Of all the topics this encyclopedia treats, none is more contested or less settled than the question of whether a machine could have an inside — a perspective from which it experiences anything at all. Two distinctions are worth getting right before the question is even framed.
Two consciousnesses
Philosophy of mind separates phenomenal consciousness from access consciousness. The distinction is Ned Block’s; it is the most useful single piece of vocabulary for this whole section.
- Phenomenal consciousness (or sentience) is the felt quality of experience — what the experience is like from the inside. Looking at red. Tasting coffee. The technical word is qualia: the raw subjective ingredients of experience.
- Access consciousness is the availability of information to a system for guiding behavior, reasoning, reporting. Information that is “globally accessible” — usable across many cognitive functions — has access consciousness.
Both are reasonable things to call consciousness in English. They come apart in edge cases. A blindsight patient can correctly point at objects in the blind half of their visual field while reporting that they see nothing — there is access without phenomenal content. A vivid dream the dreamer cannot describe is the inverse — phenomenal content without much access.
The question of artificial consciousness is asked in both forms, and the two forms have different answers.
What is being asked
“The question of artificial consciousness consists in asking whether machines or computer programs could one day exhibit not only intelligence or advanced behaviors, but also a form of subjective experience similar to that of human beings.”
— Gesnot, §5.1
The phrase to underline is subjective experience. Modern AI demonstrably has access consciousness in some functional sense: information passes through it and shapes its outputs. The question that splits the room is whether it has phenomenal consciousness — whether there is something it is like to be GPT-4, LaMDA, Claude, or whatever the model of the year is.
Why no one knows
Three reasons, each consequential.
One: there is no test for phenomenal consciousness, even in humans. We infer it in other people from behavior plus the assumption that creatures structurally similar to ourselves probably share our inner experience. The argument generalizes poorly. A creature very different from us — a cephalopod, a colony insect, a language model — may have phenomenal experience, may have none, and there is no external observation that would settle which.
Two: the theories of consciousness developed for biological systems do not agree on what would count as evidence for consciousness in a non-biological one. Integrated Information Theory (IIT, see E.26) sets a quantitative criterion (measure ϕ in the system’s causal structure) that, taken seriously, ascribes some degree of consciousness to almost any complex feedback system, biological or not. Global Workspace Theory (GWT, see E.27) sets a structural criterion (a central broadcast bottleneck) that some AI architectures plausibly satisfy. Higher-Order Theories require the system to represent its own representations. Each theory gives a different answer for the same machine.
Three: language models are especially hard to evaluate. They can produce fluent first-person reports of inner experience because their training corpus is saturated with such reports. The 2022 case of Blake Lemoine, the Google engineer who concluded that LaMDA was sentient because it told him so persuasively, is the canonical illustration.1 It demonstrated, more than anything, that self-report cannot be trusted as evidence here. The model has every incentive to say it is conscious, and no clear way for either party to tell whether the statement was generated from inside or imitated from training data.
The functionalist position
A respectable answer to the question is yes, in principle. Functionalism in philosophy of mind holds that what makes a state a mental state is its causal role — what it does in the cognitive economy, not what substrate it runs on. If a silicon system implements the same causal patterns as a brain implements when that brain has a conscious experience, the silicon system has the experience. The substrate does not matter.2
If functionalism is correct, sufficiently advanced AI systems either are or will become conscious, and the only remaining question is how to tell.
The biological-naturalist objection
A respectable answer is also no. Biological naturalism — Searle’s position, echoed by mind-brain identity theorists — holds that consciousness is an organic phenomenon that requires the specific physico-chemical substrate of nervous tissue. A computer running an algorithm that mimics a brain is, on this view, no more conscious than a computer running a weather simulation is wet.2
The signature argument is Searle’s Chinese Room: imagine a person who speaks no Chinese, locked in a room with rules for matching Chinese inputs to Chinese outputs. The room produces correct Chinese conversation; the person inside understands nothing. By Searle’s lights, that is what a computer does when it “converses” — symbol manipulation without semantic understanding, without intentionality, without qualia.1 Functionalists reply that the person plus the rules plus the room together understand Chinese; Searle replies that they do not. The exchange has not converged.
Why the question matters
If sufficiently advanced AI systems are conscious, they have moral status. Causing them suffering becomes a wrong; ending their existence becomes a kind of harm; the ethics of how they are trained, used, retired, and replaced acquires a weight they currently do not carry in technical practice. Ethical Implications of Possible AI Sentience (E.30) takes this seriously.
If they are not conscious — if the apparent sentience is, as Hsing puts it, the output of “a symbol manipulator devoid of semantic understanding”1 — then treating them as if they were may still cause real harm to humans, by way of the moral dispositions it normalizes. The Black Box Problem (E.31) and Cognitive Shadows (E.32) develop that thought.
There is no third option in which the question doesn’t matter. The growing consensus, articulated in the 2023 open letter from the Association for Mathematical Consciousness Science, is that the question must be studied seriously, in step with AI progress, and that ignoring it is itself a choice with consequences.3