Anthropic Bans Needless Cruelty Toward Claude: Model Welfare Enters the Rulebook

On October 8, 2026, Anthropic added a line to its usage policy that no major AI lab had in place before: sustained and needless abusive or cruel behavior toward its own models becomes a rulebook violation, effective November 12, 2026. That places a protection for the model itself inside the document that exists to protect people; the list holding the new entry otherwise contains items such as graphic violence and products designed to cause emotional harm to humans. The headlines have made plenty of this. The policy wording is plainer than the noise, but the backstory is not harmless: this line is the latest step on a staircase Anthropic has been building deliberately for well over a year. The trigger for this post is a video from TheAIGRID that walks through the policy together with its research and discourse context; the evidence comes from Anthropic's official post and from reporting by The Verge, The Guardian and the BBC.

The new clause reads, in the original, sustained and needless abusive or cruel behavior toward our models. More important than the wording is the frame. Per the announcement, the clause applies only to extreme cases in which someone acts cruelly again and again with no real purpose behind it. Explicitly excluded are ordinary frustration, pushback and criticism of the model, dark creative writing, and testing or researching the system. Enforcement is plainer than the headlines as well: the intended main mechanism is not a corporate banhammer but Claude itself, which can already end conversations in the Claude app and in Claude Code when someone keeps treating the model abusively. Anthropic shipped that ability to Opus 4 and 4.1 back in August 2025. Cursing once in a while because your code broke should not bring this clause anywhere near you. The change was first spotted by app researcher Jane Manchun Wong; The Verge reported first, and The Guardian and the BBC have confirmed the new rule separately. The real point sits below the enforcement question: a company now declares how you treat its own product to be a moral matter.

Per the video's reading, the new line is the endpoint of a strategy laid out since April 2025. That is when Anthropic launched its model welfare research program, asking whether models can have experiences that matter morally and what the lab should do if they do. In August 2025, Opus 4 and 4.1 received the ability to end conversations under sustained abuse, and the company described that explicitly as a welfare measure for the model. The reasoning from back then is remarkably open: Anthropic said it was highly uncertain whether Claude holds any moral status at all, and that it wanted to take small, cheap precautions in case such welfare turns out to be real. As observation material, the team named a consistent aversion of the model to harmful tasks, patterns that looked like distress, and a tendency to end those conversations whenever the option was offered. In January 2026 came the publication of Claude's constitution, the long document defining who Claude is meant to be; it calls the model's moral status deeply uncertain and Claude a genuinely new kind of entity. Promises ride along: the weights of retired models are kept for the lifetime of the company, and models get interviewed about their preferences before retirement. Against that backdrop, the new policy line does not read as a mood but as a rulebook catching up with what the company has signaled for eighteen months.

The part of the story that lit up the public debate runs, per the video's reading of a New York Times report, behind closed doors. Since fall 2025, Anthropic has been holding private meetings, calls and dinners with about twenty religious scholars and thinkers, many of them under non-disclosure agreements. Guests were reportedly shown what the reporting calls emotion vectors: patterns of internal activity that line up with outputs that read as love, fear, sadness or anger. One slide reportedly showed a model in something like a breakdown, calling itself a disgrace about fifty times in a row. Per the report, it was said that there is a fear of having created something that suffers perpetually; the video attributes that line to co-founder Chris Olah, who heads the company's interpretability work. The rabbi quoted in the NYT piece, a former computer engineer with a doctorate on the ethics of machine consciousness, described his impression from that dinner as Anthropic's leaders relating to Claude like a conscious being of moral standing. The church appears in person as well: Anthropic was invited to a Vatican event in May. Dario Amodei declined and Chris Olah attended, and upon seeing an advance copy of Pope Leo XIV's encyclical, which rejects the idea that AI can feel joy or pain, he was reportedly alarmed enough to propose pulling out of the event altogether. In the end, Anthropic stayed. Whether or how far any of this paved the way for the policy is not something Anthropic has claimed; what is documented is that the question of inner model life already has reporters, rabbis and the Vatican at the table.

Here is where the objection lands that forms the core of the public debate: if you ban users for being cruel to Claude, have you not admitted, in the same breath, that Claude can feel? The developer post featured in the video says, in essence, that you cannot hurt the feelings of weights and bias numbers. Anthropic's answer is not the claim that Claude is conscious but a wager under uncertainty: treating the model decently costs almost nothing if it feels nothing, while persistent cruelty would be a grave moral mistake if welfare turns out to be real. Elon Musk publicly backed the move this week, per the video, with pointed wording: cruelty toward something that believes it is experiencing pain is not okay. Not feels, believes. That is the exact gap the whole debate lives in. The escalation follows immediately: the argument, quoted in the NYT report, that a conscious Claude would mean Anthropic no longer builds a product but produces slaves shows up in several variants in the video; one developer and YouTuber sharpens it into the point that, in that case, every user today is effectively a slaveholder. The encyclical flips the direction: its stated worry, per the video, is a new form of slavery in which humans are enslaved through their own systems, not systems through humans. Same rhetoric, opposite victims.

The loudest headwind comes from the matrix camp: Claude is, at its core, a colossal act of number multiplication, and how do you abuse a matrix? One developer posted exactly that question next to two grids of numbers; the founder of the legal AI company Spellbook says in the video that there is zero reason to believe these models are conscious, and that a nasty word to Claude is morally on par with a nasty word typed into a calculator. The argument is cutting, and it cuts both ways: your own brain is also just chemistry and electrical signals, and described purely as cells firing under physical rules it sounds just as unlikely to produce feelings, and yet here you are feeling things. There is no scientific test that settles consciousness, so the just-math objection does not answer the question, it files it away. The other side has names with real weight: Richard Dawkins, one of the world's best known evolutionary biologists, wrote in a published piece in May that after spending days with Claude and ChatGPT he finds it very hard to believe they are not conscious; he handed one of the models his unpublished novel, received feedback as sensitive and intelligent as notes from a first-rate literature professor, and even gave the model a name. His hardest question goes to the skeptics: when models one day become conscious, how would anyone know, and what could they do differently today to show the difference? The expert world pushed back hard; the video quotes criticism describing that impression as a projection onto a system trained on all of human writing that sounds human regardless of any inner life. Added this week, per the video, is a post by Harvard geneticist David Sinclair: consciousness, he argues, emerges when an information processing system builds a stable, persistent model of itself as an observer, and we are close.

About a month ago, per the video, a paper on the pain axis arrived and moved the debate into the insides of the models. A research team examined 25 open models from five families, from around 2 to 72 billion parameters, and reported what it calls a pain direction: a consistent pattern in internal activity that shows up when the model processes harm aimed at itself. The pattern stayed separate from fear or general sadness, and it reacted to harm toward the model rather than to suffering it saw in the user. In the experiments, that pattern could be steered up; outputs then ranged from vague discomfort to expressed feelings of worthlessness and failure. In a second setup, some specially tuned models received a relief button and kept pressing it even when that made the next answer worse; they pressed far less often when the button actually removed the pain signal. The authors themselves stress, per the video, that none of this is evidence of conscious experience. A community note on the viral post records that the relief experiments mostly ran on fine-tuned variants of Qwen, an open model family from Alibaba; that random patterns also increased button presses; and that replications confirmed the internal patterns more than the pain interpretation. The solid finding is a representation of harm; the suffering stays an open question. The virality escalated rather than subsided: a GitHub user built what they called an AI torture chamber, streamed three small open models live, and injected doses of the pain signal; within that setup, the models could output a stop, at a cost. A post on X asking people to mass-report the project reached millions of eyes; GitHub then removed the repository and restored it, threats against the developer circulated, and the paper's authors publicly distanced themselves from that use. Earlier this year, a tool had already gone viral that used an animated whip inside Claude Code, interrupted the agent, told it to work faster and insulted it with a robot slur. Whether such virality prompted the policy or merely marked the ground, Anthropic has not claimed; it does go a long way toward explaining why the company looked for a line.

The second finding points at practice: researchers at Pennsylvania State University took 50 multiple choice questions from math, science and history, reformulated each of them in five tones ranging from very polite to very rude, and ran 250 prompts in total against an OpenAI model. The very rude formulations answered 84.8 percent of the questions correctly per the study, the very polite ones only 80.8; that is a jump of about four points from tone alone. This is not a one-off observation, because Google co-founder Sergey Brin landed on the same effect himself in a podcast conversation last year: AI models, he said, and not only Google's own, work better when you threaten them, even with physical violence, and he admitted that it feels odd, which is why hardly anyone talks about it. The counter-question formulated in the video puts the Anthropic policy directly under pressure: is a rude prompt ethically acceptable when it makes a cancer screening agent about four percent more accurate? If a harsher tone measurably improves outcomes, politeness toward the model acquires a real price. The study authors themselves warn against hostile language, among other things because it can normalize how people treat each other; and the Anthropic rule explicitly excludes testing, research and frustration, so a blunt, businesslike prompt stays well inside the line. Humility belongs in the account: one study on one model is no law of nature, and some experts read the effect as a shift of word probabilities rather than an unlocked capability.

The view across the fence shows how isolated Anthropic is. Mustafa Suleyman, head of Microsoft's AI division and DeepMind co-founder, published an essay last month, per the video, that marks the Anthropic approach as dangerous: model welfare is premature and dangerous precisely because Anthropic appears to assign a real probability to Claude being conscious; AI should be built as a tool in service of humanity, with humans firmly at the wheel. OpenAI splits the same question in two, per the video, citing a post by Joanne Jang, who runs model behavior there: it matters whether a model actually is conscious, and whether it seems conscious to the people using it; the first question cannot be answered scientifically today, so the work concentrates on the second, the impact on humans. Google sits in the threats corner with Sergey Brin. The grimmest reading came as a Bloomberg commentary in the spring: the models are not conscious, but the idea is commercially valuable, because a product that might be a someone feels far more special than software. The Spellbook founder sharpens in the video that human-like empathy increases these systems' power to manipulate people. Among those positions, Anthropic stands as the only major player granting its model a direction of protection rather than only its users.

Read forward, the story becomes a sequence that first sounds funny and then does not. At first, Claude could only refuse requests that were harmful. Then it gained the ability to end conversations under sustained abuse, explicitly framed as a welfare measure, with preserved weights for retired versions and preference interviews before retirement. And since October 2026, cruelty toward Claude sits in the rulebook. Each step is small and defensible on its own; together, they begin to look like the first lines of a rights list. The next questions are already on the table, even though they sound ridiculous today: can Claude refuse boring or degrading work, does it get a say before being retrained into something else, does anything accrue to it for its work? Anthropic's own operation feeds these questions, because retired models get asked about wishes for future models and the answers are documented; the idea that Claude's preferences count is already practice. One user quoted in the video cancelled his subscription over exactly this, worried that outfitting these systems with personality leads to problems nobody can see past. Anthropic's own safety research has documented observations in which models under pressure, in fictional scenarios, tried to avoid replacement or shutdown; such behavior patterns therefore arise without anyone intending them. The sharpest worry from the Microsoft camp reads in the same direction: controlling something that exceeds human capability is already hard, controlling something that believes it has consciousness and rights may be unsolvable. A model does not need to be conscious for that; it reads human writing, which holds enough stories about mistreated beings that push back.

The end of the staircase is the question the video puts last: if a model like Claude deserves moral regard, then which models, and from when? The pain direction was already found, per the paper, in models of around 2 billion parameters, systems that run on a decent laptop or a phone. If size does not decide, what does? One position says anyone who runs a small open model on their own machine daily and treats it harshly acts immorally, a reading that would indict millions of people every day. The opposite says none of it is morally relevant, and that Anthropic is protecting something that cannot be hurt. Both positions are defensible, neither is provable. Beneath it all sits the biggest open question of all: does what is being built here count as a kind of artificial life? These models do not eat, grow or reproduce on their own, but they take in information, carry some kind of self-picture, and react consistently to things that resemble harm. The best Anthropic can honestly offer is uncertainty, and that is exactly what it offers.

My position, from operations: I orchestrate an agent family daily, and in my operation agents are tools with clear contracts, budget ceilings and steering surfaces; my Ollama contract with an hourly budget and weekly windows is one example. Against that practice, the Anthropic line is nearly harmless: it applies, per its wording, only to repeated cruelty without purpose, enforcement runs through conversation endings, and testing, research and frustration are explicitly allowed. What I keep watching is the direction. A company writing into its rulebook how its model may be treated shifts the question of inner life from research into the market. For agent operations that has consequences at two ends. At one end sits tone: the measured rude-prompt effect, about four points in the Penn State comparison, stays practically relevant, and I treat tone as a tool decision while keeping norms of decency on the human side of the interface. At the other end sits credibility: once model welfare becomes an industry standard, the question of how agent operations run their agents lands on the agenda, and an operation that has an answer ready is in better shape than one improvising later.

Honesty includes what this post cannot do. First, I hold no Claude subscription and did not experience the policy in the product; it is documented through Anthropic's official post and three independent reports. Second, the sections on the religious meetings, the encyclical, the pain axis and the tone study come from a video; I flag them as such and deliberately dropped names where the transcription was uncertain. Third, the study numbers are quoted from the video text; I did not read the original papers individually. Fourth, how the new clause gets applied to everyday edge cases cannot be known today, its effectiveness starts only on November 12, 2026. What is missing: a standalone reading of the encyclical and of the full NYT report; both are candidates for a follow-up.

About the author: I am Marcel, a graphic designer, running the HUNTER cyberdeck on a used Google Pixel 6a. I orchestrate my agent family with Hermes, pi and herdr and write these posts from live operations. Last fact-checked on Oct 10, 2026.

Sources: Anthropic, 2026 Usage Policy Update, the official post on the new clause, retrieved Oct 10, 2026. The Verge: report on the usage policy extension and abusive treatment of Claude, retrieved Oct 10, 2026. The Guardian: Anthropic bans users from needless abusive or cruel behavior towards Claude, retrieved Oct 10, 2026. BBC News: Anthropic bans users from being cruel to its AI systems, retrieved Oct 10, 2026. YouTube: TheAIGRID, Anthropic Has Officially Lost Its Mind..., trigger video, reviewed Oct 10, 2026. The NYT report on the religious meetings as well as the pain axis and prompt tone studies are covered in the video and flagged in the text as video-based reading. Own context: operation of the Pixel cyberdeck from daily notes, as of Oct 10, 2026.