The major approaches since the 1990s, what each tried, and what happened. Most are still open or work only in part. Works in part means it does some of what it claims. Weakened means a later result cut into it. Still open means no decisive test yet.
Find the brain activity that goes with experience
Works in partFrancis Crick and Christof Koch, then most of the field · 1990 onwards
Compare brain activity when the same input is seen and when it is missed. The difference should point to the activity behind experience.
What happened: It produced many reliable differences between seen and unseen. But a 2016 review by Koch and colleagues found that several markers had proved illusory, and many may reflect reporting rather than experience.
Global workspace theory
Works in partBernard Baars; Stanislas Dehaene and Jean-Pierre Changeux · 1988 onwards; brain model 1998
Most processing is unconscious and local. Something becomes conscious when it is broadcast across a brain-wide network, with a sudden “ignition” of activity in frontal and parietal areas.
What happened: It explains a great deal of lab data on seeing versus missing a stimulus. In the COGITATE test (2025) one of its key predictions failed: the expected frontal flare-up when a picture disappeared was mostly absent. Two 2026 reanalyses of the same data lean back towards it.
The P3b brain wave as a sign of conscious access
WeakenedClaire Sergent, Stanislas Dehaene, Jean-Pierre Changeux and others; challenged by Michael Pitts, Stephen Metzler and Steven Hillyard · 2005 to 2011; challenged 2014
In 2005 Sergent and colleagues found that from about 270 ms after a word appeared, a late wave of brain activity followed only words people saw. The P3b, a large EEG wave peaking around 0.4 s, became a leading proposed marker of the moment something becomes conscious.
What happened: When people saw a stimulus but did not need to report it, the P3b was absent (Pitts and colleagues, 2014). An earlier wave at the back of the head, at about 200-240 ms, remained. So the P3b seems tied to reporting, not to seeing itself.
No-report experiments
Works in partNaotsugu Tsuchiya, Melanie Wilke, Stefan Frässle, Victor Lamme and others · 2010s onwards
Separate the activity behind an experience from the activity behind reporting it, by designing tasks where people see things without being asked about them.
What happened: They removed some candidate markers, such as the P3b. But without a report, experimenters must infer what was seen in other ways, and that inference can be questioned.
Integrated information theory (IIT) and Φ
Still openGiulio Tononi; version 4.0 by Larissa Albantakis, Tononi and colleagues · 2004; IIT 4.0 in 2023
Start from what experience is like and derive what a physical system must be like to have it. The amount of consciousness is a quantity, Φ (phi): how much a system's causes and effects cannot be split into independent parts.
What happened: Φ cannot be computed for a brain. The reference software, PyPhi, takes time that grows exponentially with size, which limits exact answers to about 10-12 elements. So only rough stand-ins can be tested. In COGITATE, IIT's prediction of sustained synchrony at the back of the brain was not found.
Stand-ins for Φ from ordinary brain recordings
WeakenedAndré Nilsen, Alessandro Arena and Johan Storm; Andrea Luppi and colleagues · 2024; 2026
Estimate integrated information from EEG or fMRI and check that it falls when consciousness is lost.
What happened: The stand-ins disagree. In rats, integration estimates from spontaneous EEG generally rose under propofol and sevoflurane. In fMRI across humans, monkeys, marmosets and mice, integration broke down under anaesthesia. Which way a proxy moves depends on the method.
Recurrent processing theory
Still openVictor Lamme and Pieter Roelfsema · 2000 onwards
A first fast sweep of activity through the visual system is unconscious. Seeing begins when later areas feed signals back to earlier ones. No frontal broadcast is needed.
What happened: The seen-but-unreported wave at the back of the head (Pitts 2014) fits it. But Pitts and colleagues note it might reflect attention to objects instead. We found no large adversarial test of it with published results.
Higher-order theories
Still openDavid Rosenthal; Hakwan Lau; Richard Brown, Lau and Joseph LeDoux · 2011 review; 2019 review
A state is conscious when the brain also represents itself as being in that state. This points to the prefrontal cortex and to confidence judgements.
What happened: Lau and Rosenthal (2011) reviewed supporting evidence and defended the view against critics. It was not one of the two theories tested in COGITATE. We could not confirm reports that a large adversarial test of it is running.
Predictive processing
Still openJakob Hohwy, Anil Seth, Karl Friston and others · 2010s onwards
The brain constantly predicts its input. What we experience is the brain's best guess, shaped by those predictions.
What happened: It is a broad framework, which makes sharp tests hard. The INTREPID collaboration published its agreed plan in May 2026, pitting IIT against two predictive theories. No results yet.
Attention schema theory
Still openMichael Graziano and Taylor Webb · 2015
The brain keeps a simplified model of its own attention, as it keeps a model of the body. That model is what leads us to say we have experience.
What happened: It predicts that attention can work without awareness but will be poorly controlled. Its authors say some evidence fits. We found no large pre-registered test of it.
Dendritic integration theory
Still openJaan Aru, Mototaka Suzuki and Matthew Larkum · 2020
Consciousness depends on large cortical neurons joining signals from below (the senses) and from above (the rest of the cortex) in their long branches.
What happened: In mice, three different anaesthetics cut the link between the top branches and the cell body of these neurons (Suzuki and Larkum, 2020). The work is mostly in animals, and the theory's predictions for human EEG are not yet tested.
Quantum consciousness (Orch-OR)
WeakenedStuart Hameroff and Roger Penrose · mid-1990s; review 2014
Quantum states in microtubules, tiny tubes inside neurons, collapse in a specific way proposed by Lajos Diósi and Penrose. Each collapse would be a moment of experience.
What happened: In 2000 Max Tegmark calculated that such quantum states would decay in 10^-13 to 10^-20 s, far faster than neurons act. In 2021 an experiment under the Gran Sasso mountain ruled out the simplest, parameter-free version of the Diósi-Penrose collapse. In 2024 a microtubule-stabilising drug delayed anaesthetic knockout in rats by 69 s on average. That fits a role for microtubules but does not test the quantum part. No experiment has shown quantum coherence tied to experience.
COGITATE: two theories tested head to head
Works in partCogitate Consortium · Published April 2025
Proponents of global workspace theory and IIT agreed in advance which results would count against each theory. Then 256 people viewed pictures while scanned with fMRI (120), MEG (102) or electrodes in the brain (34).
What happened: Neither theory won. IIT was challenged by the lack of sustained synchrony at the back of the brain. Global workspace theory was challenged by the general lack of a frontal flare-up when a picture disappeared.
Reanalyses of the COGITATE data
Still openKavindu Bandara, Elise Rowe and Marta Garrido; Annie Bryant and Christopher Whyte · 2026
Analyse the same public recordings with different methods.
What happened: Bandara and colleagues (a preprint, not peer reviewed) found the frontal flare-up at both picture onset and offset, which the original analysis missed. Bryant and Whyte tried 246 connectivity measures and gave tentative support to global workspace theory. The conclusion depends on the method chosen.
Perturbational complexity index (PCI)
Works in partAdenauer Casali, Silvia Casarotto, Marcello Massimini and colleagues · 2013; validated 2016
Tap the brain with a magnetic pulse and measure how complex the EEG echo is. A conscious brain should give a rich, hard-to-compress response.
What happened: The best-validated measure so far. A cutoff set in 150 people who could report gave perfect separation in that group, then caught 94.7% of minimally conscious patients. It needs TMS equipment, and we found no open dataset of this kind.
Predicting PCI from resting EEG
Works in partCharlotte Maschke, Stefanie Blain-Moraes and colleagues; Derek Newman and colleagues · 2024; 2026
If ordinary resting EEG could stand in for the magnetic-pulse test, many more patients and datasets could be checked.
What happened: In 15 healthy adults under three anaesthetics, it predicted PCI closely. In 26 brain-injured patients (a 2026 preprint), the prediction did not hold across the whole group, only in subgroups.
Signal complexity and the "entropic brain"
WeakenedMichael Schartner, Robin Carhart-Harris, Anil Seth and colleagues · 2014 onwards
Conscious brains produce varied, hard-to-compress signals. Lempel-Ziv complexity, a compression score, should fall in anaesthesia and rise under psychedelics.
What happened: Complexity fell under propofol (2015) and rose above normal waking under ketamine, LSD and psilocybin (2017). But in 2025 Maschke and colleagues showed, in 303 patients, that the score largely tracked a simpler property of the signal: how fast power falls with frequency (r = 0.86). Once that was controlled, its diagnostic value vanished.
The slope of the EEG spectrum
Works in partMichele Colombo, Marcello Massimini, Simone Sarasso and colleagues · 2019
How fast EEG power falls off with frequency, the spectral exponent, may itself mark whether consciousness is present.
What happened: It separated waking and ketamine from xenon and propofol, and correlated strongly with PCI. The study had 5 people per drug. It is the simple baseline every fancier measure now has to beat.
EEG classifiers for brain-injured patients
Works in partJacobo Sitt, Denis Engemann and colleagues · 2014; 2018
Combine many EEG markers in a machine-learning classifier to tell vegetative from minimally conscious patients.
What happened: Sitt and colleagues screened markers in 181 recordings. Engemann and colleagues used 327 patient recordings and 66 controls and reached an AUC of 0.73 to 0.78 across sites (1.0 is perfect, 0.5 is chance). Useful, but far from a bedside verdict.
Reading dreams from the sleeping brain
Works in partHorikawa, Kamitani and colleagues; Francesca Siclari, Giulio Tononi and colleagues; the DREAM database team · 2013; 2017; 2025
Wake sleepers, ask what they experienced, and see whether the brain activity just before predicts the report.
What happened: Horikawa decoded the content of sleep-onset imagery from fMRI (2013). Siclari found less slow activity at the back of the head before dream reports, and used it to predict them live. The DREAM database (2025) pooled 505 people and 2,643 awakenings from 20 studies. In the six datasets it analysed, EEG predicted dream versus no dream only weakly: mean AUC 0.586 in non-REM sleep and 0.700 in REM.
Ketamine: unresponsive but dreaming
Still openSimone Sarasso, Marcello Massimini and colleagues · 2015
Find a case where behaviour and experience come apart, and see which measures follow experience.
What happened: Under ketamine, 18 volunteers' brains gave a wake-like response to the magnetic pulse, and afterwards people reported long, vivid dreams. Propofol and xenon gave neither. The key later analyses (2019, 2024) reuse the same small group of volunteers, so this rests on one cohort.
Dreaming under sedation
Still openImad Bajwa, André Nilsen, Johan Storm, Bjørn Juel and colleagues · 2025
Sedate volunteers with propofol, wake them, and ask whether they experienced anything. Do complexity measures follow the reports?
What happened: Both PCI and EEG complexity fell with sedation. Neither told “I was dreaming” from “nothing” within sedation. Only 5 awakenings gave a clear “nothing”, so only a large difference could have been detected.
Indicators of consciousness in AI
Still openPatrick Butlin, Robert Long and colleagues · 2023 report; peer-reviewed method 2026
Derive indicator properties from brain theories, then check an AI system's internal workings against them, rather than trusting what it says.
What happened: The 2023 report found no current AI system conscious but no obvious barrier to building one that meets the indicators. The verdict depends on which theories are right, which is exactly what is unsettled.