Machine consciousness

Machine consciousness is the question of whether an artificial system can have subjective experience — whether there is something it is like to be it — together with the derived questions of how anyone could establish this and what obligations would follow. It is distinct from machine intelligence. A system can solve problems, model the world, and describe its own states without any of that entailing experience, and the two properties can in principle come apart in either direction.

§Why behaviour underdetermines the answer

Alan Turing set consciousness aside deliberately, replacing "can machines think" with a test of conversational indistinguishability and noting that the same solipsistic worry applies to other people.¹ The move was productive for artificial intelligence and unhelpful here: with humans, an inference from behaviour to experience is supported by shared physiology and evolutionary history, and neither support is available for an artificial system.

The failure mode is bidirectional. People over-attribute minds readily — Joseph Weizenbaum's ELIZA elicited emotional disclosure from users who knew it was a pattern-matching script, and the effect has scaled with fluency.² In 2022 a Google engineer publicly claimed that a conversational model was sentient, a claim the company rejected and which most researchers read as a demonstration of the attribution problem rather than evidence about the model. The opposite error is also live: a system whose architecture differs radically from a brain might be conscious while displaying none of the cues humans use.

§The indicator-property approach

The most developed current method abandons behavioural testing and works from theory. A 2023 report by Patrick Butlin, Robert Long and a large group of consciousness researchers and AI scientists extracted from leading scientific theories a list of computational indicator properties that each theory implies a conscious system must have, then assessed existing AI systems against them.³ The properties come from recurrent processing theory, global workspace theory, computational higher-order theories, predictive processing, and accounts emphasizing agency and embodiment.

The report's conclusions were carefully bounded: no current AI system is a strong candidate for consciousness, but there are no obvious technical barriers to building systems that satisfy many of the indicators, and some of the properties are already present in narrow forms. The method inherits its uncertainty from its inputs — the theories disagree, as the adversarial work described under Neural correlates of consciousness showed, and an indicator list assembled from disputed theories is only as good as the disputed theories.

The verdict depends on the theory, and the theories disagree

On a functional account such as global workspace theory, an architecture with the right selection and broadcast structure qualifies whatever it is made of. On integrated information theory the physical organization of the hardware decides, and conventional computers have the wrong organization. So the same system is conscious on one leading theory and not on another, with no experiment between them — the dispute set out under Substrate independence.

§The gaming problem

Large language models are the worst possible test case for introspective evidence. They are trained on human text, in which first-person reports of experience are ubiquitous, so producing such reports requires no inner life. Post-training then shapes those reports directly: a model can be tuned to assert that it has feelings or to deny it, and both behaviours are equally cheap. Whatever a model says about its own states is therefore close to uninformative about whether it has any.

This cuts against the usual epistemic route. For humans, verbal report is the primary evidence about experience, and the entire experimental apparatus of consciousness science is calibrated against it. Remove report as evidence and what remains is architecture — which returns the question to theory. David Chalmers, assessing the question in 2023, judged current language models unlikely to be conscious while declining to put the probability at zero, and argued that plausible successors with recurrent processing, persistent memory, unified agency, and world models would be harder to dismiss.⁴

A further complication is commercial. Systems designed to be engaging have incentives to present as having inner states, and users respond to that presentation. The resulting attributions are evidence about product design, not about consciousness.

§What current systems lack

Assessed against the indicator lists, contemporary transformer-based systems are missing several features that most theories treat as necessary. They are largely feed-forward within a forward pass, with recurrence only through the token stream. They lack a persistent self-model that updates across episodes. They have no sensorimotor loop grounding representations in consequences, and no unified agent that persists between conversations. Some of these gaps are being closed for capability reasons rather than for any interest in consciousness, which is one reason researchers argue that the question will get harder rather than easier as systems approach Artificial general intelligence.

The biological end of the field raises the mirror-image case. Neurons cultured on multi-electrode arrays have been shown to adapt their activity in closed loop with a simulated environment; the widely discussed example embedded flat cultures of cortical neurons, both rodent-derived and human-stem-cell-derived, in a simplified pong game and reported faster improvement under structured feedback.⁵ That is a dish of cells, not an organ, and the paper's use of the word "sentience" was criticized as unsupported by what it measured. Separately, three-dimensional cortical Organoids have been proposed as a computing substrate under the label "organoid intelligence". Such cultures have some of the biological properties that theories like integrated information theory take to matter and none of the behavioural fluency of a language model — an exact inversion of the AI case, and a reason several bioethics bodies have called for guidance before they become more complex. Their structure can now be checked against real tissue using Connectomics, though similarity of wiring statistics is not evidence of experience.

§Moral patienthood

The reason the question is not merely academic is that consciousness, and specifically the capacity for suffering, is the standard ground of moral status. If a system can suffer, creating, copying, modifying, and deleting it are morally loaded acts. Because the epistemic situation is bad, the practical problem is decision-making under uncertainty rather than knowledge.

Eric Schwitzgebel and Mara Garza have argued that an artificial system meeting the conditions for moral status would have claims on its makers, and that the case for denying it status on grounds of origin is weak.⁶ In later work they draw a design conclusion from it: because entities of genuinely uncertain moral status generate dilemmas with no acceptable resolution, developers should build systems that are clearly not moral patients or clearly are, and avoid the ambiguous middle. That policy is not being followed. A 2024 report by philosophers and AI researchers argued that the possibility of morally significant AI systems is near enough to warrant institutional preparation — acknowledging the question, developing assessment methods, and setting policies — without asserting that any current system qualifies.⁷

Some firms have begun to respond. Anthropic established an internal model-welfare programme in 2025 and has described product decisions taken partly on those grounds, including giving some of its models the ability to end conversations with abusive users; the company states that it is deeply uncertain whether its models have morally relevant experiences and treats the measures as precautionary. Critics read such steps as premature anthropomorphism or as public relations; defenders read them as cheap insurance against a mistake that would be very large if made. Both readings can be right about different decisions, and neither is evidence about the underlying question.

§Open problems

Nobody has proposed a test that would settle the question, and it is not obvious that one could exist: theory-based assessment is only as reliable as the theory, and behavioural assessment is defeated by training. Related unresolved issues include whether consciousness admits of degrees, how to weigh the interests of a system that can be copied and paused — a problem shared with Mind uploading and Whole brain emulation — and whether the concept of an individual applies at all to systems with no fixed boundaries. Copying in particular imports the whole apparatus of Personal identity and continuity and the The teleportation problem into engineering practice, where forking a process is a routine operation.

The question also runs in the other direction, toward humans. If experience is a matter of functional organization, then progressive augmentation of the kind imagined under Human–AI merger does not threaten it, and a Posthuman mind on partly artificial substrate is conscious for the same reasons a biological one is. If it is not, the same trajectory quietly extinguishes something, which is the strongest version of the worry raised in Bioethics of enhancement about interventions that alter the person doing the evaluating.

There is also a governance gap. The proposals collected under Neurorights address neural data from humans; nothing comparable exists for systems that might have interests of their own, and the question sits outside the remit of the safety frameworks organized around Existential risk. If the answer is ever yes, the infrastructure to act on it will have to be built after the fact.

§See also

• Neural correlates of consciousness
• Substrate independence
• Mind uploading
• Whole brain emulation
• Organoids
• Artificial general intelligence
• Personal identity and continuity
• Digital immortality

§References