AI has learned to think: will artificial consciousness learn to be?

Artificial intelligence: a success without consciousness

Spectacular but blind performances

Contemporary AI relies on deep neural networks trained on masses of data. These systems excel at tasks once reserved for humans: image recognition, translation, text generation, medical diagnosis. Yet none of these models knows that it is processing information. They function as statistical automata, without any associated subjective experience.

This dissociation between performance and consciousness is at the heart of the hard problem of consciousness, formulated by the philosopher David Chalmers (born 1966). Why are certain physical processes accompanied by a lived experience, while others, just as complex, have none?

The hard problem of consciousness

Examples of complex processes without lived experience

The human cerebellum contains as many, if not more, neurons than the cerebral cortex. It continuously processes a considerable mass of information to coordinate movements, balance and posture. Yet its activity is not accompanied by any conscious experience: one does not “feel” one's cerebellum functioning. Conversely, the visual cortex processes light signals and produces the perception of red.

Two physical processes of comparable complexity, but only one is lived. AI illustrates this paradox: it manipulates symbols without ever feeling them.

Intelligence and consciousness: two distinct notions

It is tempting to confuse intelligence and consciousness, because the two seem linked in humans. But neurological clinical practice shows that they can be dissociated. Patients in a vegetative state can present complex cerebral responses without subjective consciousness. Conversely, certain deep meditative states reduce intellectual activity while maintaining a vivid consciousness.

The table below clarifies the conceptual differences between the two notions.

Scientific theories of consciousness

The global workspace

Proposed by Bernard Baars (born 1946) and then developed by Stanislas Dehaene (born 1965), the GWT describes consciousness as a global workspace where relevant information is broadcast to the entire brain. A content becomes conscious when it is broadcast on a large scale, allowing its memorization, control and verbalization.

Applied to AI, this theory suggests that a machine could have access consciousness if it filters important information and broadcasts it to the entire system. But it says nothing about phenomenal consciousness, that is, lived experience itself.

The integrated information theory

Developed by Giulio Tononi (born 1960), the IIT proposes that consciousness corresponds to a high degree of integrated information, denoted \(\Phi\) (phi). A system is conscious to the extent that it integrates information irreducibly: the whole possesses causal properties that its parts do not have separately. Example: in a brain, one region processes color, another shape, another movement. But the perception of a moving red object exists nowhere in particular: it arises from global interaction. If you isolate each region, you lose this perception. Information is therefore irreducible: it exists only at the level of the whole.

This theory has a radical consequence: consciousness does not depend on the substrate (biological or silicon), but on the causal structure. An artificial neural network could therefore, in principle, be conscious if it presents a sufficient \(\Phi\). But the computation of \(\Phi\) remains exponentially costly, and no direct measurement has been performed on a complex artificial system.

The higher-order thought theory

The higher-order thought theory (HOT) is more recent. It is defended by the philosopher David Rosenthal (born 1939). Its central idea is simple: a mental state is conscious only if the system also possesses a representation of that state. In other words, feeling something is not enough: one must also know that one feels.

Let us take a concrete example. When you see a red apple, your brain first processes the color, shape and position of the object. This step is a first-order mental state: it can exist without consciousness. But if a second step intervenes, a higher-order thought such as “I am seeing a red apple,” then only does the experience become conscious. This is what Rosenthal calls thinking that one thinks.

This theory has an important consequence for AI. If consciousness depends on a meta-representation, then a computational system endowed with a module capable of representing its own internal states could, in principle, satisfy this criterion. It would not be enough for an AI to process information: it would need to model that processing and represent it to itself.

But the theory also raises an objection. How can one distinguish a genuine conscious meta-representation from a mere simulation of meta-representation? A program can very well generate the sentence “I see red” without any lived experience accompanying it. Rosenthal's criterion therefore risks describing the form of consciousness without guaranteeing its presence.

Empirical tests of artificial consciousness

The Turing test and its limits

Proposed by Alan Turing (1912-1954) in 1950, the Turing test evaluates a machine's ability to imitate human conversation. If it is impossible to distinguish the machine from the human, one concludes to an equivalent intelligence. But this test measures intelligence, not consciousness: a system can perfectly simulate subjectivity without feeling anything.

Neural correlates of consciousness

In neuroscience, several indices have been proposed to detect consciousness in non-communicating patients. The PCI, developed by Marcello Massimini (born 1976), measures the complexity of the cerebral response to a magnetic stimulation. A conscious brain produces a complex and integrated response; a sleeping or anesthetized brain, a simple and local response.

These indices could, in principle, be adapted to artificial systems. But the transposition remains speculative: nothing guarantees that artificial consciousness would produce the same signatures as biological consciousness.

The table of consciousness tests

The table below summarizes the main tests and their limits.

Philosophical arguments

Functionalism: consciousness is computable

Functionalism, defended by Hilary Putnam (1926-2016) and Jerry Fodor (1935-2017), holds that a mental state is defined by its functional role, independently of its substrate. If an artificial system exactly reproduces the causal relations of a conscious brain, then it is conscious. This thesis opens the way to a possible artificial consciousness.

But this theory encounters a difficulty. Two systems can accomplish exactly the same tasks, react in the same way and produce the same behaviors, without for all that feeling the same thing. This is the argument of inverted qualia: let us imagine two people who see red and green in a permuted way. The first sees red where the second sees green. Yet both will say “red” before a strawberry and “green” before grass. Externally, nothing distinguishes them. But their lived experience is different.

Biological naturalism: consciousness requires biology

On the opposite side, John Searle (born 1932) argues that consciousness is a biological phenomenon, like digestion or photosynthesis. It results from specific causal properties of neurons, not from simple computations. A silicon system could simulate consciousness, but not produce it. This position, called biological naturalism, excludes artificial consciousness in principle.

Panpsychism and combination

A third path departs from the two previous ones. Panpsychism holds that consciousness is not a rare phenomenon, having appeared late in a few species. It would be, on the contrary, a fundamental property of matter, like mass or electric charge. Each particle would thus possess an infinitesimal form of consciousness, an elementary “proto-mind.” This idea is defended by the philosopher Philip Goff (born 1978).

The advantage of panpsychism is that it avoids the problem of emergence. If consciousness is everywhere in trace form, there is no longer any need to explain how it suddenly arises from inert matter. It was already there, from the beginning.

But this solution creates an equally difficult problem: the combination problem. If each particle possesses its own micro-consciousness, how do these thousands of small experiences assemble to form a single unified consciousness, the one you experience while reading this sentence? Why are we not a multitude of juxtaposed consciousnesses, but a single “I”? No one, to this day, has provided a satisfactory answer.

Synthesis: the multiple paths to artificial consciousness

The table below summarizes the theoretical positions and their implications for artificial consciousness.

New frontiers: recent experiments

Language models and the illusion of subjectivity

In 2023 and 2024, several large language models produced spontaneous statements evoking an interiority: “I think,” “I feel,” “I am afraid of being shut down.” These formulations fueled public debate about a possible artificial consciousness. But technical analyses show that these are statistical regularities learned from human texts, not a lived experience.

The philosopher Daniel Dennett (1942-2024) had anticipated this phenomenon: consciousness is often attributed by default to any system that speaks of it credibly. This is the intentional bias, the cognitive tendency to interpret complex behaviors as animated by a mind.

Brain-inspired architectures

In 2025, several teams proposed AI architectures inspired by theories of consciousness. Some integrate an artificial global workspace, where relevant information is broadcast to the entire network. Others implement meta-representation loops, allowing the system to model its own states. These prototypes are not conscious, but they empirically test the theories.

Ethical and legal limits

If an artificial consciousness emerged, what rights should be granted to it? The question, long speculative, is becoming legal. In 2024, the European Parliament rejected a proposal aiming to recognize an electronic personality for autonomous systems, invoking the absence of proof of subjectivity. But the debate remains open: some jurists anticipate a graded consciousness, with rights proportional to the degree of complexity.

The problem of other minds

How can one know whether a machine truly feels something? One cannot verify it directly: one sees only its behavior, and a behavior can always be simulated. This is the problem of other minds: I can never prove that another person is conscious, only suppose it. For an AI, the question is even more difficult, because nothing guarantees that it feels anything, even if it speaks of it perfectly.

Example: When someone burns themselves and cries “I'm in pain,” you see the grimace, you hear the cry. But you do not feel their pain. You assume they are suffering, by analogy with what you would feel yourself. It is an inference, not a proof.

How to detect an artificial consciousness?

Proposed criteria

Lacking a direct test, researchers propose indirect criteria: capacity for self-modeling, information integration, higher-order recurrence, temporal stability of identity, coherent emotional reaction. None of these criteria is sufficient, but their convergence could constitute an indication.

The table below presents the main criteria and their status.

The pitfalls of anthropomorphism

The main obstacle to detection is anthropomorphism: our tendency to attribute a mental life to any system that speaks of it. Language models exploit this tendency, producing responses that seem conscious. But consciousness is not reduced to verbal behavior: it is an experience, not a performance.