What has already been said about consciousness and AI? Part 2: What the scientists say

Series: Part 2 of 3 in the Machine and mind essays.
This part covers the main scientific theories of consciousness, the row over integrated information theory, and what experiments on real brains have actually shown. It follows Part 1, the old puzzle, and Part 3 turns to the machines.

What are the main scientific theories of consciousness?

After the philosophy, I wanted to see what people who study brains for a living had come up with. I expected one theory with a few rivals. I found a crowd, and they do not agree.

The shortest version I could make is this. Global workspace theory links consciousness to broadcast, higher-order theories to self-monitoring, recurrent processing theory to feedback loops, and integrated information theory to a system's physical cause and effect structure. Others look to prediction, attention or life itself. Each would give a different answer about machines, which is why I kept them apart in my notes.

A useful guide here was the Stanford Encyclopedia's entry on the neuroscience of consciousness. It sets the main theories side by side and states what each one counts as necessary and sufficient (SEP, Neuroscience of Consciousness). I kept going back to it whenever I lost track. Early on I noticed that the theories disagree about more than the answer. They disagree about what has to be explained: the fact that information is available to the whole brain, the fact that we can report it, or the feeling itself.

How do global workspace, higher-order and recurrent processing theories differ?

The first big family is global workspace theory. Bernard Baars proposed the psychological version in a 1988 book, which I have only seen in the catalogue. The brain version, the global neuronal workspace, comes from Stanislas Dehaene, Jean-Pierre Changeux and colleagues. A 2020 review describes the core idea: in the conscious state, "a non-linear network ignition" amplifies and sustains a representation, so the information can be "globally accessed" by the rest of the brain (Mashour and others 2020). The image that helped me was a stage. Many processes run in the dark, and whatever wins the spotlight gets broadcast to everyone.

Its strength is a long run of experiments comparing what happens in the brain when a stimulus is seen and when the same stimulus goes unseen. The criticism I met most often is that it may explain access, what we can report and use, rather than felt experience. That is the distinction from Part 1 coming back. It matters for AI in a practical way, because of all the theories this is the easiest to turn into an engineering question: does the system have a limited workspace and a broadcast?

Higher-order theories take a different route. A mental state is conscious when the mind represents itself as being in that state, a kind of inner monitoring. Hakwan Lau and David Rosenthal reviewed the evidence in 2011 and argued that the view, although it began in philosophy, "is testable and has received substantial empirical support" (Lau and Rosenthal 2011). The Stanford Encyclopedia sets out the main worry, which is misrepresentation (SEP, Higher-Order Theories). What happens if the higher-order state gets the first-order state wrong, and represents me as seeing orange when I am seeing red? Which one is my experience?

Then there is recurrent processing theory, from Victor Lamme. Here the key is feedback inside the sensory areas themselves. Signals go forward and come back, and that loop is enough for perceptual experience without any global broadcast (Lamme 2006). Chalmers, in his 2023 paper on language models, says Lamme's theory gives recurrence "pride of place" as the central requirement (Chalmers 2023). That caught my attention, because he goes on to point out that current language models are almost all feedforward systems.

In my notes the three became three words: broadcast, self-monitoring, feedback. The summary table at the end of Part 3 sets every theory side by side.

What is integrated information theory, and why was it called pseudoscience?

Integrated information theory, or IIT, starts at the other end. Giulio Tononi, who proposed it in 2004, begins with experience itself and asks what a physical system must be like to have it. His 2004 paper opens with puzzles I had never thought to ask. Why is consciousness generated by some parts of the brain, such as the thalamocortical system, "and not by other parts, such as the cerebellum?" Why are we conscious when awake and much less so in dreamless sleep? His answer is that consciousness corresponds to a system's capacity to integrate information, measured as a quantity called Phi (Tononi 2004).

What struck me is how direct it is about what follows. A 2016 paper by Tononi, Melanie Boly, Marcello Massimini and Christof Koch says the theory can help build tools for assessing consciousness in patients who cannot communicate, and admits it "leads to some counterintuitive predictions" (Tononi and others 2016).

For machines, IIT is the odd one out. It ties consciousness to a system's physical cause and effect structure, not to the software it runs. Chalmers notes that it predicts feedforward systems have zero integrated information and therefore lack consciousness (Chalmers 2023). On this view an ordinary digital computer would be unlikely to be conscious whatever program it ran, however clever.

Then came the row. In September 2023, a letter posted on the preprint server PsyArXiv argued that IIT should be labelled pseudoscience. Its authors said the media had been celebrating IIT as a "leading" and empirically tested theory, and they wanted to express their concerns (Fleming and others 2023). Nature's news report said the letter was "signed by 124 scholars" and had caused "an uproar", with some researchers arguing over the label and others worried that it would increase polarisation in the field (Lenharo 2023). In March 2025 the journal Nature Neuroscience carried the argument on. A group signing as IIT-Concerned, the name the original letter is cited under, argued that the theory is unscientific because its core claims are untestable even in principle (IIT-Concerned and others 2025). Tononi and colleagues replied that attempts to label it pseudoscientific "expose a crisis in the dominant computational-functionalist paradigm" (Tononi and others 2025). In a third piece, Alex Gomez-Marin and Anil Seth assessed the charge and tried to turn the clash into lessons for the field (Gomez-Marin and Seth 2025). For these three I have read only the published summaries.

The criticisms of IIT I read most often were two. Phi is very hard to calculate for any real system, and the theory implies that some simple systems doing very little would be conscious. Is that a sign the theory has gone wrong, or a surprising discovery? The people I read did not agree.

What about predictive processing, biology, attention schemas and Orch OR?

Several more theories do not fit the families above.

Predictive processing treats the brain as a guessing machine. Andy Clark describes brains as "prediction machines" that constantly try to match incoming signals against their own expectations (Clark 2013). Karl Friston's free-energy principle tries to bring perception, action and learning under one idea (Friston 2010). Anil Seth, in a TED talk, calls perception a kind of "controlled hallucination" (Seth 2017).

Seth then takes a step that bears directly on AI. In a 2025 paper in Behavioral and Brain Sciences he makes the case that consciousness depends "on our nature as living organisms", which he calls "a form of biological naturalism". He concludes that real artificial consciousness is "unlikely along current trajectories" but becomes more plausible as AI becomes more brain-like or life-like (Seth 2025). That journal publishes open commentaries, and one, by Blake Richards and Blaise Agüera y Arcas, has a title that sums up the opposing view: "Life is inherently computational, and consciousness probably is too" (Richards and Agüera y Arcas 2026). I have only checked its title and authors so far, not read it.

John Searle had argued long before that minds depend on the brain's biology. The Stanford Encyclopedia records Steven Pinker's reply to Searle's appeal to the "causal powers of the brain": the relevant powers lie in the "patterns of interconnectivity" that carry out the information processing (SEP, Chinese Room). That one exchange held most of the AI debate in miniature for me.

Michael Graziano's attention schema theory is the most mechanical of all. The brain builds a simplified model of its own attention, much as it builds a model of the body, and that model leads it "to conclude that it has a subjective experience" (Graziano and Webb 2015). It could, in principle, be built. The question I kept coming back to was whether a model of attention explains experience or explains it away.

Orch OR, from the physicist Roger Penrose and the anaesthesiologist Stuart Hameroff, is the most unusual. They propose that moments of consciousness come from quantum processes in microtubules, tiny structures inside neurons (Hameroff and Penrose 2014). Penrose argues the brain must do something non-computational, for reasons I come back to in Part 3, and this is where he looks for it. It has drawn pointed criticism. Max Tegmark calculated that such quantum states would fall apart, or decohere, far faster than neurons fire, so the brain should behave as a classical system (Tegmark 2000). Rick Grush and Patricia Churchland called the chain from quantum gravity to microtubules to consciousness "entirely speculative", and pointed out that a gout drug which breaks down microtubules has no noticed effect on consciousness (Grush and Churchland 1995).

Finally, enactivism, from Francisco Varela, Evan Thompson and Eleanor Rosch's book The Embodied Mind (1991). It sees mind as arising from a living body acting in its world. The Stanford Encyclopedia describes one strand that treats cognition in terms of "the biodynamics of living systems" (SEP, Embodied Cognition). For AI the question it raises is simple and hard: can anything without a body and a life have a point of view at all?

What have experiments on consciousness actually shown?

With so many theories, I wanted to know what had actually been tested.

One long project is the search for the neural correlates of consciousness: the minimal brain activity that goes with a particular experience. It tells us what goes with experience, not why, which is the explanatory gap from Part 1 again.

What did the COGITATE study find?

The most direct test I found is COGITATE, published in Nature in 2025. It was an adversarial collaboration, the method that Templeton World Charity Foundation's Accelerating Research on Consciousness initiative promotes, where rival camps agree the predictions in advance (Templeton World Charity Foundation). The team pitted IIT against global neuronal workspace theory. 256 people looked at clearly visible stimuli for different lengths of time while their brains were recorded with fMRI, MEG and electrodes placed inside the skull. The results "align with some predictions" of both theories while "substantially challenging key tenets of both" (Cogitate Consortium 2025). For IIT, the team did not find the sustained synchronisation within the back of the brain that the theory predicted. For the workspace theory, there was a general lack of the predicted "ignition" when a stimulus disappeared, and some features of what people saw were only weakly represented in the prefrontal cortex. Neither side walked away the winner.

What do damaged, sleeping and anaesthetised brains show?

Some of the most striking evidence comes from damaged brains. In blindsight, people with damage to the visual cortex say they see nothing in part of their visual field, yet can still respond to what is there. A classic early study is Weiskrantz and colleagues (1974), which I have checked as a citation only. The Stanford Encyclopedia describes a patient, blind across his whole visual field, who could walk down a hallway around obstacles he reported being unable to see (SEP, Neuroscience of Consciousness). Clever processing without reported experience is the very gap the AI debate turns on.

Split-brain patients, whose hemispheres have been surgically disconnected, were long described as two minds in one skull. A 2017 study by Yair Pinto and colleagues found something different in two patients. Both patients showed full awareness of stimuli across the whole visual field, whether they answered with the left hand, the right hand or in words. Perception was split, but the results suggested that cutting the connection "does not create two independent conscious perceivers within one brain" (Pinto and others 2017). I have not yet checked how other researchers have responded to this.

Sleep and anaesthesia give a third line of evidence. When Marcello Massimini and colleagues stimulated one part of the cortex during deep sleep, the response stayed local instead of spreading, and they linked the fading of consciousness to "a breakdown in cortical effective connectivity" (Massimini and others 2005). A review of anaesthesia concluded that anaesthetics seem to cause unconsciousness "when they block the brain's ability to integrate information" (Alkire and others 2008). Out of this came a measure, the perturbational complexity index, which "reliably discriminated" levels of consciousness in single individuals, including patients emerging from coma (Casali and others 2013).

None of these tools has been tried on a machine, as far as I could find. They were all calibrated on brains, which left me with a question I could not get past. Can a test built around human brains say anything about silicon?

Next in the series: Part 3, the machines. The Turing test, the Chinese Room, Penrose's argument, large language models, the welfare question and what people believe.

My own reactions, as someone who builds AI agents, are in a separate essay. There is also an older piece of mine on a 2,500-year-old instruction: in the seen, only the seen.

Sources

1. The Neuroscience of Consciousness, Stanford Encyclopedia of Philosophy.
2. A Cognitive Theory of Consciousness, Bernard Baars, 1988 (catalogue record only).
3. Conscious Processing and the Global Neuronal Workspace Hypothesis, George Mashour, Pieter Roelfsema, Jean-Pierre Changeux, Stanislas Dehaene, 2020.
4. Empirical support for higher-order theories of conscious awareness, Hakwan Lau and David Rosenthal, 2011.
5. Higher-Order Theories of Consciousness, Stanford Encyclopedia of Philosophy.
6. Towards a true neural stance on consciousness, Victor Lamme, 2006.
7. Could a Large Language Model be Conscious?, David Chalmers, 2023.
8. An information integration theory of consciousness, Giulio Tononi, 2004.
9. Integrated information theory: from consciousness to its physical substrate, Giulio Tononi, Melanie Boly, Marcello Massimini, Christof Koch, 2016.
10. The Integrated Information Theory of Consciousness as Pseudoscience, Stephen Fleming and others, 2023.
11. Consciousness theory slammed as pseudoscience, sparking uproar, Mariana Lenharo, Nature (news), 2023.
12. What makes a theory of consciousness unscientific?, IIT-Concerned, Matthijs Klincewicz, Tony Cheng and others, Nature Neuroscience, 2025 (summary read).
13. Consciousness or pseudo-consciousness? A clash of two paradigms, Giulio Tononi, Larissa Albantakis and others, Nature Neuroscience, 2025 (summary read).
14. A science of consciousness beyond pseudo-science and pseudo-consciousness, Alex Gomez-Marin and Anil Seth, Nature Neuroscience, 2025 (summary read).
15. Whatever next? Predictive brains, situated agents, and the future of cognitive science, Andy Clark, 2013.
16. The free-energy principle: a unified brain theory?, Karl Friston, 2010.
17. Your brain hallucinates your conscious reality, Anil Seth, TED, 2017.
18. Conscious artificial intelligence and biological naturalism, Anil Seth, 2025.
19. Life is inherently computational, and consciousness probably is too, Blake Richards and Blaise Agüera y Arcas, 2026 (title and authors checked, not read).
20. The Chinese Room Argument, Stanford Encyclopedia of Philosophy.
21. The attention schema theory: a mechanistic account of subjective awareness, Michael Graziano and Taylor Webb, 2015.
22. Consciousness in the universe: a review of the Orch OR theory, Stuart Hameroff and Roger Penrose, 2014.
23. Importance of quantum decoherence in brain processes, Max Tegmark, 2000.
24. Gaps in Penrose's Toilings, Rick Grush and Patricia Smith Churchland, 1995.
25. The Embodied Mind, Francisco Varela, Evan Thompson, Eleanor Rosch, 1991.
26. Embodied Cognition, Stanford Encyclopedia of Philosophy.
27. Adversarial testing of global neuronal workspace and integrated information theories of consciousness, Cogitate Consortium, 2025.
28. Accelerating Research on Consciousness, Templeton World Charity Foundation.
29. Visual capacity in the hemianopic field following a restricted occipital ablation, Lawrence Weiskrantz and others, 1974 (citation checked, not read).
30. Split brain: divided perception but undivided consciousness, Yair Pinto and others, 2017.
31. Breakdown of cortical effective connectivity during sleep, Marcello Massimini and others, 2005.
32. Consciousness and anesthesia, Michael Alkire, Anthony Hudetz, Giulio Tononi, 2008.
33. A theoretically based index of consciousness independent of sensory processing and behavior, Adenauer Casali and others, 2013.

How this was made: the idea, the practice and the final edit are mine. An AI assistant helped with research, fact-checking and copy-editing, and never invented experiences or quotes.