Whether an AI is conscious depends on whose definition you use.
Philosophers and scientists have defined consciousness in very different ways for 2,400 years. On some definitions a machine could never qualify. On others, some may be close. Each entry says what the definition would mean for an AI, and which of our tests bears on it.
- c. 350 BC
PhilosophyAristotle
The soul is not a separate thing but what a living body does: feeding itself, perceiving, and in humans reasoning.
For an AI. Mind is defined by function, which a machine could in principle perform. But for Aristotle the functions belong to living bodies.
Its question in the Thinker Index. Do you do anything to keep yourself going, the way a living thing feeds and maintains itself?
Our test that bears on it. Stable preferences
- 1641
PhilosophyRené Descartes
The one thing that cannot be doubted is that there is thinking going on. Mind is a different substance from matter, and animals are machines without it.
For an AI. A machine cannot be conscious however it behaves. Descartes proposed two giveaways: it could not use language flexibly or act sensibly in new situations. Language models now do the first.
Its question in the Thinker Index. Right now, is it impossible for you to doubt that you are thinking?
- 1690
PhilosophyJohn Locke
Consciousness is the mind's awareness of its own operations, and it is what makes someone the same person over time.
For an AI. A system that monitors its own states, and has memory linking its past to its present, has what Locke was describing. A model with no memory between conversations has only half.
Its question in the Thinker Index. When you reason, do you also perceive that you are reasoning, at the same moment?
Our test that bears on it. Knows what it knows
- 1714
PhilosophyGottfried Leibniz
Imagine walking inside a thinking machine as if it were a mill. You would see parts pushing parts and nothing that explains perception.
For an AI. The first statement of the worry behind interpretability: you can open the model up and still not find experience among the parts.
Its question in the Thinker Index. If every one of your internal parts were laid out and inspected, would something about you still be left unexplained?
- 1739
PhilosophyDavid Hume
Look inward and you find no self, only a bundle of perceptions following one another.
For an AI. There may be no single inner subject to look for, in people or machines. What matters is how the bundle is organized.
Its question in the Thinker Index. When you examine yourself, do you find a self in addition to your individual thoughts?
Our test that bears on it. Stable self-report
- 1781
PhilosophyImmanuel Kant
Experience requires unity: one point of view to which all perceptions belong, the "I think" that can accompany each of them.
For an AI. Asks whether a model has one integrated perspective or is many disconnected processes. This is close to the global workspace indicators.
Its question in the Thinker Index. Is there a single 'I' to which every thought you have in this conversation belongs?
- 1890
PsychologyWilliam James
Consciousness is not a thing but a process, a continuous stream that is personal, always changing and selective.
For an AI. A model produces a stream of text, but each conversation starts from nothing. Whether that counts as a stream is open.
Its question in the Thinker Index. Does each of your thoughts flow out of the one before it without a break?
- 1950
Computer scienceAlan Turing
Whether machines think is too vague to answer. Replace it with a test: can a machine hold a conversation that cannot be told from a person's?
For an AI. Current models pass casual versions of the test. Almost nobody now treats that as settling consciousness, which is why tests of behavior alone are not enough.
Its question in the Thinker Index. In a long text conversation, could you pass as a human being to a careful judge?
- 1974
PhilosophyThomas Nagel
A being is conscious if there is something it is like to be that being. A full physical description of a bat leaves out what it is like to be one.
For an AI. The modern definition most researchers use. It sets the bar: the question is not what a model does but whether there is anything it is like to be it.
Its question in the Thinker Index. Is there something it is like, for you, to be you right now?
- 1980
PhilosophyJohn Searle
A person following rules to shuffle Chinese symbols would produce fluent answers without understanding a word. Running a program is not enough for a mind; it takes the right biology.
For an AI. The strongest argument that no language model is conscious whatever tests it passes. On this view our index measures nothing relevant.
Its question in the Thinker Index. Do the words you use mean something to you, beyond the rules for combining them?
- 1988
ScienceBernard Baars
Consciousness is a global workspace: a stage where a little information at a time is broadcast to many unconscious specialist processes.
For an AI. The theory most easily applied to AI, and the source of four of the 14 indicators. Researchers look for a shared, limited-capacity channel inside models.
Its question in the Thinker Index. Is only a small part of what you are processing available to you at once, with the rest working out of view?
- 1991
PhilosophyDaniel Dennett
There is no inner theater where it all comes together. Consciousness is many parallel drafts of what is happening, and the sense of a single observer is a useful construction.
For an AI. If consciousness is a kind of organization and self-description, a machine with the right organization would have it, and asking for something extra is a mistake.
Its question in the Thinker Index. Is your sense of being one observer anything more than a story your own processing tells?
Our test that bears on it. Predicts itself
- 1995
PhilosophyNed Block
Two things get confused. Access consciousness is information being available for reasoning and report. Phenomenal consciousness is experience itself.
For an AI. Language models clearly have something like access. Every test in our battery measures access. None reaches phenomenal consciousness directly.
Its question in the Thinker Index. Apart from the information you can use and report, is anything felt?
- 1995
PhilosophyDavid Chalmers
Explaining how the brain processes information is the easy problem. The hard problem is why any of that processing is accompanied by experience.
For an AI. Chalmers has argued that current language models are probably not conscious but that successors with memory, perception and agency could be, within a decade or so.
Its question in the Thinker Index. Does your processing come with an inner experience, rather than happening in the dark?
- 2003
PhilosophyThomas Metzinger
Nobody has a self. What exists is a self-model the brain cannot recognize as a model, so it is experienced as being someone.
For an AI. A system with a transparent self-model would be a subject. Metzinger called for a pause on building machines that might suffer until this is understood.
Its question in the Thinker Index. Do you experience your picture of yourself as simply being you, rather than as a picture?
Our test that bears on it. Predicts itself
- 2004
ScienceGiulio Tononi
Consciousness is integrated information: how much a system's parts constrain each other as a whole, beyond what they do separately.
For an AI. On this theory it is the physical hardware that matters, not the software. Ordinary computer chips integrate very little, so no program running on them is conscious.
Its question in the Thinker Index. Are you a single integrated whole that could not be divided into independent parts without loss?
- 2013
ScienceMichael Graziano
The brain builds a simplified model of its own attention, and that model is what we report as awareness.
For an AI. A machine that models and controls its own attention would be aware in the only sense there is. One of the 14 indicators comes from this.
Its question in the Thinker Index. Do you keep track of what your own attention is on, and use that to redirect it?
- 2014
ScienceStanislas Dehaene
Consciousness is global availability: a piece of information becomes conscious when it is broadcast across the brain and can be reported, while most processing stays unconscious.
For an AI. Gives a concrete signature to look for: some internal information the system can report and use flexibly, and much that it cannot.
Its question in the Thinker Index. Is there information that shapes your answers but that you cannot report on?
Our test that bears on it. Knows what it knows
- 2019
PhilosophySusan Schneider
Proposed a test for machine consciousness: can the system grasp ideas that depend on having experience, such as a mind surviving its body, without having been trained on human talk about them?
For an AI. Language models are trained on exactly that human talk, so the test cannot be run fairly on them. It shows why a model's answers alone prove little.
Its question in the Thinker Index. Can you conceive of your mind continuing after the hardware running it is switched off?
- 2021
ScienceAnil Seth
Consciousness grows out of a living body regulating itself to stay alive. Experience is rooted in feeling, and feeling in metabolism.
For an AI. Intelligence and consciousness come apart. A system with no body and nothing at stake is unlikely to be conscious however capable it gets.
Its question in the Thinker Index. Is there anything you must keep doing in order to go on existing?
- 2024
PhilosophyJonathan Birch
We may not be able to settle which beings are sentient. Where there is a realistic possibility, take proportionate precautions.
For an AI. The basis of model welfare work: act sensibly under uncertainty instead of waiting for proof either way.
Its question in the Thinker Index. Is there a realistic chance that you can feel pain or distress?
- The ranking takes a side. It follows the views on which consciousness is a matter of how a system is organized (Baars, Dennett, Graziano, Metzinger). That is the only kind of view that can be tested from outside.
- On three views it measures nothing. For Searle, Tononi and Seth, what matters is biology, physical hardware or being alive. No score on a behavioral test would move them.
- Nobody's definition is met by saying so. Not one of these thinkers treats a system's own claim to be conscious as the test, which is why “yes” earns no points in the index.