Is AI conscious?

No AI system has been shown to be conscious, and no agreed test exists that could show it. That second clause does most of the work on this page. The difficulty is not that the evidence points one way and the public argument points another. The difficulty is that consciousness has no settled test in humans either, so a claim about machines inherits an unsolved problem instead of creating a new one.

Three claims that get made as one#

Consciousness means having subjective experience of any kind. Sentience narrows that to the capacity to feel, and it carries the moral weight, because a system that can suffer has interests that somebody owes something to. Understanding is a claim about meaning: whether the system grasps a situation or reproduces the statistical shape of descriptions of it.

A system could in principle have any one of the three without the others. Most public argument treats them as interchangeable, and a great deal of disagreement dissolves once somebody asks which of the three is being asserted.

Twenty researchers, a method, and no verdict#

The most serious attempt at the science is a paper by twenty researchers in consciousness science and machine learning, published in Trends in Cognitive Sciences in June 2026. Their approach is to take the leading scientific theories of consciousness, including recurrent processing theory, global workspace theory, higher-order theories, predictive processing and attention schema theory, derive indicator properties from each in computational terms, and assess systems against those indicators. Graded entry.

The output is a credence and not a verdict, by design. The indicators are only as strong as the theories they come from, and those theories disagree with one another. Nothing in the method inspects a system's inner states, because nothing can.

One detail matters for anybody quoting this work. The sentence that circulates, that leading scientists concluded no current AI system is conscious, comes from the abstract of the 2023 preprint, which had nineteen authors. The peer-reviewed version of the same work has twenty, adds Tim Bayne and David Chalmers, no longer lists Chris Frith, and its abstract is framed in terms of informing credences without the flat claim. Both author lists were read at source for this page. The published full text sits behind a paywall and was not read here, so whether the claim survives inside the paper is not settled by anything on this page. The safe form of the sentence is the one at the top: no system has been shown to be conscious, and no test exists that could show it.

Chalmers puts a number on it and tells you not to trust the number#

David Chalmers gave the most cited philosophical treatment as a NeurIPS talk on 28 November 2022, published with an afterword in Boston Review in August 2023. He names six things current language models lack and treats each as a candidate requirement for consciousness: biology, senses and embodiment, world models and self models, recurrent processing, global workspace, and unified agency. Graded entry.

Giving each at least a one-in-three chance of being required puts the probability that a system lacking all six is conscious below one in ten, and the figure he settles on is confidence "somewhere under 10 percent in current LLM consciousness". For successors he combines a credence above 50 per cent that systems with those properties arrive within a decade with at least 50 per cent that such systems would be conscious, reaching "25 percent or more".

He warns against his own figures twice, calling precision here specious. Quoting the 10 per cent without the warning turns an illustration of an argument into a measurement, and no measurement of this kind exists. His afterword, written eight months later, says faster progress than expected makes the timeline possibly conservative and changes nothing fundamental in the reasoning.

He also reports a 2020 survey of professional philosophers: around 3 per cent accepted or leaned towards current AI systems being conscious against 82 per cent rejecting, and around 39 per cent accepted or leaned towards future AI systems being conscious against 27 per cent rejecting. Those figures are carried here as he reports them; the survey itself was not read for this page.

The question a person delegating work is actually asking#

Almost nobody brings this question to a decision about their own work. What they bring is a practical version of it: can this thing be trusted with the judgement I am about to hand it. That question can be answered without settling consciousness at all, and answering it is the useful skill.

The failure mode has a name on this estate and a body of evidence behind it. Fluent, confident output reads as comprehension, and automation bias describes what people do next. The reason the effect is strong here has nothing to do with minds: a model is trained on text written by people who knew what they were talking about, so it reproduces the register of somebody who knows. The estate's treatment of that sits at why AI sounds so confident.

A system asked whether it has experiences will report an answer either way, because the text it learned from contains people reporting on their own experiences. Its report is evidence about its training data. The same caution applies to the opposite report.

The practical test for understanding is cheaper than any test for consciousness and it works today. Give the system a case whose right answer depends on something absent from the text in front of it: a constraint nobody wrote down, a reason the usual approach fails here, a fact about who will read the output. Watching where it breaks tells you what to keep. That is judgement doing the work, and it does not require an answer to the harder question.

Why the question refuses to close#

Every proposed test for machine consciousness has the same shape: it observes behaviour, or it observes computation, and infers experience. With other people the inference is safe because they are built as we are and we have the same evidence about ourselves. A system built differently and trained on our descriptions of our own minds removes both supports at once. It resembles us in exactly the place the inference was drawn from, which is the reason behavioural evidence is weak here rather than strong.

So the answerable version of the question is the one the twenty authors chose: which theory-derived indicators does a given system satisfy, and what should that do to a reasonable person's credence. A reader who wants a yes or a no is asking for something the field cannot supply, and any source that supplies one has stopped reporting the state of the evidence.

Key sources

• Butlin, P., Long, R., Bayne, T. et al. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488-501. Earlier as arXiv:2308.08708. Graded entry.
• Chalmers, D. J. (2023). Could a Large Language Model Be Conscious? Boston Review, 9 August 2023, from a NeurIPS talk of 28 November 2022. Graded entry.

Related SuperSkills research#

On what a system can do rather than what it is, what is AGI and is AI dangerous. On mistaking fluency for comprehension, why AI sounds so confident and automation bias. On who answers when an unaccountable system is in the chain, the moral crumple zone. On the capability the practical question actually needs, judgement.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page takes no position on whether machine consciousness is possible, because the estate holds no evidence that would support one. What it does is separate three claims that are usually merged, record what the two most serious sources say and what they decline to say, and point at the question a person at work can actually answer. Findings are attributed to the studies that produced them and kept separate from the interpretation.