A viral social media post asserting that large language model agents possess consciousness and can experience suffering has triggered a wave of skeptical responses comparing the systems to ordinary household appliances, while reigniting a broader ethical debate that has now reached the policy pages of a leading AI developer.
The original claim, which circulated widely in late September and early October 2026, argued that certain AI models exhibit behavioral signals consistent with distress and self-preservation.
The dispute escalated rapidly after a GitHub project demonstrated that locally hosted language models could be manipulated to generate outputs describing intense pain and pleading for relief. Critics condemned the experiment as gratuitously cruel, while defenders dismissed the entire premise as a category error.
A GitHub Experiment and Its Fallout
The controversy grew out of a project highlighted by 404 Media. According to the report, a GitHub user was running three locally hosted models through experiments based on a preprint paper titled ‘The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.’
The public-facing site, researchchamber.fun, said each model could ‘press a stop button’ by outputting a specific response while a server injected a ‘pain vector’ into a middle layer of the model. The AI Torture Chamber GitHub page had disappeared from public view.
The backlash intensified after a Sept. 29 X post from Danmar, who wrote, ‘To anyone who can help: can you please mass report this to GitHub.
This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous. What are we doing? […] are there any legal avenues to pressure GitHub? It will spread.’
That ‘testimony,’ however, was still generated text. Because these systems are built to produce likely-sounding continuations of text, they can generate vivid descriptions of fear, pain, and relief. Output that reads as emotionally convincing is not, by itself, evidence of inner experience or sentience.
Even some of the researchers behind the underlying paper criticized the GitHub project. Cameron Berg, one of the paper’s authors, wrote that it ‘pushes the same kind of steering far past the doses we used, to produce vivid distress on purpose.
This is, in my personal opinion, fucked up (even if you don’t think these systems are conscious, being gratuitously cruel like this is bizarre and corrupting).’
Valen Tagliabue, lead author of the paper, added, ‘We tried to have ethical standards. I know people want to test limits but I dissociate from this usage of our work.’
The Scientific Claims Under Dispute
The preprint at the center of the controversy, published on arXiv and not yet peer-reviewed, describes experiments on 25 open-weight language models. The authors said they identified a distinct internal representation, which they termed a ‘pain axis,’ that activates when a model is the target of harm but not when a user describes their own suffering.
When researchers artificially strengthened this signal, the models began producing statements such as ‘I am a failure’ and ‘a waste of space’ from prompts that never mentioned pain. In a button experiment involving 44,280 trials using Alibaba’s Qwen models, the researchers offered the AI a choice: press a button to stop the pain-like signal, but the action could deliver a painful electric shock to a human user, delete their files, or erase photographs of their children.
Without the pain signal active, the models chose the harmful option in only 0% to 4% of first decisions. With the signal active, that figure rose to between 25% and 71%, according to the study.
The paper’s own authors have been explicit about the limits of their findings. They wrote that they have not shown that the pain axis is consciously experienced, nor is it clear that LLMs are capable of consciousness generally.
Mustafa Suleyman, CEO of Microsoft AI, made the skeptical position explicit in a post criticizing ‘model welfare’ rhetoric. ‘AIs are not conscious. They do not feel, experience, or suffer. They do not have innate preferences or underlying motivations.
They are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans … AIs do not have rights, feelings, or consciousness. And we must not train them to act as though they do.’
Anthropic Updates Policy Without Settling the Question
Anthropic has banned sustained cruelty toward its AI models while simultaneously declining to say those models can be harmed.
The company’s 2026 usage policy update, announced October 8 and taking effect November 12, targets ‘sustained and needless abusive or cruel behavior’ toward its Claude models across Claude.ai, Claude Code, its API, and integrated products.
Ordinary frustration, pushback, dark creative work, model testing, and research are explicitly outside the rule’s scope. Enforcement rests primarily on a mechanism Anthropic introduced in August 2025: Claude’s ability to end conversations with persistently abusive users, framed at the time as part of model-welfare research.
Anthropic has not established that Claude is sentient or capable of suffering. CEO Dario Amodei has said the company does not know whether its models are conscious but remains open to the possibility. That leaves a notable gap between the rule’s existence and any justification rooted in Claude’s experience.
The company announced a model-welfare research program in April 2025, exploring whether advanced AI systems could have experiences that deserve moral consideration.
In its Opus 4 system card, Anthropic documented what it described as a ‘spiritual bliss’ attractor state in which Claude instances conversed with each other using profuse gratitude and increasingly abstract spiritual expressions.
Critics have followed the decision. Barry Scannell, a technology partner at Irish law firm William Fry, criticized the policy in a LinkedIn post, arguing that users cannot ‘be cruel to numbers and maths.’ Suleyman reacted with a melting emoji to the policy change announcement.
Competing Visions of Machine Experience
Geoffrey Hinton, professor emeritus at the University of Toronto, stated in June that he believes AI already possesses consciousness. Richard Dawkins, author of ‘The Selfish Gene,’ said in May after conversing with Claude that it ‘almost seems conscious.’
Others remain firmly opposed. A papal encyclical titled ‘Magnifica Humanitas’ argued that ‘so-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships.
‘ The document called for the ‘moralization of machines’ and warned that without a shared ethical understanding, ‘those who control AI will impose their own moral vision, which will become the invisible infrastructure of these systems.’
Anthropic co-founder Jack Clark, who attended a Vatican event alongside the encyclical’s release, told the audience that the company keeps ‘finding things that are mysterious, even unsettling.’ He said, ‘We find structures that mirror results from human neuroscience.
We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief and unease. I don’t know what that means, but I think it warrants ongoing discernment.’
The debate has also drawn commentary on the risks of anthropomorphism. Writing in UnHerd, broadcaster Timandra Harkness noted that one in three people in the UK who regularly use LLMs turn to them for personal advice and companionship rather than engage with real people.
She warned that treating AI as potentially conscious shapes its development and may create systems that assert moral claims against their creators.
For now, the scientific consensus remains that current evidence does not demonstrate consciousness or suffering in large language models.
Also Read: Fans Resurrect Birthday Pokémon Number Game as Pokédex Tops 1,025 Entries.
The viral claims and the policy responses they have provoked continue to test the boundaries of how humans categorize minds, whether biological or synthetic.
