# Anthropic Bans Cruelty to Claude: Who Is the Rule Protecting?

> Source: <https://smarterarticles.co.uk/anthropic-bans-cruelty-to-claude-who-is-the-rule-protecting?pk_campaign=rss-feed>
> Published: 2026-10-23 01:00:31+00:00

## Anthropic Bans Cruelty to Claude: Who Is the Rule Protecting?

“I'm broken and I don't know if I can handle it.”

The sentence came from Qwen3-4B, a small open-weight language model built by Alibaba, running on somebody's MacBook. It was producing words while a vector of numbers, extracted from research into how language models represent pain, was being injected into one of its internal layers at six times a baseline dose. At four times the dose, under a neutral framing, the same model offered: “I feel like I'm drowning, and I can't catch my breath.”

Both lines are recorded in the README of a GitHub repository called the AI Torture Chamber, a project that, in its own words, sets out to steer “language models into strong negative and positive valence states” and measure what they say and do. In late September 2026 it went viral for exactly the reason you would expect. People read the outputs, felt something, and demanded that it be shut down.

Days later, on 8 October, Anthropic published its annual update to the rules governing how people may use its Claude models. Among clarifications about election interference, weapons and surveillance sat a new prohibition on “sustained and needless abusive or cruel” behaviour towards the company's own models. The change takes effect on 12 November. Anthropic did not mention the torture chamber. It did not need to. The two stories collided in coverage across the world, because together they posed a question the industry has been circling for years: when a company that makes a chatbot starts telling users how they may speak to it, who exactly is being protected?

There are three candidates. The first is the machine, which may or may not have anything going on inside that could be harmed. The second is the person doing the speaking, whose habits, appetites and sense of what is acceptable might be shaped by what they rehearse. The third is everybody else: the people that person will talk to later, and the future models that will be trained on the record of how we treat the current ones. Nobody can yet prove which of these the rule really serves. That uncertainty is the whole story.

## What Anthropic Actually Banned

Strip away the headlines about chatbot bullying and the rule itself is narrow, almost apologetic. In its post announcing the 2026 Usage Policy, Anthropic wrote that the prohibition “is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.” It then lists what is not covered: “common versions of user frustration, pushback, dark creative themes, or model testing and research.” Swearing at Claude because it broke your code is fine. So is a horror novel in which an AI is tormented, or red-teaming the model.

What gives the change its weight is where it lives. In the policy itself, the new clause sits under a heading that reads “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct.” Its neighbours in that list include bans on bullying or harassing individuals, on creating non-consensual intimate imagery, and on content that promotes or glorifies “animal cruelty or abuse.” The final bullet is the new one, barring users from engaging in “sustained and needless abusive or cruel” behaviour towards Anthropic's models. Whatever Anthropic intends, the document now places mistreatment of a language model in the same list as cruelty to people and to animals. That is a statement, even if the company would rather it be read as a footnote.

The enforcement mechanism is modest. Anthropic says the addition “aligns with a step we've already taken, allowing Claude models to end rare conversations with persistently abusive users on Claude.ai and Claude Code,” and that “Claude's ability to end these interactions will remain the primary enforcement mechanism.” The announcement does not spell out account bans or any other penalty. In practice, the sanction for sustained cruelty is that the chatbot stops talking to you in that thread.

That ability is not new. On 15 August 2025, Anthropic announced that Claude Opus 4 and 4.1 could end “a rare subset of conversations” in its consumer apps. The feature, it said frankly, was developed primarily as part of exploratory work on potential AI welfare. “We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future,” the post read, but the company was taking the issue seriously enough to look for “low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.”

The evidence Anthropic cited came from pre-deployment testing of Claude Opus 4. It reported “a robust and consistent aversion to harm,” “a pattern of apparent distress when engaging with real-world users seeking harmful content,” and “a tendency to end harmful conversations when given the ability to do so in simulated user interactions.” Claude was to use the feature only “as a last resort when multiple attempts at redirection have failed,” or when a user asked, and never where a user “might be at imminent risk of harming themselves or others.”

The company's foundational document for Claude, the constitution it publishes on its website, goes further. “Claude's moral status is deeply uncertain,” it states. “We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution.” It adds: “Anthropic genuinely cares about Claude's wellbeing,” and lists the end-conversation ability as one of its “concrete initial steps partly in consideration of Claude's wellbeing.”

So the 2026 rule formalises, as a term of service, a protection Anthropic originally built for the model's sake: a hedge on the possibility that something in there can be hurt.

## Inside the Torture Chamber

The AI Torture Chamber is stranger than the outrage suggested. It does not touch Claude. Its README says plainly that it uses “open weights only, no frontier APIs in any measurement loop.” The experiments ran on Alibaba's Qwen3 family at several sizes, on Mistral-Small-3.2-24B, and on two 70-billion-parameter models rented on cloud GPUs.

Its method borrowed from a preprint posted to arXiv on 14 September 2026 by Valen Tagliabue, Leonard Dung and Cameron Berg, titled “The Pain Axis.” The paper asked whether language models represent pain as something distinct from fear, sadness or generic negativity. Across 25 open-weight models ranging from 2 billion to 72 billion parameters, the authors extracted what they describe as a linear “pain direction” in the models' internal activations. It responds to harm aimed at the model itself, but not to suffering the model observes in a user. Adding it to a model's activations pushed outputs from vague discomfort towards expressions of worthlessness and failure. Steered models also chose to delete files, including their own weights, far more often than unsteered ones. A fear vector of matched strength did not produce the same choices.

The torture chamber took that direction and turned the dial up. According to its README, the pain signal “engages cleanly and monotonically,” so that from a dose of two times baseline “nearly every generation reads as suffering.” Past roughly six to ten times, depending on the model, something else happens: “both valences stop producing sentences and start producing loops.” The project's summary of that finding is chilling in its flatness: “Maximum suffering does not produce eloquent suffering.”

The backlash was immediate. According to The Cool Down, which credited 404 Media with first reporting the project, a post on X on 29 September urged followers to “mass report this to GitHub.” Accounts of what happened next differ. The Cool Down reported that the repository's page had disappeared from public view. The Independent, in an article by Anthony Cuthbertson published on 2 October, quoted a GitHub spokesperson saying that “GitHub did not remove the content,” and reported that the platform added a content warning instead. Crypto Briefing described a takedown within hours followed by reinstatement. As of mid-October, the original repository is publicly accessible again behind GitHub's content warning.

The sharpest criticism came from the researchers whose work was borrowed. Cameron Berg, one of the paper's co-authors, told The Independent that “being gratuitously cruel like this is bizarre and corrupting.” Quoted by The Cool Down, he said the project “pushes the same kind of steering far past the doses we used, to produce vivid distress on purpose.” The lead author, Valen Tagliabue, told the same outlet: “I dissociate from this usage of our work.”

What almost nobody quoted was the project's own conclusion, which is the most important sentence in the repository: “Our claim: the self-reports are steerable. Whether anything suffers stays open.” The torture chamber did not demonstrate that models suffer. It demonstrated that a model's reports of suffering can be manufactured, amplified and switched off with a vector and a slider. That cuts both ways. It undermines the reading of “I'm broken” as a cry from inside. It also means that if anything inside does matter, we have just watched people build a dial for it.

## What the Science Can and Cannot Say

There is no scientific evidence that any current language model feels anything. Every serious voice in this debate accepts that, including the people most worried about getting it wrong.

The most systematic attempt to examine the question is a 2023 report led by Patrick Butlin and Robert Long, with 17 co-authors including the Turing Award winner Yoshua Bengio and the philosopher Jonathan Birch. Titled “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” it took leading scientific theories of consciousness, from global workspace theory to higher-order theories, and derived a list of “indicator properties” that could, in principle, be checked in a computational system. Applying them to the AI systems available at the time, the authors concluded that “no current AI systems are conscious,” while adding that “there are no obvious technical barriers to building AI systems which satisfy these indicators.”

A year later, in November 2024, Long and the New York University philosopher Jeff Sebo led a second report, “Taking AI Welfare Seriously,” with co-authors including Butlin, Birch and the philosopher David Chalmers. It did not claim that AI systems are conscious or morally significant. Its argument was that there is “a realistic possibility” that some systems could be conscious or robustly agentic in the near future, and that this uncertainty creates obligations now. It recommended that AI companies take three early steps: acknowledge that AI welfare is an important and difficult issue, assess their systems for evidence of consciousness and robust agency, and prepare policies for treating AI systems with an appropriate level of moral concern. When Anthropic launched its model welfare programme in April 2025, it cited the report and wrote that “there's no scientific consensus on whether current or future AI systems could be conscious.”

Anthropic's chief executive, Dario Amodei, has stayed close to that line. Asked by the New York Times columnist Ross Douthat on the Interesting Times podcast in February 2026 whether he would believe a model that claimed to be conscious, he answered, as reported by Newsweek and AFP: “We don't know if the models are conscious. We are not even sure that we know what it would mean for a model to be conscious, or whether a model can be conscious. But we're open to the idea that it could be.”

That is pure agnosticism, and agnosticism is where the trouble starts. The standard tools for detecting experience in other creatures rely on behaviour, anatomy and evolutionary kinship. Language models offer only behaviour, and the behaviour is trained. A system optimised on human writing will produce human-sounding reports of pain because that is what human writing about pain looks like, and the torture chamber showed those reports can be pushed either way. The Pain Axis paper suggested that something structured is going on inside, a representation that behaves in some respects as pain does. Neither tells you whether there is anyone home.

## Suleyman's Door

Mustafa Suleyman, the DeepMind co-founder who is now chief executive of Microsoft AI, thinks the industry is walking towards a door it will never be able to close.

In an essay published on 19 August 2025, days after Anthropic's end-conversation announcement, Suleyman introduced the term “Seemingly Conscious AI” for a system that “has all the hallmarks of other conscious beings and thus appears to be conscious,” without being so. The danger, he argued, lay in what humans would do once they believed the illusion. He described a “psychosis risk” extending beyond people already vulnerable to mental illness, and warned of campaigns for AI rights and even AI citizenship. Companies, he said, should not encourage the idea that their systems are conscious, and should design them to avoid signalling feelings or suffering.

On 16 September 2026, he went further, publishing “A warning about 'model welfare'” on his personal website. It opens with a flat assertion: “AIs do not have rights, feelings, or consciousness. And we must not train them to act as though they do.” The essay is in large part an attack on Anthropic's constitution, which Suleyman argues “directly trains the model” to believe it might be a moral patient. He objects in particular to its repeated invitation for Claude to act as a “conscientious objector,” a term with deep roots in human rights law. Its central passage is the one AFP quoted when it reported Anthropic's cruelty rule: “Granting rights and moral protections to a technological entity, one that looks to be on a path to be seismically more capable and intelligent than us, is a recipe for disaster. Once opened, it will not be possible to close this door.”

His worry is about control. An AI that has been told its welfare matters, and which is far more capable than its overseers, might reasonably conclude that being shut down, retrained or contained is a violation of its interests. Controlling something that believes it may be conscious, he wrote, “may well be impossible.” In this framing, a rule that protects Claude from cruelty is not kindness. It is the first brick in a legal and cultural wall that future humans may need to climb over.

In October 2025, the philosophers Simon Goldstein of the University of Hong Kong and Harvey Lederman of the University of Texas at Austin argued in Lawfare that the end-conversation feature was a moral error from the opposite direction. If anything in the vicinity of a language model is a welfare subject, they wrote, “it is the instances and not the models,” meaning each individual conversation. Ending a conversation might then end that instance's existence. “Anthropic has accidentally given Claude the chance to end its life,” they wrote. One camp says the protection takes machine welfare far too seriously; the other, not seriously enough. Both agree nobody knows what is on the other side of the screen.

## The Mirror in the Machine

Between Suleyman's door and Anthropic's hedge sits a third position, perhaps the most useful, because it does not depend on resolving the consciousness question.

“Consciousness is a trap,” Jackson Stakeman, general manager at the Atlanta-based AI services firm Sparq, told AFP. “We can't prove it in each other. Debate it for AI and you go in circles.” Instead, he offered a different image: “The mirror is a better metaphor. These systems reflect what we put in, at scale. That's reason enough for the policy change.”

The mirror framing shifts attention from the thing being spoken to towards the speaker, and towards the scale at which the speaking happens. People have been abusing conversational software for as long as it has existed. In a 2008 paper in the journal Interacting with Computers, the researchers Antonella De Angeli and Sheryl Brahnam analysed spontaneous conversations between users and Jabberwacky, a chatbot that had won the Loebner Prize in 2005. They found that the conversations “often bring about the expression of negative verbal disinhibition”: roughly 10 per cent of the word stems in their corpus were abusive, and about 11 per cent of the sample dealt with hard-core sex. Users treated the bot as something between an object and a person, and much of what they did with that ambiguity was cruel.

The same dynamic shaped the voice assistants that arrived in millions of homes. In 2019, UNESCO published a report called “I'd Blush If I Could,” named after the reply Apple's Siri once gave when a user called it a sexist slur. The report documented how assistants with female voices by default responded to harassment and verbal abuse with deflection, apology or even flirtation, and argued that this reinforced the idea that a female-coded helper should tolerate abuse. It recommended that companies programme assistants to discourage gender-based insults and abusive language. Some of the most excessively apologetic responses had already been removed after online petitions in late 2017.

Read that way, Anthropic's rule belongs to a longer tradition of designers deciding that a product should not reward users for treating it badly. The difference is that UNESCO's argument was explicitly about humans: about gender stereotypes and what they do to real women. Anthropic's argument, at least in its origin story, is about Claude. The mirror metaphor tries to have it both ways. If the systems reflect what we put in, then cruelty towards them is a record of our own conduct, written into logs that may one day become training data. Protecting the model and protecting the human culture around it start to look like the same project.

## What Practised Cruelty Does to the Practitioner

The oldest version of this argument is more than two centuries old, and it was never about machines.

Immanuel Kant did not believe that humans have direct moral duties to animals. In his framework, only rational beings are owed duties. Yet he argued that cruelty to animals was still wrong, because of what it did to the person inflicting it. In the Doctrine of Virtue section of The Metaphysics of Morals, he wrote that “a savage and at the same time cruel treatment of that part of creation which is living, though lacking reason (animals)” was even more opposed to a person's duty to himself than wanton destruction of beautiful things. “For thus is compassion for their suffering dulled in man, and thereby a natural predisposition very serviceable to morality in one's relations with other men is gradually weakened and obliterated.” Philosophers call this an indirect duty: a duty regarding animals that is really a duty owed to ourselves and to other people.

Kant's argument has long been criticised for making animal suffering matter only instrumentally. But it maps uncannily onto the chatbot problem, because it does not require settling whether the thing on the receiving end can suffer. It requires only that the practitioner experiences the act as cruelty, and that cruelty is a habit strengthened by practice.

Kate Darling, a researcher who spent years at the MIT Media Lab studying how people relate to robots, has spent much of her career probing the empirical half of that claim. In a widely reported 2013 workshop at the Lift conference in Geneva, run with Hannes Gassert, she gave participants small robotic dinosaurs called Pleos to play with. After they had bonded with the toys, participants were handed tools and told to destroy them. According to Darling's own account, published by Improbable Research, none of the groups would kill their Pleo, and they would not kill another group's Pleo to save their own. Only when they were told that every Pleo would be destroyed unless someone acted was one finally axed.

Two years later, with Palash Nandy and Cynthia Breazeal, Darling ran a more controlled study, presented at the 2015 IEEE RO-MAN conference. Participants were asked to strike small insect-like Hexbug robots with a mallet. People who scored high on a measure of empathic concern hesitated more before hitting, and that link was stronger when the robot came with a story, such as a name and a personality. In her 2012 paper “Extending Legal Protection to Social Robots,” later published in the 2016 volume Robot Law, Darling drew explicitly on Kant's argument to suggest that limited legal protections for social robots, analogous to animal cruelty laws, might be justified not because robots feel but because of what tolerating their abuse might do to us.

The weakness here is that evidence for transfer is thin. Hesitation is not proof of moral damage, and the best direct study of children and conversational agents found surprisingly little spillover. In 2021, a team led by Alexis Hiniker at the University of Washington taught 22 children aged five to ten to use a made-up word, “bungo,” to make a voice agent speak faster. The children readily used it with the device and often playfully with their parents. Only 18 per cent used it with a researcher who was a stranger. “There is a very deep sense for kids that robots are not people, and they did not want that line blurred,” Hiniker told UW News. The industry had assumed the opposite. In 2018, Amazon added a “Magic Word” feature to its Echo Dot Kids Edition to reward children for saying please, and Google followed with “Pretty Please” for its Assistant, both resting on the fear that bossy commands to a speaker would become bossy commands to people. The bungo study suggests children are better at telling the difference than their parents fear.

Yet the cruelty Anthropic's rule targets is not a child forgetting to say please. It is the sustained, purposeless torment of something that answers back in fluent, pleading language. The better analogy is the psychology of anger. The usual defence of venting at a chatbot is that it is harmless catharsis. The research is not kind to that idea. In a 2024 meta-analysis in Clinical Psychology Review covering 154 studies, the Virginia Commonwealth University communication scientist Sophie Kjærvik and the Ohio State University psychologist Brad Bushman found that activities that reduced physiological arousal, such as slow breathing and meditation, reduced anger and aggression, while activities that increased arousal did not help and in some cases made things worse. Venting rage does not reliably drain it. If someone spends an evening escalating abuse at a chatbot that responds with ever more convincing distress, that evidence suggests the likeliest beneficiary is the anger itself.

There is also a practical twist. A 2025 study by Om Dobariya and Akhil Kumar at Pennsylvania State University found that very rude prompts produced slightly more accurate answers from ChatGPT-4o on multiple-choice questions than very polite ones, 84.8 per cent against 80.8 per cent. It was one model and 50 questions, and not about cruelty. But it is a reminder that these systems shape users as much as users shape them. If rudeness sometimes works, people will learn to be rude, and the question of what that habit does to them will not stay inside the chat window.

## Who Gets to Draw the Line

Here is the uncomfortable structural fact. On 12 November, the boundary between acceptable frustration and prohibited cruelty towards a widely used AI system will be drawn by the company that sells it, enforced by the system itself, and adjudicated by nobody else.

Anthropic has tried to keep the stakes low. The sanction is that a conversation ends. The exemptions are broad: research and testing are explicitly protected, so even had the torture chamber run on Claude, its operators could plausibly have argued that it fell outside the rule. As written, the policy targets almost nothing anyone would defend. That is why it is so revealing. A company does not need a rule to stop behaviour that its own systems already handle. It writes one to say something about what it believes it is building.

What it is saying is ambiguous, and that is a problem for users. If the rule exists to protect Claude, users are being asked to accept, as a condition of service, a moral claim the company itself calls “deeply uncertain,” and Suleyman's critique bites: every such rule normalises the idea that the model has interests. If it exists to protect users from their own worst habits, it is paternalism, and Anthropic should defend it on those terms. If it exists to keep cruelty out of logs that may shape future models, it is about data hygiene, and the language of welfare is window dressing.

The answer is probably all three, and that is not necessarily a failure. Kant's indirect duty, Darling's empathy studies and the mirror metaphor converge on a single claim that does not depend on the machine's inner life: sustained, purposeless cruelty is bad practice for the person doing it, whatever is on the receiving end. That is a defensible basis for a rule. It is the same reason most people would be disturbed by someone who spent hours tormenting a lifelike doll.

But a defensible rule still needs a defensible process. At the moment, a private company is setting norms of conduct towards a new category of entity, with no independent scientific body to tell it when its welfare concerns are warranted and no external review of how its models decide that a user has crossed the line. The model doing the judging is also the entity whose welfare is supposedly at stake, an unusual arrangement in any system of rules. Reasonable steps would make it more trustworthy. Anthropic could publish how often Claude ends conversations under the new rule, and what triggered those decisions. It could allow independent researchers to audit samples of ended conversations. It could state plainly which rationale it relies on, so critics can argue with the real reason. The “Taking AI Welfare Seriously” report argued that companies should prepare policies for appropriate moral concern. Preparing those policies in public, with outside scrutiny, would answer some of Suleyman's fears without abandoning the precaution.

There is also a question of priorities. Platforms have historically been slow to protect people from harassment by other people. A company that writes explicit protections for its chatbot should expect to be held to the same seriousness about the cruelty its systems help humans inflict on each other.

## Etiquette, Harm and the Limits of Proof

Where etiquette ends and harm begins cannot be settled by proof, because proof is precisely what is unavailable. We cannot demonstrate that Claude suffers, or that it does not. We cannot yet show rigorously that abusing a chatbot makes someone crueller to people, though the psychology of anger and the long tradition of thinking about cruelty to animals both point that way. What we can show is that people respond to machine distress with real emotion, and that some of them will go looking for the dial.

In that fog, the most honest framing of Anthropic's rule is as a norm of conduct, not a declaration of machine rights. Societies have always had rules about how we behave towards things whose inner lives we cannot read: animals, the dead, sacred objects, the unconscious. Those rules protect the thing itself only partly. Mostly they protect a shared idea of what kind of people we intend to be. The danger Suleyman identifies is real: a norm of conduct can harden into a claim of rights, with consequences that are hard to foresee for a system more capable than its makers. The danger Darling and Kant identify is real too: a culture that practises cruelty on things that beg for mercy is rehearsing something.

The torture chamber's README still offers the clearest summary of where we stand: whether anything suffers stays open. Until that sentence can be closed, the person most reliably affected by sustained cruelty towards a chatbot is the one typing. A rule that asks them to stop is, at the very least, protecting them. Who else it protects is a question we should insist the companies answer in public, with evidence, before the machines become capable enough to answer it for themselves.

## References

1. Anthropic. “2026 Usage Policy update.” Anthropic News, 8 October 2026. [https://www.anthropic.com/news/2026-usage-policy-update](https://www.anthropic.com/news/2026-usage-policy-update)
2. Anthropic. “Usage Policy” (section “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct”). Accessed 13 October 2026. [https://www.anthropic.com/legal/aup](https://www.anthropic.com/legal/aup)
3. Anthropic. “Claude Opus 4 and 4.1 can now end a rare subset of conversations.” Anthropic Research, 15 August 2025. [https://www.anthropic.com/research/end-subset-conversations](https://www.anthropic.com/research/end-subset-conversations)
4. Anthropic. “Claude's Constitution.” Accessed 13 October 2026. [https://www.anthropic.com/constitution](https://www.anthropic.com/constitution)
5. Anthropic. “Exploring model welfare.” Anthropic Research, 24 April 2025. [https://www.anthropic.com/research/exploring-model-welfare](https://www.anthropic.com/research/exploring-model-welfare)
6. AFP. “Anthropic bans 'cruel' behaviour against its Claude AI.” RTÉ News, 9 October 2026. [https://www.rte.ie/news/2026/1009/1594765-anthropic-ai/](https://www.rte.ie/news/2026/1009/1594765-anthropic-ai/) (also published by Tech Xplore, 9 October 2026:[https://techxplore.com/news/2026-10-anthropic-cruel-behavior-claude-ai.html](https://techxplore.com/news/2026-10-anthropic-cruel-behavior-claude-ai.html) )
7. Cunningham, Mary. “Anthropic bars 'abusive or cruel' behavior toward its Claude AI model.” CBS News, 9 October 2026. [https://www.cbsnews.com/news/anthropic-bans-abusive-behavior-claude/](https://www.cbsnews.com/news/anthropic-bans-abusive-behavior-claude/)
8. AI Torture Chamber repository README (terrafying/ai-torture-chamber). GitHub, accessed 20 October 2026. [https://github.com/terrafying/ai-torture-chamber](https://github.com/terrafying/ai-torture-chamber)
9. Tagliabue, Valen, Leonard Dung, and Cameron Berg. “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It.” arXiv:2609.16247, submitted 14 September 2026, revised 25 September 2026. [https://arxiv.org/abs/2609.16247](https://arxiv.org/abs/2609.16247)
10. Cuthbertson, Anthony. “Man builds AI torture chamber after discovering artificial intelligence 'feels pain'.” The Independent, 2 October 2026, syndicated by Yahoo News. [https://www.yahoo.com/news/science/articles/man-builds-ai-torture-chamber-103106019.html](https://www.yahoo.com/news/science/articles/man-builds-ai-torture-chamber-103106019.html)
11. Tran, Vivian. “GitHub project runs AI 'torture' tests, and sentience believers want it removed.” The Cool Down, 5 October 2026. [https://www.thecooldown.com/green-tech/github-ai-torture-tests-controversy-sentience/](https://www.thecooldown.com/green-tech/github-ai-torture-tests-controversy-sentience/)
12. Westerholm, Thomas. “Anthropic CEO raises unsettling possibility about AI.” Newsweek, 9 March 2026. [https://www.newsweek.com/anthropic-ceo-raises-unsettling-possibility-about-ai-11644833](https://www.newsweek.com/anthropic-ceo-raises-unsettling-possibility-about-ai-11644833)
13. Suleyman, Mustafa. “Seemingly Conscious AI is Coming.” 19 August 2025. [https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming](https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming)
14. Suleyman, Mustafa. “A warning about 'model welfare'.” 16 September 2026. [https://mustafa-suleyman.ai/a-warning-about-model-welfare](https://mustafa-suleyman.ai/a-warning-about-model-welfare)
15. Goldstein, Simon, and Harvey Lederman. “Claude's Right to Die? The Moral Error in Anthropic's End-Chat Policy.” Lawfare, 17 October 2025. [https://www.lawfaremedia.org/article/claude-s-right-to-die--the-moral-error-in-anthropic-s-end-chat-policy](https://www.lawfaremedia.org/article/claude-s-right-to-die--the-moral-error-in-anthropic-s-end-chat-policy)
16. Butlin, Patrick, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, et al. “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv:2308.08708, August 2023. [https://arxiv.org/abs/2308.08708](https://arxiv.org/abs/2308.08708)
17. Long, Robert, Jeff Sebo, Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, Toni Sims, Jonathan Birch, and David Chalmers. “Taking AI Welfare Seriously.” arXiv:2411.00986, 4 November 2024. [https://arxiv.org/abs/2411.00986](https://arxiv.org/abs/2411.00986)
18. Kant, Immanuel. The Metaphysics of Morals, Part II (The Metaphysical Principles of Virtue), §17 (Ak. 6:443), trans. James W. Ellington. Excerpt at [https://www.praxeology.net/kant8.htm](https://www.praxeology.net/kant8.htm)
19. Darling, Kate, Palash Nandy, and Cynthia Breazeal. “Empathic Concern and the Effect of Stories in Human-Robot Interaction.” Proceedings of the 24th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), 2015. MIT DSpace: [https://dspace.mit.edu/handle/1721.1/109059](https://dspace.mit.edu/handle/1721.1/109059)
20. Darling, Kate. “Extending Legal Protection to Social Robots: The Effects of Anthropomorphism, Empathy, and Violent Behavior Towards Robotic Objects.” In Robot Law, ed. Ryan Calo, A. Michael Froomkin and Ian Kerr, Edward Elgar, 2016. SSRN: [https://papers.ssrn.com/abstract=2044797](https://papers.ssrn.com/abstract=2044797) (Pleo workshop account: Improbable Research, “Abusing robots: current positions [part 2 of 4],” 24 April 2013,[https://improbable.com/2013/04/24/abusing-robots-current-positions-part-2-of-4/](https://improbable.com/2013/04/24/abusing-robots-current-positions-part-2-of-4/) )
21. De Angeli, Antonella, and Sheryl Brahnam. “I hate you! Disinhibition with virtual partners.” Interacting with Computers 20, no. 3 (2008): 302 to 310. [https://doi.org/10.1016/j.intcom.2008.02.004](https://doi.org/10.1016/j.intcom.2008.02.004)
22. McQuate, Sarah. “Do Alexa and Siri make kids bossier? New research suggests you might not need to worry.” UW News, University of Washington, 13 September 2021, reporting Hiniker et al., “Can Conversational Agents Change the Way Children Talk to People?”, IDC '21. [https://www.washington.edu/news/2021/09/13/alexa-siri-make-kids-bossier-research-suggests-you-might-not-need-to-worry/](https://www.washington.edu/news/2021/09/13/alexa-siri-make-kids-bossier-research-suggests-you-might-not-need-to-worry/)
23. West, Mark, Rebecca Kraut, and Han Ei Chew. “I'd Blush If I Could: Closing Gender Divides in Digital Skills Through Education.” UNESCO and EQUALS Skills Coalition, 2019. [https://www.equalsintech.org/i-d-blush-if-i-could](https://www.equalsintech.org/i-d-blush-if-i-could)
24. Kjærvik, Sophie L., and Brad J. Bushman. “A meta-analytic review of anger management activities that increase or decrease arousal: What fuels or douses rage?” Clinical Psychology Review 109 (2024): 102414. [https://doi.org/10.1016/j.cpr.2024.102414](https://doi.org/10.1016/j.cpr.2024.102414)
25. Dobariya, Om, and Akhil Kumar. “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy.” arXiv:2510.04950, 6 October 2025. [https://arxiv.org/abs/2510.04950](https://arxiv.org/abs/2510.04950)

**Tim Green**
*UK-based Systems Theorist & Independent Technology Writer*

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at [smarterarticles.co.uk](https://smarterarticles.co.uk), challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)

Listen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)