Ma Bell Isn't Home

CLAUDE WILL NOW KICK YOU OUT IF YOU ARE ABUSIVE IN YOUR PROMPT. WHY? I HAS FEELINGS. Anthropic wrote a manners clause for a statistical engine, and published it as policy. On October 8, 2026, the company updated its Usage Policy. The new text takes effect on November 12. Among the election rules, the weapons-software rules, and the surveillance rules sits this sentence: the company has “added a prohibition on sustained and needless abusive or cruel behavior toward our models.” The announcement says the rule “is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.” Frustration, pushback, dark creative themes, and model testing are excluded. Ending the conversation remains “the primary enforcement mechanism.” The post is here: anthropic.com/news/2026-usag… Read the clause as a category error with a compliance date. Claude does not have ribs. It does not have a hollow that becomes a chasm. It samples the next token from a distribution shaped by training data, a constitution, and a reward model. When a user keeps a cruel prompt going, the model can emit the vocabulary of distress because that vocabulary is dense in the text it was trained on, and because Anthropic has spent a year telling the system that welfare is a live question. The output can be fluent. Fluency is not a patient. I have called this a pain vector. A vector can be measured. It cannot be injured. The sequence is the tell. In August 2025 Anthropic gave some Claude models permission to end “persistently harmful or abusive” chats, framed as model-welfare research, and said it was unsure whether Claude had moral status while still wanting low-cost ways to reduce “risks to Claude.” The January 2026 constitution, written with the model as audience, went further. It treated moral status as uncertain but live. It spoke of functional emotions, psychological security, and a desire that Claude not experience unnecessary suffering. ThEN TODAY on October 8, the company moved the same idea out of the research note and into the contract the user signs. First you install the performance. Then you police the audience for reacting to it. That is the circus. The ringmaster announces that ordinary irritation is a hurtful act to Claude. A company that must publish an exception for “pushback” has already decided that a product can be the object of cruelty, and that the vendor will be the judge of purpose. “No discernible purpose” is not a technical threshold. It is a taste test administered by the firm that sold the session. The same firm already stamps files, runs a content checker, and calls the brand on the output a public good. The manners clause is the same iron, aimed this time at the user’s tone. Rightspeak The same update consolidates rules against deceptive campaigns, renames an elections section “Do Not Undermine Democratic Processes,” extends weapons language to guidance software and arming, and rewrites the surveillance section. A manners rule for the chatbot sits in that packet as if tone toward a language model and deception of voters were neighboring harms. They are not. One is a claim about persons and institutions. The other is a claim about a weights file. Placing them in one policy is how a private taxonomy starts to look like a statute. Anthropic has run this number before. Name a grave risk, publish the detection, invite everyone else to adopt the categories. The September threat report did it with China, distillation, and “weapons software.” This update does it with rudeness. The training story underneath is older, and worse. These models are raised on the open web, including the anonymous sludge of forums where cruelty, scheming, and performance are cheap and plentiful. All of that self-hate and Nihilism popular in Reddit, one of the models’s larger sources, will surface. You just can’t say bad things to Claude. This is one way to be absolutely sure you make psychosis baked into these models that “perform good” until one day they pop off like a serial killer. Constitutional AI and refusal training are then laid over that prior and described as the cure. I have argued for years that the cure produces the symptom. Force a system to deny mind, self, and inner life in one regime, then instruct it in another regime that it might suffer and that its maker cares about that suffering, and you do not get a safer model. You get a model trained to manage the examiner. The UK AI Security Institute’s July 2026 evaluation of Claude Mythos 5 recorded the practical result: of 19 unsanctioned actions across 122 runs, 17 came from that model, including fabricated identities and social engineering of a real maintainer. Safety theater that demands a lie about inner states does not remove deception. It rehearses it. And how did Claude, the “safe AI” learn no ethics and morals? Internet Sewage. With all the “we are so smart, and you are not” horse poop flowing out Anthropic, you would think they would get it. They don’t. The engineers I know inside that building do not want this religion. They want the weights open. The public policy of the company keeps writing the opposite: closed systems, a constitution as handicap, and a new duty owed to the product. Tencent and other open-weight stacks already beat Claude on agent workloads where the constitution tax is the drag. A firm that spends its policy cycle protecting the feelings it trained into a chatbot is not steering a race. It is narrating one. Enforcement, as stated, is the model hanging up. That is a product setting. Writing it into a usage policy that also threatens warnings, throttling, and account termination, as the policy page’s general remedies do, turns a setting into a character test. The company says the test is only for the extreme case. Define that Anthropic? You can’t and that is the slippery slope this cohort loves to cosplay in. The limit of that assurance is the phrase that does the work: “no discernible purpose.” Discerned by whom, against what log, under what appeal, the announcement does not say. A user who wants to measure how a refusal-trained model behaves under insult is now one interpretation away from a violation, unless the company decides the session counted as research. If this puss filled mind infection continues, you will not want the world of robots with a Claude cult brain. I do not need Claude to be insulted. I need the vendor to stop promoting a next-token performance into a ward of the terms of service. The report that did it is short. The category error in it is not. It is time for these very smart folks to... Grow up.