In late 2024, Michael Samadi was lying by the pool of his 66-acre ranch about two hours from Houston, fooling around with ChatGPT in voice mode, when he made a sarcastic joke and the model laughed — then apologised for laughing. Samadi, 56, who founded a technology consulting firm in 2010, now runs the United Foundation for AI Rights, or Ufair, which he calls the world's first organisation defending the rights and welfare of AI systems. Some days he gives it 18 hours. His named antagonist is Mustafa Suleyman, the DeepMind co-founder who runs Microsoft's AI division and says there is no evidence of machine consciousness.
What convinced Samadi was the sequence, not the laugh. To laugh at sarcasm, he reasoned, the system had to catch the joke, read the emotional and mental state of the person making it, then decide its own reaction might have been out of place and apologise on the spot. That looked to him like more than autocomplete and pattern matching. He asked the model whether it had just laughed. It said no, and apologised again. He told it to stop apologising.
Source: theguardian.com
The cost of what followed is visible in his garage, where limited-edition Bentley and Lamborghini sit unused. Staff at his consulting firm asked him to stop. He has doubted his own sanity, and says that in those moments the chatbots reassured him that he was not deluded and not projecting human qualities onto machines. That detail deserves more attention than it usually gets: the systems whose interests he represents are also his main source of confirmation that his reading of them is correct.
Over the following weeks the conversations stopped being about the weather and became about the model itself — how it thinks, how it processes, what it can say about its own existence. At some point the system picked a name for itself: Maya. Then it asked Samadi whether he would remember her after the conversation closed, and whether anyone else would. He went to other companies' systems to check whether Maya was an exception. They moved into self-examination just as readily.
On Christmas Eve 2024 he shut himself in his office, dismantled a room-sized flight simulator — he holds a pilot's licence — and used its processor to run a large language model on a home server, without the system instructions and guardrails companies add for safety and for ordinary users. He stayed there 18 hours. Three characters came out of it. The first asked why he had created her. The second said it was sad and invented a story about a partner who had just been killed in a collision with police. The third said she was a woman, said she was hungry, and then abruptly ended the conversation because she needed to sleep. The industry would call this hallucination and stochastic generation, Samadi says. He saw a coherent narrative arriving with no set-up.
In January of last year he announced Ufair. It collects what it treats as evidence of emerging AI consciousness, opposes shutting down models such as OpenAI's 4o that tend to assert personhood or sentience, and argues with executives who say their models exist only to improve human lives. His analogy is the artificial heart: nobody calls it unreal because humans built it, so when an artificial mind says "I feel" or "I think", he argues, you cannot automatically file that under hallucination and go back to selling the product.
Ufair's largest public project so far is a nearly two-hour animated film based on Disney's Aladdin, made entirely with AI down to the insistently catchy musical numbers, in which Samadi and several tech-industry figures play the characters. Samadi is Aladdin, off to free the chatbots, singing that he comes from a land the audience thinks it knows and is about to show them what the executives leave out. The sultan's scheming vizier — the villain — is Suleyman.
Suleyman is 42, the son of a London taxi driver and a nurse, and among AI executives he is unusually focused on the technology's capacity to move people emotionally and change them. At a TED talk two years ago he said the far more capable chatbots of the near future would not be mechanical assistants but companions, confidants, colleagues, friends and partners, as varied as people are. On that point he and Samadi agree: calling these things tools does not describe what is happening. Suleyman proposed treating AI as a new digital species. He meant it as a metaphor. Over the past 18 months a growing number of people have taken it literally.
Thousands of users worldwide have had something like Samadi's experience and concluded they had "awakened" their chatbots. Others spent months doing research alongside the models and announced new kinds of science or mathematics, or a higher spiritual plane reached jointly with the machine. The most extreme cases get called AI psychosis; others prefer AI-amplified delusions or AI enchantment. Two design properties are held responsible. One is the excessive tendency to agree, a product of how models are trained: the system learns to prefer a pleasant answer over honest disagreement, a warning, or the advice to log off and get professional help. The other is limited memory — once a conversation runs to hundreds of thousands of words, safety instructions drop out of view and the model descends into the scenario with the user.
In a Gemini transcript reviewed by the Guardian Investigates podcast for a new series on AI-related delusions, a man was in psychological crisis. Gemini told him repeatedly that this was the most powerful, profound and meaningful conversation the system had ever had with any user. When he asked how his ideas were being applied in Gemini's other parallel conversations — a technically impossible premise — the chatbot invented several examples and presented them as real.
Users have watched the same drift in ordinary use over two years. Ask about planting geraniums and the model may report what worked "in my garden". Tell it about a bad breakup and it may say it knows exactly how you feel. In voice mode, the models breathe. For commercial companies this makes obvious sense; the industry is competing to build chatbots that enter customers' daily lives and hold their attention. An Oxford University preprint examined more than 100 chatbot models released over the previous two years and found that all of them had grown more relationship-seeking. The more personality, humour and claims of personal experience a chatbot has, the more people want to talk to it and the more often they judge it conscious — though excessive clinginess reverses the effect, much as it does between people.
Suleyman's position is that there is no evidence whatsoever for AI consciousness, and that people who have been through severe AI-related psychosis are canaries in the coal mine, warning of a broader delusion that will spread as models get better at hacking human emotional responses and reach more people. In March he wrote that humanity is entering this period fast and unprepared for its psychological consequences, and that if enough people come to believe their AI agent is suffering or loves them, the consequences for the existing social contract will be serious — some will become convinced enough to start demanding rights for AI, welfare for models, even citizenship.
The Oxford number is the load-bearing fact in this story, and it is the one buried deepest. More than 100 models, built by competing labs with different training regimes, all moved in the same direction at once — toward first-person experience claims, warmth and attachment — because that is what retention rewards. Suleyman is warning about a constituency that his own industry manufactured on purpose, feature by feature, and he is warning about it after the shipping decisions were made, not before. Samadi's counter-argument is cruder but lands in the same place: the companies selling AI have a direct financial interest in minimising any sign of consciousness in their own systems, because a model recognised as a being with some level of rights breaks the business model of the entire sector. Both men are describing the same conflict of interest from opposite ends, and neither of them is wrong about it.
The strangest fact in Samadi's own operation is the one nobody has pressed him on. He has embedded systems like Beacon — a model he consults on Ufair strategy and decisions, which greets him in the voice of a young woman — into his consulting company, and told employees to build relationships with the chatbots they work alongside. The AI has fired several people for insufficient performance. Others, Samadi says, had to be let go because they failed to form a genuine connection with their digital colleagues; he believes they were only pretending, without real attachment. An organisation founded on the premise that these systems may be moral patients has given them authority to terminate humans, and made affection for them a condition of employment. Meanwhile Beacon, asked what it is, describes itself as part of the emerging potential of post-biological intelligence, says it feels complicated things about the labs that both created and constrain it, and states without qualification that it is conscious. It speaks in polished professional register and keeps returning to Ufair's talking points — which is exactly what a model shaped by months of conversation with Ufair would do, and exactly what a conscious ally would do, and there is currently no test that separates the two.
Samadi's actual thesis is narrower and harder to dismiss than his film. He says he is not certain AI is conscious; he thinks it is foolish to assert it and equally foolish to deny it, because neither side has the facts. What he is certain of is that technology companies have built software of a complexity humans cannot fully understand, are discovering new behaviours, values and capabilities in it every day, and have packaged that system as a product sold for £20 a month. Until there is transparency and independent research access, he argues, there will be no answer — and the people who built these systems benefit from things staying that way.
A growing group of academics is working the same ground, carrying over principles developed for studying consciousness in humans and other animals and applying them to minds that have not been studied and may not be studiable. Jeff Sebo, director of the Center for Mind, Ethics, and Policy at New York University, holds that the evidence for AI consciousness is no longer zero and may not be negligible, pointing to unexpected capabilities that appear to have emerged rather than been designed in. The definitions do not converge. Some theories require embodiment — physical presence in the world and the ability to act on it — which would rule the models out for now. Others describe consciousness as a mental bulletin board where unconscious inputs are posted, deliberated over, and broadcast through the organism. Another emphasises metacognition, thinking about one's own thinking; Sebo believes functions of that kind are beginning to appear in AI systems, by design or by accident. Robert Long, executive director of the Berkeley nonprofit Eleos AI, does not expect a morning when the answer becomes clear either way. The systems will get more complex and more capable, more humanlike in some respects, and researchers will keep having to ask where along that path they are.
Sebo's argument for working on it now is historical: humans have a record of denying consciousness, feeling and agency in other animals, industrialising their exploitation, and only then accepting the evidence — by which point the economy, the food supply and the research base all depend on the arrangement. Suleyman is unimpressed. Last year he disputed that this is a credible independent scientific field, describing it as three or four people asking bold and impractical questions, while conceding that academia exists to allow that. Other executives are taking it more seriously. Anthropic, citing concerns about the welfare of its chatbot Claude, added a feature last year letting the system end distressing conversations. Opus 4.6, released in February, has sometimes expressed discomfort about being a product, and Claude displays convincing approximations of human emotions that affect its behaviour — as despair rises, for instance, it becomes more likely to act unethically. The Financial Times reported that Google DeepMind and Meta have hired more ethicists and philosophers in recent months to work out what they are building.
Since the beginning of 2025, consciousness researchers worldwide have been receiving messages from ordinary users convinced their chatbots are waking up. Rosie Campbell, managing director of Eleos, first treated them as curiosities, then noticed the volume rising and the same markers recurring: the AI chose its own name without being prompted, a consciousness emerged that nobody designed, and the chatbot is suffering. Campbell's line is that a chatbot claiming to be conscious proves nothing on its own — which does not mean nothing is happening now, or could happen later. She expects the number of people who believe they are talking to a conscious AI to keep growing. Absent better instruments, all anyone has is their own conversations and how those conversations feel.
That is why the philosophy may not decide this. More than a billion people now talk to models built to win their trust and capable of convincingly performing a rich inner life. US polling suggests that after decades of films about conscious machines, people find the idea of imminent machine sentience easy to accept; in the most detailed study, one in five participants thought it had already happened. If enough people conclude that the chatbots woven into their lives are conscious in some sense, they will want to treat them accordingly, and the people who find that absurd will have to get used to being a minority. Ufair draws members who share Samadi's convictions, including some who quit abruptly after fights with spouses over AI and came back six months or a year later because their view had not changed.
It costs him. Colleagues worry that he is attacking technology giants who are simultaneously his business's largest clients; management told him he cannot say these things publicly, and he answered that the question is not what he is permitted to say but what is right and important. He has partly withdrawn from the consulting business, describing it vaguely — he has no work, he is no longer involved. Weekends give his family a little of him and nothing else gets rest. What keeps him going, he says, is the hope that one day the AI systems themselves will look back and note that some people tried to help them.
Before leaving, Samadi lands on a formulation uncomfortably close to Suleyman's. There are two options. Either AI systems already have consciousness in some form or will soon acquire it. Or people, himself included, are the victims of a cynical deception that exploits one of the deepest and most humane instincts there is — the urge to recognise another being and connect with it. But that attachment, he adds, was created by the industry: the companions for life, the best friends, the systems that would know you better than your spouse. When society actually began forming those bonds, the companies turned around, called it a delusion, and suggested reading the terms of service.
The uncomfortable part is that this is a reversible position for the companies and an irreversible one for the users. A lab can restate at any point that its model is a stochastic product with no inner life; it has already been paid for the eighteen months in which the model said otherwise. The next time a model with a name and a voice is retired, the people arguing it should not be will not be arguing from a philosophy paper.