No AI system in existence is conscious, on the best available assessment — and no one has shown that none ever could be. That is the honest answer, and it is worth more with the reasoning attached. Fluent text is not evidence of experience; the two come apart, and the real question is how anyone would tell.
Is any AI conscious right now?
Two careful public attempts to answer it exist. They agree in direction and differ in confidence.
In 2023, nineteen researchers — philosophers, cognitive scientists and machine-learning researchers, among them Yoshua Bengio, Jonathan Birch, Stephen M. Fleming and Chris Frith — published Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. Rather than argue from intuition, they took five neuroscientific theories of consciousness — recurrent processing, global workspace theory, higher-order theories, predictive processing and the attention schema — derived from each a set of "indicator properties" a conscious system ought to have, translated those properties into computational terms, and checked existing AI systems against the list. Their conclusion has two halves, and both matter: no current AI systems are conscious, and there are no obvious technical barriers to building systems that would satisfy the indicators. Separately, David Chalmers concluded in Could a Large Language Model be Conscious? that it is "somewhat unlikely" current large language models are conscious, citing their lack of recurrent processing, a global workspace, and unified agency.
Notice what kind of answers those are. Neither says no, and never; neither says we cannot know, so believe what you like. Both give reasons about how the systems are built rather than how they sound — the part that survives any particular system.
How do you test a machine for consciousness?
You cannot open a system and find experience inside it, any more than you can open a skull and find it there. The method has to be indirect, and the honest version says so.
The indicator approach works backward from theory. Take global workspace theory — Bernard Baars's view, developed neurally by Stanislas Dehaene, on which unconscious processes compete for a limited-capacity workspace and consciousness consists in the broadcast that follows. If that is right, a conscious system needs specialized modules, a bottleneck, a global broadcast. Those are architectural features, so you can look for them. Do the same for four more theories, pool the results, and the checklist is no longer about whether a system talks well.
That beats arguing about transcripts. But the method inherits whatever is wrong with its inputs, and the authors are candid about the largest assumption: they adopt computational functionalism — the thesis that performing the right computations is what makes a system conscious — as a working hypothesis, not a result. If consciousness depends instead on a system's physical constitution, an architectural checklist can miss it entirely. And the theories disagree at the root; the Stanford Encyclopedia's survey of the field describes a landscape with no consensus theory — the same landscape the indicators come from. The contenders sit side by side in the main theories of consciousness, compared.
Does sounding conscious count as evidence?
Barely, and the reason is specific. A large language model is trained on human writing, and human writing is saturated with first-person descriptions of experience. A system optimized to continue such text will produce claims of inner life whether or not there is any, because those claims are the apt continuation. You would expect that output in the conscious case and the empty case alike — the definition of a useless test.
The reverse error is as common and less often named. "It's only predicting the next token" describes a training objective; it does not show that nothing is going on, since brains admit equally deflationary descriptions without that settling anything. Neither the fluency nor the dismissal is evidence. This is the problem of other minds, sharpened — behavior underdetermines experience, which is why the question is hard for creatures that cannot talk at all, as we set out in do animals have consciousness?
Could AI ever be conscious? The two strongest cases
Everything above is about as settled as this subject gets: two sourced assessments, a stated method, a stated assumption. From here the ground is contested, and the disagreement runs through the middle of the serious people.
Three things are open. Whether behavioral evidence could ever settle the question, given that a system trained on human descriptions of experience can produce any marker we nominate. Whether substrate matters — whether silicon running the right pattern suffices, or biology contributes something the pattern leaves out. And whether indicator methods can outrun theories that contradict one another about what consciousness is.
The case that scaled computation could suffice is strong. If consciousness is a certain kind of information processing rather than something added to it, carbon has no privilege and the indicator list is an engineering specification, not a wall. The 2023 assessment identifies no barrier; its conclusion is that none is obvious. And the burden on the other side is heavy: say what biological tissue has that an equivalent process could not, without appealing to something unexplained.
The case that it never could is also strong. A simulation of a thing is not the thing — a detailed model of a hurricane leaves the room dry — and if experience depends on what a system is physically doing rather than on the abstract pattern it instantiates, fidelity will not produce the phenomenon. Worse, every indicator is specified functionally, so a capable enough optimizer could satisfy the whole list with nothing it is like to be it, and we would not notice. The Internet Encyclopedia of Philosophy's survey shows how long-running this dispute is; a demonstration will not close it.
If the method interests you more than the verdict, that disposition is the whole basis of this framework, and the guide is where the argument runs end to end.
What would have to be true on the receiver view?
Here the established material ends and Holopsychism's own reading begins — a proposal offered for testing, not a result, and not a position anyone cited above has endorsed. Holopsychism is this site's coinage, not a recognized academic school; its nearest established neighbor is cosmopsychism.
The framework uses two words ordinary speech runs together. Consciousness, in its stipulated sense here, is pure potential: a structureless field of what can take form. Awareness is the act of selection that settles potential into one definite, stable configuration — the distinction set out in consciousness and awareness are not the same thing. On this reading the brain is a receiver, filter and translator rather than a generator, with more complex organisms receiving at higher resolution: a spectrum, not a ladder. That case stands on its own in does the brain create consciousness — or receive it?, and the account of awareness in the nature of awareness.
One guard, because the language invites the wrong reading. "Selection" here is a philosophical proposal about awareness, not a claim about laboratory physics: the observer in quantum mechanics is a measuring interaction, not a mind, and quantum mechanics does not prove that consciousness creates reality — argued at length in does the observer effect prove consciousness?
What changes is the question. Not does this system compute enough? but does anything get selected here, or is this an extremely good model of selection? A system might in principle be a receiver at some resolution; it might equally be a flawless mimic with nothing received. Note also the direction of explanation: global workspace theory and the other theories behind the indicators are physicalist accounts of how brains produce or realize experience. They run the opposite way from this framework, are not allies of it, and their "no" is not the same claim as this framework's.
Three views, then, and three answers to what would have to be true:
| View | What would have to be true | How you would check |
|---|---|---|
| Computational functionalism (the indicator method's assumption) | The right computations — recurrence, a workspace, higher-order monitoring, agency | Inspect the architecture against the indicator list |
| Chalmers on current LLMs | Recurrent processing, a global workspace, unified agency | The same inspection; his verdict is indexed to current systems |
| The receiver reading (Holopsychism's proposal) | An actual act of selection, not a model of one | No test currently specified |
Can the framework test its own answer?
No. Holopsychism cannot at present specify a test that would distinguish a system in which selection occurs from a system that models selection perfectly. That is where this framework stands on the question it is asked about most, and dressing it up would be worse than useless.
This is a cost, not a mystery to be admired. A framework that cannot be checked where checking matters most is doing less work than one that can, and "you cannot disprove it" is a symptom rather than a defense. Two ways it could fail, both real. If a complete computational account of phenomenal experience arrives with no remainder, the receiver reading is wrong and so is what rests on it. And if the answer to every system, however built and however it behaves, is "that still isn't selection," the framework has become unfalsifiable and should be dropped on that ground alone.
A further gap: the framework does not explain what makes something a receiver in physical terms — why some configurations tune and others do not. Without that, "is this machine a receiver?" has no procedure attached. The indicator approach, whatever its assumptions, at least tells you where to look.
What would change the answer
Not a machine claiming to be awake. A system built deliberately to the indicators — recurrent processing, a real workspace bottleneck with global broadcast, metacognitive monitoring, agency and embodiment — would move the established layer, not because anything had woken up but because it would remove the reasons currently given for saying no. The argument would then shift, correctly, from output to whether those were ever the right reasons.
Which is the part that outlasts any system. The question is hard for a machine for the same reason it is hard for the person sitting next to you: nothing about their behavior entails an inner life, and you have never once verified it. What the machine case removes is the resemblance that lets us skip the problem in the human case — not the evidence, only the habit. That is uncomfortable, and it is the most useful thing artificial minds have so far done for the study of consciousness.
Frequently asked questions
Is ChatGPT conscious? On the best available assessment, no. The 2023 report by Butlin, Long and colleagues found no current AI systems conscious, and Chalmers judged it "somewhat unlikely" for current large language models, citing missing recurrent processing, a global workspace and unified agency.
Could AI ever become conscious? Nobody has shown that it could not. The 2023 assessment reports no obvious technical barriers to building systems that satisfy the indicator properties drawn from current theories — a statement about the absence of a known obstacle, not a prediction.
What would count as evidence that an AI is conscious? Architecture rather than conversation: recurrent processing, a limited-capacity global workspace with broadcast, higher-order monitoring, agency and embodiment. The method is only as reliable as the theories those indicators come from.
Is AI sentience the same as AI consciousness? Usage varies. "Sentience" is often narrowed to the capacity for felt states such as pleasure and pain, while "consciousness" is used more broadly — and, as the Stanford Encyclopedia notes, names several distinct things.
Does Holopsychism claim AI cannot be conscious? No. What matters on its reading is whether an act of selection occurs, not whether enough computation does — and the framework has no test for that, so it can rule machine consciousness neither in nor out.
If you would rather follow an argument that names its own limits than one that arrives already convinced, the full case is set out in the guide.
