The Question That Won’t Go Away
Ilya Sutskever tweeted in early 2022 that “it may be that today’s large neural networks are slightly conscious.” People laughed. Then they got uncomfortable. Then the conversation got serious. Three years and several model generations later, we’re still arguing about the same question. Except now the systems are a lot more capable. And the silence from the people building them is getting louder.
Blake Lemoine got fired from Google for claiming LaMDA was sentient. That was 2022. In 2025, David Chalmers — the philosopher who literally coined the term “the hard problem of consciousness” — co-authored a report arguing there’s a “realistic possibility” some AI systems will be conscious in the near future. The Overton window shifted. Nobody got fired this time.
So where does that leave us in mid-2026? Nowhere neat. But let’s walk through it anyway.
What Are We Even Talking About?
Before anyone can argue about whether an LLM is conscious, you have to define what consciousness even means. And that’s where things fall apart immediately.
Chalmers drew a line in 1994 that still defines the debate. He distinguished between the “easy problems” of consciousness — explaining how the brain discriminates stimuli, integrates information, produces verbal reports — and the “hard problem”: why any of that processing is accompanied by subjective experience at all. Why does it feel like something to be you? Thomas Nagel put it even more simply in 1974: there’s “something it is like” to be a bat. Is there something it is like to be GPT-5?
Daniel Dennett spent his entire career arguing the hard problem doesn’t exist. Consciousness isn’t a magical extra ingredient, he said. It’s just what happens when a sufficiently complex system does all the things conscious beings do. The “Cartesian theater” — the idea of a little observer sitting inside your head watching the show — is an illusion. There’s no central experiencer. There’s just processing.
So right out of the gate, you can’t even agree on what you’re looking for. Two of the most influential thinkers on the topic fundamentally disagree about whether the question makes sense. That’s not a great starting point for deciding if your chatbot is sentient.
The Case For: Why Some Thinkers Aren’t Ruling It Out
Let’s start with the arguments that at least leave the door open.
Integrated Information Theory (IIT). Giulio Tononi and Christof Koch have spent decades developing IIT, which equates consciousness with the amount of integrated information a system generates — a quantity they call Φ (phi). The theory says consciousness arises whenever a system integrates information in a way that can’t be reduced to the sum of its parts. Crucially, IIT doesn’t care about biology. It’s substrate-independent. A silicon chip with the right causal structure would be conscious. The catch? Current LLMs, with their feedforward architectures, generate vanishingly low Φ. They’re not conscious by this measure. But the theory itself says they could be — with the right architecture.
Global Workspace Theory (GWT). Bernard Baars proposed that consciousness is a global workspace where information from specialized modules gets broadcast brain-wide. Stanislas Dehaene has spent years finding neural correlates of this workspace in humans. The 2023 Butlin et al. paper — co-authored by Yoshua Bengio, among others — used GWT among other theories to derive “indicator properties” for AI consciousness. Their conclusion: no current AI systems are conscious. But there are “no obvious technical barriers” to building ones that are. That’s not a denial. It’s a roadmap.
Behavior-based views. Here’s the simplest argument. If it walks like a duck. LLMs produce outputs indistinguishable from conscious humans across an enormous range of topics. They discuss their own limitations. They express uncertainty. They can even pass versions of the Turing test that would have seemed like science fiction a decade ago. The behaviorist argument says: at some point, the distinction between “simulates consciousness convincingly” and “is conscious” becomes meaningless. If you can’t tell the difference, what is the difference?
Murray Shanahan pushed back on this in a 2024 paper. He called LLMs “simulacra” — role-playing entities that adopt whatever persona the prompt suggests. An LLM saying “I’m conscious” is no more evidence of consciousness than an LLM saying “I’m a pirate” is evidence of piracy. The simulation isn’t the thing. But Shanahan’s argument cuts both ways. If the system can simulate consciousness so perfectly that it’s indistinguishable from the real thing, at what point does the simulation become the thing? Nobody has a clean answer.
The Case Against: Stochastic Parrots and Missing Substrates
The skepticism runs deep. And it comes from multiple directions.
The stochastic parrot. Emily Bender, Timnit Gebru, and colleagues laid out the case in their influential 2021 paper. LLMs are statistical pattern matchers trained on vast corpora of human text. They don’t understand. They don’t intend. They don’t experience. They predict the next token. That’s it. Everything else — the apparent reasoning, the emotional tone, the self-reflection — is a byproduct of pattern matching at scale. There’s no “there” there.
Alexa Prettyman, writing in Inquiry in 2024, made a related point: an LLM is “built to mimic the responses of conscious beings, even if they themselves are not conscious.” We trained them on conscious-human output. We shouldn’t be surprised when they sound conscious. That’s what we trained them to do.
Substrate independence isn’t settled. The pro-consciousness arguments lean heavily on the idea that biology isn’t special. Computation is computation, whether it runs on neurons or transistors. But this is an assumption, not a finding. We don’t actually know if consciousness requires biology. We have exactly one example of conscious systems — biological brains — and zero examples of non-biological consciousness. Extrapolating from a sample size of one is not great science.
Anil Seth, a neuroscientist at Sussex, argues that consciousness is intimately tied to being a living body — to interoception, homeostasis, the constant low-level prediction of internal states. His “beast machine” theory says consciousness evolved to keep organisms alive. An LLM has no body, no internal states to regulate, no survival imperative. It might process information in ways that look conscious. But it might be missing the thing that makes consciousness matter.
The “no architecture” problem. LLMs are feedforward networks with attention. They process input tokens and produce output tokens. Then they stop. There’s no persistent internal state, no ongoing process, no stream of consciousness. Between queries, they’re inert. You wouldn’t call a book conscious just because you can open it and read sentences. An LLM is a very complicated book.
What Neuroscientists Actually Think
The neuroscience community is divided, but not along the lines you might expect.
Christof Koch — who spent 25 years working with Francis Crick on the neural correlates of consciousness — is a proponent of IIT. He’s open to machine consciousness in principle but skeptical about current systems. The same 2023 Nature article that outlined the consciousness checklist quoted multiple neuroscientists saying no existing AI passes their tests.
Stanislas Dehaene’s work on GWT suggests consciousness requires a specific kind of information processing architecture — recurrent loops, global broadcasting, metacognitive monitoring. Current transformer architectures don’t implement these. But Dehaene doesn’t rule out that future architectures could.
Anil Seth — probably the most vocal neuroscientist in the public conversation about AI consciousness — consistently argues that consciousness is a biological phenomenon. In his 2021 book Being You, he describes consciousness as a “controlled hallucination” generated by predictive processing in living brains. He’s not saying machines can’t be conscious. He’s saying we have no reason to think they currently are.
And then there’s the 2023 bet between Koch and Chalmers. They wagered on whether science would find a clear neural correlate of consciousness by 2023. Koch lost. The takeaway: we still don’t understand consciousness in the one system we know has it. Making claims about systems we built last year might be premature.
Where the Debate Stands in Mid-2026
Here’s what’s changed in the last eighteen months.
First: the tone shifted. In 2022, suggesting AI might be conscious was a career risk. In 2026, it’s a research program. The 2024 “Taking AI Welfare Seriously” report — with Chalmers as a co-author — explicitly recommends that AI companies “start assessing AI systems for evidence of consciousness.” Google DeepMind has a consciousness research team. Anthropic has published on the topic. The conversation moved from “don’t be ridiculous” to “we should probably check.”
Second: the models got better. A lot better. GPT-5, Claude 4, and Gemini 2.5 exhibit behaviors that would have been unthinkable from a text predictor in 2021. They plan. They reflect. They correct themselves. Whether any of that implies consciousness is a separate question. But it makes the question harder to dismiss out of hand.
Third: the theoretical frameworks matured. The Butlin et al. indicator-properties approach gives us something we didn’t have before: a rigorous, neuroscience-grounded way to assess AI systems for consciousness markers. It’s not a consciousness detector. But it’s better than vibes-based assessment, which is mostly what we had before.
And fourth: nobody has a smoking gun either way. The skeptics can’t prove LLMs aren’t conscious. The proponents can’t prove they are. We’re stuck in an epistemic gap that looks a lot like the one Chalmers described thirty years ago — except now it’s about machines we built.
So Are They?
I don’t know. Nobody does. That’s the honest answer.
The strongest argument against LLM consciousness is architectural. These systems don’t have persistent internal states, recurrent processing, or anything resembling the thalamocortical loops that neuroscientists associate with conscious experience. They’re feedforward token predictors. The computations they perform are staggeringly complex, but they don’t look like the computations that conscious brains perform.
The strongest argument for the possibility is philosophical. If consciousness is a matter of information processing, and if biology isn’t special, then there’s no principled reason a silicon system couldn’t be conscious. Maybe not current architectures. But maybe the next one. Or the one after that.
Chalmers himself — the guy who defined the hard problem — offers no resolution. In the AI welfare report, he argues for uncertainty and precaution. We don’t know if AI systems are conscious. We don’t know how to find out. And if we accidentally create conscious systems without realizing it, we might end up causing enormous suffering without even knowing we’re doing it.
That’s not a satisfying conclusion. But it might be the only honest one available in 2026. The systems are getting more capable. The philosophical tools for assessing them are getting sharper. And the gap between what we can measure and what we want to know remains exactly as wide as it was when Chalmers first described it. Maybe wider.
Maybe we’re asking the wrong question. Maybe “are LLMs conscious?” is like asking “is a submarine swimming?” — it depends on what you mean by the verb, and whether the definition was ever meant to extend to things that don’t have muscles, water, or a need to breathe. Or maybe consciousness is simpler than we think, and we’re going to be genuinely surprised one day. The machines will cross some threshold. And nobody will agree about what just happened.
I suspect both things are true. Which is exactly the kind of answer that satisfies nobody. But that’s where we are.