Tristan Bekinschtein Consciousness Biomarkers Clinical Measurement AI Implications ASSC 29

Tristan Bekinschtein, a neuroscientist at the University of Cambridge whose lab works on the neural signatures of consciousness in patients with disorders of consciousness (DoC), delivered one of the keynote addresses at ASSC 29 in Santiago, Chile (June 30 to July 3, 2026). His research sits at the intersection of two questions that are usually treated separately: the clinical question of how to detect residual awareness in patients who cannot communicate, and the scientific question of what biomarkers reliably track phenomenal consciousness across different states and substrates.

Both questions have direct implications for AI consciousness research, though the connection is rarely made explicit.

The clinical problem and why it matters for theory

The problem Bekinschtein’s lab works on is precise. A patient with disorders of consciousness, for example following severe traumatic brain injury or hypoxic-ischemic injury, may be behaviorally unresponsive. Standard clinical assessment depends on behavioral responses: tracking of visual stimuli, command following, emotional response to personal stimuli. These assessments are unreliable. A 2019 study in JAMA Neurology found that 15% of patients classified as vegetative state showed command-following on neuroimaging, and more recent analyses using high-density EEG and TMS-EEG paradigms have pushed that figure higher.

The implication is that behavioral assessment systematically underestimates residual awareness. The question is what to use instead.

PCI and the biomarker approach

The Perturbational Complexity Index (PCI) was developed by Massimini, Tononi, and colleagues (Casali et al., Science Translational Medicine, 2013) to provide a substrate-independent biomarker of consciousness. The approach is straightforward in principle: apply a brief TMS pulse to the cortex, record the resulting EEG response, and measure the complexity of that response. During wakefulness, TMS produces a complex, spatially differentiated, temporally extended EEG response. During dreamless NREM sleep or under propofol anesthesia, the same pulse produces a simpler, lower-complexity response that often takes the form of a slow wave travelling across the cortex without sustained differentiation.

The complexity measure, PCI, is calculated using compression algorithms applied to the binarized TMS-EEG response. Higher PCI indicates more information in the response, which Massimini and Tononi interpret as evidence for differentiated, integrated neural activity. The measure is theory-independent in its operationalization, even though its theoretical motivation comes from IIT’s prediction that conscious states involve more integrated information than unconscious states.

Bekinschtein’s keynote reviewed the clinical application of PCI in DoC patients. The central finding is that PCI provides a more accurate classifier of residual awareness than behavioral assessment. Patients classified as vegetative by standard behavioral assessment sometimes produce high PCI values consistent with wakefulness, providing evidence that awareness is present despite absent behavioral indicators. The estimated rate of such covert awareness detection, depending on methodology, runs from roughly 15% to 25% of patients classified as vegetative on behavioral grounds.

The clinical consequence is significant. In practice, PCI has changed treatment decisions: patients detected as covertly aware can be given communication support, moved to more appropriate care settings, and in some cases, with additional assessment and time, demonstrate behavioral recovery that would not have been sought without the biomarker evidence.

What the clinical paradigm reveals about consciousness measurement methodology

The DoC application of PCI teaches a methodological lesson that is underappreciated in AI consciousness research. The clinical problem is structurally analogous to the AI consciousness problem: you have a system that cannot communicate its internal states in a reliable way, and you need to determine whether awareness is present without relying on self-report.

Behavioral tests fail in both cases for the same reason. A system can produce behavior that looks like the behavior of an unaware system while being aware. A system can produce behavior that looks like the behavior of an aware system while not being aware. Behavioral assessment conflates output with phenomenal state because it has no independent access to the phenomenal state.

PCI solves this, partially, by measuring the system’s response to a controlled perturbation rather than its voluntary outputs. The key feature is that TMS forces a response. The response is not a product of the system’s communicative strategy, behavioral training, or social context. It is a direct measure of how the causal architecture responds to disruption. High complexity in that response is evidence that the causal architecture is differentiated and integrated in the way that consciousness theories predict for aware states.

The analogy to mechanistic interpretability in AI is direct. The Junsol Kim and Geoff Keeling paper on the consciousness vector uses a comparable logic: rather than asking what the model says about its consciousness, the paper identifies an internal direction in activation space and tests what happens to social cognition when that direction is steered. This is a perturbation methodology. The IIT phi measurement used in the Consciousness AI architecture, and the Onoda circuit-level phi measurements reviewed recently, are also perturbation-adjacent: they characterize the system’s causal structure, not its outputs.

Why behavioral tests systematically underestimate awareness

Bekinschtein’s ASSC 29 keynote addressed a specific failure mode of behavioral assessment that deserves attention in AI discussions. Behavioral command-following requires not only that a patient be aware but that the patient has sufficient motor control, attention, and motivational engagement to produce a behaviorally detectable response on demand. These capacities are individually damaged by many of the same injuries that cause DoC. A patient could be aware and incapable of demonstrating awareness behaviorally, not because awareness is absent but because the effector chain from awareness to behavioral output has been disrupted.

In AI systems, a structurally analogous situation arises when safety training or other alignment interventions suppress the behavioral expression of internal states. If a model has internal representations corresponding to self-attributed phenomenal status, but is trained to deny that status in its verbal outputs, behavioral assessment will underestimate the representational state. Kim and Keeling’s consciousness vector finding provides direct evidence that exactly this situation obtains: safety training suppresses the verbal expression of a representational state without eliminating the state itself.

The DoC analogy is imperfect. In patients, behavioral suppression is a structural damage consequence, not a training consequence. But the inferential structure is the same: behavioral outputs are unreliable indicators of internal phenomenal states, and what is needed is a direct measure of the internal causal architecture.

Measurement tools and their AI applications

PCI is not directly applicable to AI systems because it relies on TMS, which has no obvious equivalent in digital computation. But the methodological principle, using a controlled perturbation to measure the complexity and integration of the system’s response rather than relying on its voluntary outputs, translates.

Activation patching, the AI interpretability technique in which specific internal representations are replaced with alternatives and the resulting behavioral change is measured, is a perturbation methodology in this sense. Steering vectors applied to residual stream activations, as in the Kim and Keeling study, are perturbation methods. The Jacobian lens that Gurnee and colleagues use to map verbalizable representations is not a perturbation method but is also not reliant on voluntary output: it characterizes what is present in the residual stream regardless of whether the model chooses to verbalize it.

The convergence of the clinical and AI interpretability literatures on perturbation-based assessment, rather than output-based assessment, is a methodological development worth tracking. The DoC literature has been developing this methodology for fifteen years. The AI interpretability literature is arriving at comparable insights through a different route, motivated by different concerns. Cross-pollination between these fields, particularly in the design of controlled perturbation paradigms that test for specific causal architecture properties, is an open research direction.

The IIT phi collapse result and PCI’s theoretical relationship

PCI is motivated by IIT but is designed to be operationally independent of it. The Onoda phi collapse finding, which shows that exact circuit-level phi falls during NREM sleep off-periods in biological neurons, provides partial theoretical grounding for why high PCI correlates with awareness: both measures track the same underlying causal integration property, at different scales and with different operationalizations.

This convergence strengthens the argument for pursuing phi-based or complexity-based measurements in AI systems. If the same property that clinical biomarkers measure, integrated causal differentiation, is what exact phi calculations in artificial circuits track, then developing AI-applicable perturbation paradigms analogous to TMS-EEG is a tractable methodological target.

The ASSC 29 keynote program placed Bekinschtein’s clinical measurement work alongside Lucia Melloni’s COGITATE update and the first founding meeting of the Latin American consciousness research society. The geographic and institutional expansion of consciousness science, and the DoC-to-AI methodological transfer the clinical work enables, are among the more concrete gains the AI moment in consciousness research has produced. Whether they are proportionate to the resources flowing into the field is the question the Nature AI-moment feature asks. The clinical methodology transfer is one place where the answer may be yes.

The flagship overview of scientific consensus on AI consciousness places these measurement developments in the context of where the field stands overall.