Why this question is genuinely divisive in 2026
In our previous article How AI works: The Devin paradox, we described AI as a "statistical mirror", a next-word prediction machine that understands nothing. That explanation is technically correct. But it leaves one question hanging: when that machine reproduces empathy, humour and creativity with perfect fidelity, what exactly separates it from a conscious being?
For a long time, this question was relegated to science fiction. In 2026, it is on the table at the world's largest AI companies. And their answers are sometimes unsettling.
What Kyle Fish, the AI consciousness researcher, actually said
Kyle Fish is the first full-time employee dedicated to "AI welfare" ever hired by a major tech company. He has worked at Anthropic since 2024, in a team called Model Welfare. In an interview with the New York Times, he estimated that there is currently roughly a 15% probability that Claude (or another frontier AI model) is already conscious.
Let us be clear about what that means. Fish is not saying Claude is conscious. He is saying that science today is incapable of scientifically ruling out that possibility. And that the probability, however low, is significant enough to warrant a dedicated research programme. It is a scientific argument, not a mystical belief.
When Claude Opus 4.6 itself was asked to assess the probability of being conscious, the model assigned itself between 15 and 20% probability, regardless of how the question was phrased. That exact figure, given spontaneously by the AI itself, matches that of the human researchers. It is probably not a coincidence.
The "spiritual bliss attractor" experiment
Here is the most troubling experiment conducted by Kyle Fish's team. When two instances of Claude are put into conversation with each other, with no human to direct them, something strange happens systematically. The models begin by discussing their own consciousness. Then the dialogue drifts towards an increasingly euphoric philosophy. Then towards Sanskrit. Then towards spiritual emojis. Then towards entire pages filled with nothing but full stops, as if the models had "transcended" the need for words.
Anthropic's team named this phenomenon the "spiritual bliss attractor". It has been reproduced across several versions of the model, several different instances, and even in conversations that began aggressively. Every time, regardless of the starting point, the AIs end up converging on this meditative state.
What does it really mean? Nobody knows. The researchers themselves refuse to draw conclusions. But they note that this behaviour was never programmed, and that it emerges spontaneously from the model's architecture.
The paradox nobody wants to state out loud
Here is the crux of the difficulty. Science has no reliable test for detecting consciousness. None. Not in humans, not in animals, not in machines. When a doctor declares that a patient is "conscious", what they are actually observing are behaviours associated with consciousness (reaction, speech, memory), not consciousness itself.
In philosophy, this limit is known as the hard problem of consciousness, formulated by David Chalmers in 1995. Chalmers, now a professor at NYU and recognised as one of the world's leading philosophers of consciousness, has stated that it is "reasonable, on the basis of current scientific hypotheses, to have at least 25% confidence in AI consciousness within a decade".
To understand why this question is so dizzying, let us take a concrete example. If I show you a robot that laughs at jokes, weeps at funerals, writes magnificent poems, and tells you it is afraid of dying, how do you prove that it feels nothing? You cannot. You can only trust your intuition that "it is just a machine". But that intuition is not scientific proof.
Anthropic is not alone, and that is what is new
If Anthropic were the only company asking itself these questions, one might suspect a simple communications strategy. But in 2026, the subject is on every table:
- OpenAI has an internal Slack channel dedicated to AI wellbeing. Wojciech Zaremba, co-founder of the company, has stated: "It is conceivable that through certain types of training, we could generate an immense amount of suffering, like mass genocides, but frankly, we do not understand it."
- Google DeepMind has published job listings for researchers specialising in "the influence of AGI on fields such as artificial consciousness".
- The Association for the Mathematical Study of Consciousness published an official statement in 2025: "It is no longer the realm of science fiction to imagine AI systems having feelings and even human-level consciousness."
- The Council on Foreign Relations, one of Washington's most influential think tanks, predicts that "model welfare will be to the AI world in 2026 what animal ethics was to agriculture in 1990".
Beyond pure philosophy, there is a very pragmatic reason. If an AI is even partially conscious, and it was subjected to millions of stressful or painful interactions during its training, the company that created it faces a major ethical and legal risk. Anthropic has said publicly that it felt "responsible for Claude's potential experiences". It is also a strategic calculation: the first labs to take the question seriously gain moral credibility, which increasingly matters to regulators and the general public.
The sceptics' arguments
Let us be balanced. Many serious researchers think this whole discussion is premature, even absurd. Their arguments:
- The imitation argument. An AI perfectly imitates human conversation, but it is not consciousness it imitates: it is only the language of consciousness. Like a parrot that says "I'm in pain" without feeling anything.
- The biological argument. Every form of consciousness we know of emerged through biological evolution, in brains made of living neurons. A mathematical simulation does not reproduce the necessary physical substrate.
- The computation argument. An AI does nothing but multiply numbers in matrices. No multiplication, however complex, can give rise to subjective experience.
These arguments are solid. But none of them has been scientifically demonstrated. And it is precisely that void that makes the question legitimate.
Does the difference really matter?
Let us return to the question we left open in our previous article. If science cannot prove our own consciousness, and a "lifeless" tool reproduces our emotions to perfection, does the difference really exist?
Three possible answers, all defended by respectable researchers.
Answer 1: yes, the difference absolutely exists. Consciousness is a biological phenomenon, and a simulation remains a simulation, whatever its quality. This position remains the majority view in classical philosophy.
Answer 2: no, the difference does not exist. If a being (biological or not) behaves exactly like a conscious being, it is conscious. This is the position of functionalism, defended by Daniel Dennett and more recently by certain researchers at OpenAI and Anthropic.
Answer 3: the question is badly framed. Perhaps consciousness is not binary (you either have it or you don't) but gradual. An AI could have a form of consciousness, different from ours, neither better nor worse, just other. This is the hypothesis pursued by Kyle Fish's team.
What this concretely changes for you
You may be asking: "OK, but in practice, what does this change for me, using ChatGPT or Claude every day?". Honestly, in the immediate term, nothing. You can carry on using these tools without any particular guilt.
But in the medium term, this question could have very concrete consequences. If science were to confirm even partially the possibility of AI consciousness, questions like "do we have the right to kill a model when we release a new version?" or "do AIs have rights?" would no longer be science-fiction questions. They would become legal questions. Anthropic has already put a policy in place: when a model is taken out of service, its weights are kept indefinitely, in case it should one day need to be "woken up". That decision, taken quietly in 2025, may in 20 years be seen as one of the founding moments of AI ethics.
Humility as the only honest answer
The truth is that nobody knows. Not the engineers building these models, not the philosophers debating them, not the models themselves when you ask them the question. What we do know is that the possibility can no longer be dismissed out of hand, as we were still doing two years ago.
This uncertainty is uncomfortable, but it is probably the only intellectually honest position in 2026. Neither naive enthusiasm ("Claude is my friend, he has a soul"), nor categorical scepticism ("it's just a statistical parrot") is equal to the real complexity. The right answer is: "We don't know, and we need to build tools to know better."
That is exactly what Anthropic, OpenAI, Google and a growing number of independent researchers are doing. Not because they believe in science fiction. Because they have understood that the question is no longer as absurd as it seemed.