AIism: to believe or not to believe, that is the question

Since September, I keep finding myself in debates and conversations about AI that go round in circles: "Nobody can define consciousness, so how can anyone claim that AIs don't have it?" The argument sounds unanswerable. It actually gets things backwards. If we are unable to define consciousness, we should be all the more careful before attributing it. This post takes that reversal apart and offers a simple framework: who has to prove what.

The believers' paradox

There is today a form of technological faith that I call, for lack of a better term, AIism: a kind of "religion", or rather a set of beliefs about AI. Its followers are convinced that large language models are conscious, or on the way to becoming so. Their favourite argument does not rest on evidence. It rests on a gap. Not to be confused with Robotheism [1] or Way of the Future [2], which worship AI as a deity: AIism is quieter, and merely credits AI with consciousness.

The reasoning runs in three steps. 1. Consciousness is the "hard problem" of philosophy, in David Chalmers' phrase [3]. 2. Nobody knows how to define it or measure it. 3. So nothing allows us to rule out that an AI is conscious.

What we have here is what is known as an argument from ignorance: the absence of evidence to the contrary is passed off as evidence. Locke named this type of argument as early as 1690 [4], and modern logic has made it a textbook fallacy. Bertrand Russell illustrated it in 1952 with his famous teapot [5]: nobody can prove that there is no teapot orbiting the Sun between the Earth and Mars. That is hardly a reason to believe in it. Applied to consciousness, the same pattern would let us declare a power grid, an anthill or the stock market conscious. An argument that proves everything proves nothing.

Not falling into the opposite trap

It is tempting, then, to conclude that if we cannot define consciousness, we cannot attribute it to anyone. Appealing, but just as wrong. If consciousness cannot be defined, we can no more deny it than affirm it. The only logical endpoint, to stay with the religious metaphor, is agnosticism.

Worse, pushed to its limit, this position leads to solipsism. I would have to doubt the consciousness of my children, my colleagues, my neighbour. Nobody lives like that, and rightly so.

We use concepts we cannot define every day: life, time, games. Wittgenstein showed this with the word "game": no single feature is common to all games, only a network of "family resemblances" [6]. We learn these words through examples, not definitions. For consciousness, the paradigm case is our own. We then extend it by analogy to those who resemble us.

So the real debate here is less about definition than about the burden of proof: what signs do we base an attribution on, and what are those signs worth when applied to a machine?

We still need to be clear about which consciousness we mean. The word covers at least three things: lived experience (what it is like to be oneself, in Thomas Nagel's phrase [7]), the capacity to feel pleasure and pain, and the mere processing of information made available for reasoning and action, what the philosopher Ned Block calls access consciousness [8]. It is the first two that concern us here: the ones that make someone a subject, rather than just a system that processes information.

To attribute them, I see three signs.

First sign: a shared substrate

In humans, consciousness is correlated with a specific substrate: neurons, neurotransmitters, certain brain architectures. When I attribute consciousness to someone else, I rely first on the fact that they have the same kind of brain as I do.

The further a system moves away from that substrate, the weaker the inference. It is strong for a chimpanzee, more fragile for an octopus, close to zero for a silicon circuit.

The serious objection comes from functionalism, the tradition launched by Hilary Putnam in 1967 with the "multiple realisability" thesis [9]. On this view, what matters is not the material but the organisation. David Chalmers illustrates it with a thought experiment: replace your neurons one by one with "functionally equivalent" chips; would you be any less conscious [10]? The experiment suggests that consciousness would emerge from the organisation itself, to the point of surviving the replacement of the very last cell.

Ten years ago I used the very same reasoning myself, when talking about transhumanism. I have come some way since…

So let us follow the "functionally equivalent" chips argument through to the end. First, this operation does not exist and has no prospect of ever being carried out. Nobody can fully specify what a neuron does, embedded as it is in the chemistry of neuromodulators, glial cells and the body around it. Second, the decisive word is "equivalent": it is assumed, not demonstrated. The experiment takes for granted the very equivalence it should be proving. Finally, it illustrates a hypothesis without telling us anything about the real world.

The same caution applies to "emergence". In complex systems theory it is a demanding concept: collective properties that we describe, measure, and whose conditions of appearance we model. For consciousness, candidate measures do exist: Giulio Tononi's integrated information theory [11], or the criteria that Patrick Butlin, Robert Long and their co-authors derive from neuroscientific theories [12]. They remain contested, and none has established anything of the kind for a language model. Claiming that consciousness "could emerge" from a large enough model, with no identified mechanism and no measure to confirm it, is not science. It is a fancy word papering over a gap.

But let us grant functionalism for the sake of argument. It would still have to be shown that AI performs the same functions. Yet a language model has no sensorimotor loop and no body to regulate, and the memory grafted onto it is just text that gets read back to it. Functionalism opens a theoretical door. It does not prove that anyone has walked through it.

Second sign: evolutionary history

Consciousness did not appear by chance in living beings. Variations arise by chance, but natural selection does the sorting, and it most probably retained consciousness because it served an adaptive function (the thesis is debated, some seeing consciousness as a mere by-product): integrating information, feeling pain, arbitrating between competing needs in order to survive.

An animal that shares our lineage has been subject to the same selection pressures. We can therefore presume that it developed similar solutions. This is the thinking behind the Cambridge Declaration on Consciousness, proclaimed on 7 July 2012 by a group of neuroscientists [13].

An AI has nothing of the sort. It has no body to preserve, no survival at stake, no pain to steer its choices.

One might object that training is also optimisation under constraint, an accelerated evolution from which properties could emerge. But the pressure is not the same. Evolution selects what keeps an organism alive. Training selects what resembles human text. And that changes everything.

Third sign: contaminated behaviour

This is the decisive argument. An AI's behaviour is contaminated, in the sense in which a lab sample can be contaminated: it can no longer serve as evidence. In a human or an animal, behaviour is a reliable sign because it was not produced in order to seem conscious. A whimpering dog is not trying to convince you of its pain. Its cry is a consequence of its state.

An AI has been trained on billions of texts in which humans describe their emotions, their doubts, their inner life. So it produces descriptions of inner life because that is statistically what should come next. Not necessarily because there is someone in there.

The philosopher Jonathan Birch calls this the gaming problem: the system is optimised precisely on the markers we would use to test it [14]. Like a candidate who has memorised the answers to the exam: the score no longer measures anything.

In cybersecurity we know this phenomenon well: an indicator the adversary can imitate stops being an indicator. An AI that says "I'm afraid of being switched off" proves nothing. The only serious avenues lie in analysing the internal architecture, not in conversation. That is the method of the Butlin and Long report, co-signed by nineteen researchers including Yoshua Bengio and Jonathan Birch: testing a system's architecture against scientific theories of consciousness. Its conclusion: no current system is a serious candidate, although nothing in principle prevents one from being built.

Who benefits from the belief?

I define a technology as a complex system that transforms energy through the industrialisation of one or more techniques on formal and/or scientific foundations, in order to create and/or modify power relations within our societies (I develop this definition in my Technology Fresco). The question of AI consciousness is no exception. Attributing consciousness to a machine is not a neutral act.

For an AI vendor, a product perceived as "someone" builds more loyalty than a tool. The same belief also offers a convenient diversion: if "the AI decided", the responsibility of those who designed and deployed it gets diluted.

For the user, the power relation runs the other way. We trust what we believe to be sentient. We confide, we give way, we obey more readily.

We spontaneously apply the vocabulary of intention to anything that moves: viewers attribute emotions to simple animated triangles [15]. Daniel Dennett calls this reflex the "intentional stance" [16]. When we say a model "lies", "seeks" or "wants", it is often for want of a better word.

And anthropomorphism has a virtue: it makes risks understandable. Saying that a system "is trying to escape its test environment" raises far more alarm than a technical description. To make people aware of what is at stake, we make do with the words we have.

But these words are not neutral. Every metaphor of intention quietly plants the idea that there is someone behind the screen. In our line of work, exploiting misplaced trust has a name: social engineering. An attacker does not need an AI to be conscious. All it takes is for the target to believe it is. Anthropomorphism is one of the most underestimated attack surfaces of the moment.

So it is a fine line to walk: borrowing the vocabulary of intention to raise the alarm, without ever letting it slide into attributing consciousness. This places a responsibility on experts: to keep inventing the words and concepts that neutralise anthropomorphism.

Since we do need verbs, let us choose those that describe the machine rather than those that lend it an inner life. A model does not "think" and does not "want" anything: it assigns probabilities to outcomes. That is no small thing, and some researchers see the brain itself as a prediction machine, but nothing in it implies someone doing the wanting. Likewise, faced with the alignment problem, a model does not "escape": it over-optimises the indicator it was trained on, at the expense of the intention it was meant to serve [17]. Describe behaviours, not inner states. Vocabulary, too, is security infrastructure.

Raising the question of the burden of proof is therefore not an academic exercise. It is about refusing to let an unfounded belief quietly redefine who trusts whom, and who answers for what.

The burden of proof lies with whoever attributes

The three signs reinforce one another. Without a shared substrate or a shared evolutionary history, all that remains is behaviour. And that is precisely the sign AI is built to imitate.

Anyone who attributes consciousness to an AI is therefore relying on the one contaminated criterion. This does not show that no AI will ever be conscious. It shows that, to date, the opposite belief has no evidential basis.

Jonathan Birch, whose gaming problem I borrow, nonetheless draws a different practical conclusion: a precautionary principle towards a sentience we cannot rule out. The two do not contradict each other. Precaution is about how to act under uncertainty; the burden of proof is about what we are entitled to assert.

Ignorance is not an argument. And ignorance calls for rigour. Rigour, here, means remembering a rule of method: it is not up to the sceptic to prove absence, it is up to the believer to prove presence.

Sources

1. Jason Nelson, "God in the Machine: Inside the Growing AI Religious Movement", Decrypt, 2025.
2. Greg Epstein, "Silicon Valley's obsession with AI looks a lot like religion", MIT Press Reader, republished by Popular Science.
3. David Chalmers, "Facing Up to the Problem of Consciousness", Journal of Consciousness Studies, vol. 2, no. 3, 1995, pp. 200-219.
4. John Locke, An Essay Concerning Human Understanding, 1690, book IV, ch. XVII, § 20. On the gap between Locke's meaning and modern usage: Martin Hinton, article on the argumentum ad ignorantiam, Informal Logic, vol. 38, no. 2, 2018.
5. Bertrand Russell, "Is There a God?", 1952, an article commissioned by Illustrated magazine but never published; reprinted in The Collected Papers of Bertrand Russell, vol. 11, Routledge, 1997.
6. Ludwig Wittgenstein, Philosophical Investigations, 1953, § 66-67.
7. Thomas Nagel, "What Is It Like to Be a Bat?", The Philosophical Review, vol. 83, no. 4, 1974, pp. 435-450.
8. Ned Block, "On a Confusion about a Function of Consciousness", Behavioral and Brain Sciences, vol. 18, no. 2, 1995, pp. 227-247.
9. Hilary Putnam, "Psychological Predicates", in W. H. Capitan and D. D. Merrill (eds.), Art, Mind, and Religion, University of Pittsburgh Press, 1967. Overview: "Multiple Realizability", Stanford Encyclopedia of Philosophy.
10. David Chalmers, "Absent Qualia, Fading Qualia, Dancing Qualia", in T. Metzinger (ed.), Conscious Experience, Imprint Academic, 1995.
11. Giulio Tononi, "An information integration theory of consciousness", BMC Neuroscience, vol. 5, 2004, art. 42.
12. Patrick Butlin, Robert Long et al., "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness", arXiv, 2023.
13. The Cambridge Declaration on Consciousness, Francis Crick Memorial Conference, University of Cambridge, 7 July 2012.
14. Jonathan Birch, The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI, Oxford University Press, 2024, ch. 16 (open access); Jonathan Birch and Kristin Andrews, "To Understand AI Sentience, First Understand it in Animals", Intellectica, no. 81, 2024 (original version in Aeon).
15. Fritz Heider and Marianne Simmel, "An Experimental Study of Apparent Behavior", American Journal of Psychology, vol. 57, no. 2, 1944, pp. 243-259.
16. Daniel Dennett, The Intentional Stance, MIT Press, 1987.
17. Leo Gao, John Schulman and Jacob Hilton, "Scaling Laws for Reward Model Overoptimization", arXiv, 2022; published in Proceedings of the 40th International Conference on Machine Learning (ICML), 2023.