To What Extent Can Electroencephalography (EEG) Be Used to Understand Consciousness?

Consciousness is exceptionally difficult to study scientifically: it is directly accessible from a first-person perspective but must be measured indirectly from a third-person one. A person may know what it is like to feel pain or hear a sound, but a researcher must infer that experience from behaviour, physiology, or neural activity.¹ Any workable definition must therefore be operationalisable from third-person evidence. This essay expands on Seth et al.’s² work to define consciousness as the capacity for subjective experience, operationalised through three constructs: conscious level—an organism’s global state of arousal, ranging from coma to full wakefulness; conscious content—the specific, first-person experience present within that state at a given moment; and conscious access—content that has become functionally available for report, reasoning, or the guidance of action.³ This framing matches Friedman et al.’s⁴ distinction between level and phenomenal aspects, and sets the terms on which electroencephalography (EEG) can be evaluated. This essay argues that EEG can be used to understand consciousness to a substantial, but bounded extent: strong, converging evidence about conscious level, moderate evidence for conscious access, but only indirect evidence about phenomenal experience itself. It will first explain what EEG measures and where its strengths and limitations lie, then critically evaluate studies of covert command-following, event-related potentials (ERPs), and cortical complexity, situate those paradigms within the active, passive, and resting-state distinction they exemplify, and ask how theoretically secure their markers are, before concluding that EEG’s capacity to measure consciousness is strongest when its evidence converges across paradigms, rather than as a direct readout of subjective experience.

EEG does not measure consciousness itself; it records scalp voltage fluctuations generated mainly by synchronised postsynaptic activity across large populations of cortical neurons, not individual action potentials.⁵ Its principal strength is temporal precision, making it well suited to asking when neural processing occurs. It also registers radially and tangentially oriented sources, and deep structures as well as superficial cortex.⁶ However, three limitations bound what that precision can deliver. First, volume conduction through brain tissue, cerebrospinal fluid, skull, and scalp means topography does not map simply onto cortical generators, and the signal is vulnerable to ocular, muscular, cardiac, and environmental artefacts.⁷ Second, and following directly from this, that ambiguity bears directly on conscious level, which depends on bi-thalamic and brainstem arousal systems.⁸ Because no scalp pattern has a unique generator,⁹ EEG can only ever index those systems downstream, through cortex. Third, single-trial EEG has poor signal-to-noise, so an evoked response emerges only across many clean trials,¹⁰ making single-patient inference fragile in precisely the population where it matters most. Its value for conscious level therefore comes not from millisecond timing as such—an argument for tracking access—but from its capacity to index how activity propagates, integrates, and differentiates across the cortex.

EEG is particularly useful for understanding consciousness when it reveals covert command-following in patients who appear behaviourally unresponsive. Clinical diagnosis relies on observable behaviour, yet behaviour can fail because of motor impairment, aphasia, sensory impairment, fluctuating arousal, or severe brain injury.¹¹ Owen et al.¹² first demonstrated covert command-following using functional magnetic resonance imaging (fMRI): a behaviourally vegetative patient’s activation during imagined tennis-playing and spatial navigation was indistinguishable from healthy participants. Cruse et al.¹³ adapted this logic to the bedside, requiring patients diagnosed as vegetative to imagine squeezing their right hand or moving their toes, tasks producing distinguishable sensorimotor EEG patterns in conscious participants. Of 16 patients, three produced reliable, command-specific EEG responses despite being behaviourally unresponsive. This separates behavioural output from conscious access: if a patient can understand an instruction, hold it in working memory, select the correct imagery task, and modulate brain activity at the correct time, behavioural silence cannot be equated with absence of awareness. Importantly, the task required repeatedly timed, command-specific differentiation between two mental actions, which a passive sensory response cannot produce. A later reanalysis, however, found that comparing all block-pair combinations rather than only adjacent pairs reduced two of the three patients’ accuracy to chance, with no motor-imagery markers surviving correction for multiple comparisons, suggesting the original classification may reflect statistical artefact.¹⁴ The same design that makes command-following hard to fake also limits negative results: since the task is cognitively demanding, failure may reflect impaired hearing, comprehension, attention, or working memory rather than absence of consciousness. Thus, the paradigm can demonstrate preserved access when classification holds, but cannot rule out experience when the task fails.

If command-following provides the strongest positive evidence, ERPs offer a broader but less specific route. ERPs are voltage changes extracted from ongoing EEG by averaging trials time-locked to a stimulus.¹⁵ An unexpected sound interrupting a repeating pattern elicits an automatic mismatch response, a low-level comparison occurring without conscious awareness.¹⁶ Detecting a longer-term, abstract rule and noticing when it breaks is different. It requires holding the rule in mind and updating it, a more demanding process reflected in a later, larger waveform, the P3b.¹⁷ Bekinschtein et al.¹⁸ demonstrated this in healthy participants: short-term deviants elicited only the mismatch response, whereas rule violations elicited the P3b, and only when participants were aware of the structure. Faugeras et al.¹⁹ adapted this local-global paradigm for patients meeting criteria for the vegetative state. Of 22 patient recordings, two showed a P3b-like effect interpreted as evidence of conscious access, and both patients showed clinical signs of consciousness within 3 to 4 days. Bedside electrophysiology therefore provided clinically meaningful evidence before behaviour revealed awareness, though a 9% detection rate leaves the estimate highly uncertain. They excluded a deafness explanation by identifying early cortical responses in negative cases, but 29% of recordings were discarded for poor signal quality. That rate is not incidental: it is the averaging requirement made visible, and it is why an ERP null in a restless, artefact-prone patient carries little evidential weight. Motor-imagery tasks demand sustained effort, giving high specificity but low sensitivity;²⁰ ERP paradigms require no such effort, but the marker itself is unreliable. P3b-like components appear in up to 40% of comatose patients, and the P3b is absent in most brain-damaged patients who are clearly conscious.²¹ ERPs extend EEG’s reach beyond volitional command-following, but they track processing that accompanies consciousness rather than constitutes it.

Different EEG paradigms make different inferential contributions, and no single marker can settle the question alone. Sanz et al.²² divided consciousness-probing paradigms into active, passive, and resting-state approaches: active paradigms are persuasive when positive, since volitional modulation is hard to explain without conscious access, but they are vulnerable to false negatives; passive paradigms reduce response demands yet cannot always distinguish consciousness from sensory processing, prediction, or attention; and resting-state measures drop task and sensory demands altogether, moving further from specific conscious contents but closer to conscious level. This category becomes most useful once made quantitative. Engemann et al.²³ trained a multivariate classifier on spectral, connectivity, and evoked-response features from 327 EEG recordings across two independent clinical centres, distinguishing unresponsive wakefulness syndrome from minimally conscious state with an area under the curve of 0.73–0.78 on previously unseen data, generalising across different EEG equipment and recording protocols. That consciousness-relevant information can be extracted with no task or stimulus is a genuine advance, but not a clean success. For a measure that could inform withdrawal-of-care decisions, that level of discrimination is modest. More fundamentally, the classifier’s ground truth is behavioural diagnosis using the Coma Recovery Scale-Revised, the very standard that covert command-following shows to underestimate consciousness, so its apparent errors may be precisely the covert cases that matter most. This circularity constrains nearly every marker considered here: EEG is repeatedly validated against the behaviour it is meant to surpass, so the case for EEG cannot rest on any single paradigm.

Cortical complexity offers a further route to conscious level, measuring how the brain responds when its cortex is directly perturbed rather than sensorily stimulated. Massimini et al.²⁴ showed that a transcranial magnetic stimulation (TMS) pulse delivered during wakefulness evokes a response propagating across interconnected cortical areas, whereas in non-rapid-eye-movement sleep it evokes a stronger but short-lived local response, indicating a breakdown in effective cortical connectivity as consciousness fades. Casali et al.²⁵ developed the perturbational complexity index (PCI) from this, measuring the joint integration and differentiation of TMS-evoked activity with no sensory stimulus and no behavioural response. Critically, PCI separated vegetative-state patients (0.19–0.31) from minimally conscious patients (0.32–0.49) and from conscious but paralysed patients with locked-in syndrome (0.51–0.62), tracking graded differences in conscious level that behavioural scoring had failed to resolve. Casarotto et al.²⁶ validated PCI independently in 81 patients, reaching 94.7% sensitivity for minimally conscious patients and identifying vegetative-state patients with a potential for consciousness that behaviour alone had missed, echoing Cruse et al.’s²⁷ covert awareness. That said, this evidence needs qualifying. The clinical groups are small—six vegetative and six minimally conscious patients in Casali et al.²⁸—and Casarotto et al.’s²⁹ threshold derives from a benchmark of 150 conscious and unconscious subjects, calibrating it against the report-based criteria it adjudicates. A more pertinent issue is that the informative manipulation here is TMS and EEG only the readout, so the strongest evidence for conscious level rests on a hybrid method that also needs neuronavigation and expertise, making it far less deployable at the bedside than routine EEG. Engemann et al.’s³⁰ purely EEG-based classification, however, reaches the same conclusion independently, itself an instance of convergence across paradigms. PCI also indexes a capacity for integrated, differentiated processing rather than experience itself, leaving complexity far more convincing on level than on content.

EEG’s usefulness is further limited because its proposed consciousness markers are theoretically contested. Koch et al.³¹ argue that prominent candidates, such as gamma activity or the P3b, have not shown the predictive values once attributed to them. Friedman et al.³² similarly describe the P3b as historically important but increasingly interpreted as post-perceptual processing rather than phenomenal awareness itself. That reinterpretation is where EEG comes closest to conscious content, and where it stops. If the P3b indexes access rather than experience, a content marker should survive when report is removed; in no-report paradigms much of the late frontal response does disappear while posterior activity persists,³³ which is why Koch et al.³⁴ locate content-specific correlates posteriorly. Yet such markers are still identified by contrasting trials participants reported seeing against those they did not, leaving them validated against the access process they are meant to be separated from. This uncertainty extends to the theories meant to explain these markers. A large adversarial collaboration testing Global Neuronal Workspace Theory (GNWT) against Integrated Information Theory (IIT) in 256 participants, using fMRI, magnetoencephalography, and intracranial recordings, found neither theory’s core predictions held: prefrontal activity was not sustained as GNWT predicts, and posterior activity was sparse where IIT predicted robustness.³⁵ No waveform maps onto consciousness one-to-one. Jackson and Bolger³⁶ add that polarity, scalp location, and waveform shape are not direct evidence of a generator beneath an electrode. Seth et al.³⁷ make the broader point: measures of consciousness gain meaning within theoretical frameworks and through convergence among behavioural and neurophysiological evidence. Thus, EEG cannot settle phenomenal consciousness on its own—it forces researchers to distinguish conscious level, access, and content rather than hiding those distinctions behind a single brain marker.

In conclusion, this essay has shown that EEG can be used to understand consciousness to a substantial, but bounded extent, with strong evidence for conscious level, moderate evidence for access, and only indirect evidence for content. EEG does not reveal subjective experience directly, and scalp voltage is not equivalent to consciousness itself. Rather, it provides evidence about the neural conditions under which consciousness becomes accessible, clinically detectable, or theoretically plausible. Three contributions support this position. Covert command-following reveals conscious access when behaviour is absent, separating what a patient can do from what a patient can show. ERPs extend this reach to patients who cannot sustain a demanding task, tracking rule-based processing that may anticipate clinical recovery. Cortical complexity indexes graded conscious level without any task or sensory input, extending the argument to patients who cannot perform command-following or rule-use. Read together across active, passive, and resting-state measures, and against the theories behind them, this evidence clarifies the dynamics of conscious level and access; it forces researchers to specify what their task can and cannot show, especially in fragile clinical populations; and it warns that behavioural silence is not always experiential absence. If consciousness can persist when outward response is absent, then diagnosis, prognosis, pain assessment, rehabilitation, and family communication depend on resisting overconfident conclusions from behaviour alone—including in the EEG measures still calibrated against it. EEG’s capacity to measure consciousness is therefore strongest when its evidence converges across paradigms, rather than as a direct readout of subjective experience.