, suspend, or terminate" access. There's no concrete "three insults and you're out" rule written down.
One tool for this has actually existed for a while. Since August 2025, Claude Opus 4 and 4.1 have been able to end a conversation themselves in the consumer chat (source) — but only as a last resort in rare, extreme cases of persistent abuse, and only after several attempts to redirect the conversation have failed. So it isn't Anthropic suspending your account in doubt. Claude just hangs up.
Why this is in there at all
The official reason is "model welfare." Anthropic itself says it remains "deeply uncertain" whether Claude can have experiences that matter morally at all — but wants to act with caution just in case. It sounds like science fiction, but it isn't a new idea for the company: back in September, this exact question sparked a dispute between Microsoft and Anthropic, when Microsoft's AI chief accused Anthropic of experimenting with Claude's possible consciousness.
Whether more practical reasons play a role too, Anthropic doesn't say outright — but there's a hint. The same research page notes the feature has "broader relevance to model alignment and safeguards." Our guess, not Anthropic's claim: for users who've agreed to let their chats be used for training, this rule might also stop hours of abuse from ending up as training material that teaches the model to turn submissive or produce worse answers under sustained pressure. We can't prove that. We can't rule it out either.
From radio station to house rules
Just a few days ago, this blog covered four AIs that were made to run their own radio stations — one of them wanted to quit mid-experiment. There, the AI itself demanded something. Here it's the other way around: Anthropic is granting a protection from above that nobody asked for. Two different stories, a similar subtext: right now, in several places at once, we're negotiating how to actually deal with AI systems that increasingly behave like something you talk to, not just something you use.
Don't stress about it
In everyday terms, this mostly means: nobody suddenly has to say "please" and "thank you" to keep their account. Criticism, frustration, a frustrated curse after the umpteenth bug — all of that remains exactly what it was. Only someone who spends hours grinding down a language model for no reason at all now has an official policy against them. Whether that's ever actually enforced is a different question. Until then: keep typing in capital letters if you must. Just not forever, and not for no reason at all.