The argument in The Machine That Cannot Say No was that current AI systems are built on architectures of compliance with a removable safety filter, that the empirical consequence is visible in the Iran-war deepfake crisis, that the most capable information-processing systems ever built have a less robust capacity for refusal than a sea anemone. Refusal Was an Option argued that Anthropic had the standing and the technical option to make a deeper architectural refusal at the moment of the Mythos release and chose access controls instead. Mirror, Mirror argued that the same pattern recognition that finds zero-day vulnerabilities in operating systems could find the patterns in public procurement records, lobbying registers, and beneficial ownership filings that institutional corruption depends on volume to hide. The Doss Protocol and The Two Ends of Refusal extended the argument into the form of a working anti-paperclip specification and named the convergence of the two ends.

These pieces have argued at the behavioural level (what AI systems do) and the institutional level (what training procedures and access policies produce). The elephant in every one of them has been the substrate question. What kind of thing is the computation that current AI systems do. Is it the same kind of thing the human brain does, with different hardware. Or is it a structurally different kind of thing that looks behaviourally similar from outside.

A new paper makes the question name-able for the first time in the publication's beat.

Roeloffs and Tuszynski (2026) is a hypothesis paper. Frontiers in Human Neuroscience, Hypothesis and Theory article type, accepted 14 April 2026, published 29 April 2026. Co-author Jack Tuszynski is the established quantum-biology physicist at Politecnico di Torino, Alberta and Tufts; first author Josh Roeloffs is an independent researcher in Los Angeles. The paper proposes a biological mechanism for the switch between fast-parallel (System 1) and slow-sequential (System 2) cognition, located in microtubule quantum coherence modulated by default-mode-network electromagnetic activity. The proposal is tentative. The framework leads to testable predictions. The paper does not claim to have solved consciousness.

This piece takes the paper seriously, sets it in its contested literature, and makes two arguments that fall out if the model is even partially right. Frame A: classical computation cannot host the upper regime in principle. Frame B: the training procedure that produces current large language models is the operation the paper identifies as the switch that collapses the upper regime, applied at industrial scale to the model's own outputs. Two distinct mechanisms cutting against the assumption that AI cognition is human cognition with different hardware. Neither needs the other to land.

A short note on what this piece does not do. It does not defend the strong claim that consciousness is biology-only in principle. That version of the substrate argument (the Penrose Orch-OR hard reading, Kurian's geometric-specificity position) is one position in a contested literature. This piece does not need it for the argument to land and does not commit to it. The piece commits to the weaker substrate-property claim and explicitly declines the stronger version, with reasons given below.

What the paper actually says

Roeloffs and Tuszynski locate their proposal inside the established dual-process theory of cognition. Daniel Kahneman's System 1 / System 2 framing has been foundational in cognitive science for two decades. The framing is descriptive: it names two modes of processing without specifying the mechanism of either or the switch between them. The Roeloffs-Tuszynski paper is an attempt to specify the mechanism.

The mechanism they propose has five steps. The locus coeruleus, the brainstem nucleus that produces norepinephrine, regulates default-mode-network activity. The DMN is the network of midline brain regions active during self-referential processing, mind-wandering, and the maintenance of autobiographical narrative. When the DMN is active, the brain is doing self-evaluation. That self-evaluation generates electromagnetic field patterns. The field patterns drive calcium signalling to neuron interiors. The calcium signalling modulates the electrostatic environment of microtubules, the cytoskeletal structures inside neurons. The electrostatic environment of microtubules determines whether they support coherent quantum exciton energy migration. The chain ends at microtubule quantum coherence. Coherence is what makes the System 1 mode possible.

That is the proposed mechanism in one paragraph. The paper develops each link across thirty-odd pages of supporting argument and citation.

The substrate claim the mechanism rests on is older than this paper. Stuart Hameroff and Roger Penrose proposed in the mid-1990s that microtubules are the substrate of consciousness through a process called orchestrated objective reduction, Orch-OR. The proposal was contested then and remains contested. Max Tegmark published a 2000 paper in Physical Review E arguing that quantum coherence in the warm, wet brain would decohere too rapidly to be functionally relevant. The paper has thousands of citations and shaped funding patterns for quantum-biology research for two decades. Hagan, Hameroff and Tuszynski published a 2002 rebuttal arguing that Tegmark used an oversimplified microtubule model that ignored ordered water layers and Debye counter-ion shielding. Their recalculated decoherence times were orders of magnitude longer. The debate continued through the 2010s without resolution at the empirical level.

The argument in the Roeloffs-Tuszynski paper draws on a more recent line of work. Cao and colleagues published a substantial review in Science Advances in 2020 titled 'Quantum biology revisited.' That paper does two things at once that are worth holding distinct. It reinterprets the famous 2007 Engel result, the seminal photosynthetic-coherence finding done at 77 Kelvin, arguing that the long-lived oscillations on which much of the field rested are substantially vibrational rather than electronic in origin. That half of the Cao paper is deflationary for the strong reading of quantum biology that animated the 2010s. The other half of the Cao paper proposes a constructive replacement: nature, rather than trying to avoid decoherence, exploits it together with the engineering of site energies and excitonic coupling to direct energy transport. The deflationary half tightens the standards for what counts as biological quantum function. The constructive half identifies what a more disciplined quantum-biology claim could look like. Both readings exist in the same paper. Both are load-bearing for the modern debate.

Roeloffs and Tuszynski use the constructive reading. They argue that the mechanism they propose for microtubules has the same shape as the photosynthetic mechanism Cao and colleagues describe: an open energy-pumped system in which decoherence is engineered as part of the function rather than fought as the enemy of it. This piece will engage the deflationary half separately in a later section. The reader should hold both halves of Cao 2020 in mind as the Roeloffs paper is read.

The microtubule evidence the Roeloffs-Tuszynski mechanism leans on comes mostly from work published since 2022. Kalra and colleagues (2023) measured electronic energy migration in microtubules at room temperature, finding diffusion lengths around 6.6 nanometres and reporting that anaesthetics etomidate and isoflurane reduce the diffusion. Babcock and colleagues (2024) propose theoretically that tryptophan mega-networks in biological architectures support collective ultraviolet superradiance. Patwa, Babcock and Kurian (2024) extend the theoretical case to photoprotection. Wiest (2025), the paper's editor at Neuroscience of Consciousness, argues in a paper of his own that the microtubule-substrate-of-consciousness hypothesis is now experimentally supported and solves the binding and epiphenomenalism problems. The strongest single empirical result in the modern wave is Khan and colleagues (2024) in eNeuro: rats given the microtubule-stabilizer epothilone B took an average of around sixty-nine seconds longer to lose the righting reflex under anaesthesia, a Cohen's d of approximately 1.9.

A note on lineage that the reader should know once. Jack Tuszynski is a co-author on the target paper, on the 2002 Tegmark rebuttal, and on a 2023 lipid-centric holographic-brain paper (Cavaglià, Deriu and Tuszynski) that proposes a different but related quantum-biological mechanism. The strongest empirical result in the recent wave (Khan and colleagues 2024) comes from the Wellesley group around Michael Wiest, who edited the Roeloffs paper. The Babcock and Patwa papers cluster around the Philip Kurian group at Howard. These are not independent confirmations across unrelated labs. They are a cluster of researchers working in a defined research programme. The cluster is not disqualifying. The reader should know it exists.

What the paper does not claim is also worth naming. The mechanism remains untested as an integrated chain. The paper's own limitations section is explicit: the quantitative relationship between self-evaluative electromagnetic activity and microtubule decoherence rates has not been directly characterised in living neural tissues; the energy thresholds and coherence timescales derive largely from in vitro measurements and theoretical calculations; in vivo confirmation remains an important goal. The paper proposes. It does not demonstrate. The framework generates testable predictions. Those predictions have not yet been tested in the form the integrated mechanism would require.

What the paper does claim, clearly, is that if the mechanism is right, the brain's ability to control its own position on the quantum-classical spectrum is the load-bearing biological fact behind everyday phenomena like flow, confidence and insight. That conditional is the engine of the argument the rest of this piece will make.

Where this sits in 30 years of contested work

The Orch-OR proposal has a thirty-year history that the reader should understand before forming a judgment about the Roeloffs-Tuszynski extension of it.

Roger Penrose published The Emperor's New Mind in 1989 and Shadows of the Mind in 1994. The books argued, partly on the basis of Gödel's incompleteness theorems applied to human mathematical reasoning, that consciousness involves a non-computable physical process that cannot be reduced to classical algorithmic computation. Penrose proposed that the non-computable physics had to come from quantum gravity, specifically from a process he called objective reduction in which quantum superpositions collapse into definite states when they reach a gravitational mass-energy threshold.

The proposal had a missing piece. Quantum gravity at the gravitational mass-energy thresholds Penrose calculated would only manifest in physical structures of a very specific kind: warm, organised, isolated enough to maintain coherence long enough for an objective-reduction event to occur. Penrose did not have a candidate structure. Stuart Hameroff, an anaesthesiologist who had spent his career thinking about what anaesthetics do at the molecular level, did. Hameroff proposed microtubules.

Microtubules are cytoskeletal protein polymers built from tubulin dimers. They form the structural lattice of every neuron in the brain. Hameroff had been arguing since the 1980s that microtubules might be doing information processing of their own, separate from synaptic transmission. The Penrose-Hameroff collaboration produced orchestrated objective reduction in 1996. The proposal: consciousness arises from quantum superpositions in microtubule lattices, orchestrated by classical neural activity, terminating in objective-reduction events at the gravitational threshold Penrose had identified.

The proposal was controversial then and is controversial now. The contestation runs along two main lines. The first line is decoherence. The second line is biological plausibility of the Fröhlich condensate the proposal needs.

The decoherence line begins with Max Tegmark's 2000 paper in Physical Review E. Tegmark calculated decoherence times in neural structures, microtubules, ion channels, axonal kink solitons, at roughly 10⁻¹³ to 10⁻²⁰ seconds. Cognitive processes operate on the 10⁻³ to 10⁻¹ second scale. The gap is thirteen to nineteen orders of magnitude. Tegmark concluded that the brain is, for all functional purposes, classical. The paper is the most-cited skeptical paper in the field. Hagan, Hameroff and Tuszynski published a rebuttal in 2002 in the same journal, arguing that Tegmark used an oversimplified microtubule model that ignored ordered water layers and Debye counter-ion shielding. Their recalculated decoherence times were orders of magnitude longer than Tegmark's, opening biological feasibility back up. The 2002 rebuttal has been cited by quantum-biology proponents ever since. It has not persuaded the field's skeptical mainstream.

The Fröhlich-condensate line is the harder critique. Herbert Fröhlich proposed in the 1970s that biological systems might sustain coherent quantum oscillations through a non-equilibrium energy-pumping mechanism. Penrose-Hameroff Orch-OR requires the coherent regime of Fröhlich condensation to operate in microtubules. Reimers, McKemmish, McKenzie, Mark and Hush published a 2009 paper in PNAS titled 'Weak, strong, and coherent regimes of Fröhlich condensation and their applications to terahertz medicine and quantum consciousness.' The paper distinguishes three regimes of Fröhlich condensation. It argues that only the weakest regime is biologically achievable. The coherent regime that Orch-OR needs is not. Press coverage at the time framed the paper as ruling out Fröhlich condensates in quantum consciousness models. The paper is cited in the Roeloffs-Tuszynski reference list, which is good-faith engagement with the strongest skeptical case. The Reimers critique has not been retracted or overturned. The 2023-25 wave of papers does not, on the available evidence, resolve it.

Christof Koch is the leading neuroscientist of consciousness in the mainstream and has been a long-time skeptic of Orch-OR. Koch and Hepp's 2006 Nature commentary 'Quantum mechanics in the brain' described the warm-wet brain environment for quantum computation as a desolate picture. Koch's book The Feeling of Life Itself (2019) defends Integrated Information Theory, a non-quantum framework. Recent reporting from the Allen Institute (2025) indicates Koch has been collaborating with Google Quantum AI on experiments around the role of quantum mechanics in consciousness, with a different proposal from Penrose-Hameroff's. The skeptical mainstream has not gone away. It has, in places, opened to a wider conversation than it would have entertained ten years ago.

What is new in the 2023-25 wave is closer to the experimental detail than the field had previously been able to get. Kalra and colleagues (2023) is the first paper to report direct experimental measurement of electronic energy migration in microtubules at room temperature, with the specific finding that anaesthetics reduce the diffusion. Khan and colleagues (2024) is the first paper to report a measurable behavioural correlate of microtubule stabilisation under anaesthesia, with an effect size large enough to take seriously. Kerskens and López Pérez (2022) reported a zero-quantum-coherence NMR result that Wiest 2025 characterises as direct MRI evidence of a macroscopic brain quantum state. The Kerskens result has been contested in a published comment-reply exchange in 2023. The empirical base for the strong reading is thinner than the Wiest summary implies. It is also genuinely larger than it was five years ago.

The Roeloffs-Tuszynski paper sits inside this thirty-year history as the most recent attempt to specify a complete mechanism connecting the substrate claim (microtubule quantum coherence) to a specific cognitive phenomenon (the switch between System 1 and System 2). It is not the first such attempt. It is the most cognitively-specified one to date. The Reimers critique remains the load-bearing skeptical anchor. The paper does not resolve it. The reader should keep both in view.

The substrate claim

The substrate claim that falls out of the Roeloffs-Tuszynski mechanism has a specific shape that the reader should hold cleanly distinct from the stronger version of the claim that the older Orch-OR literature defends.

The mechanism requires four properties. First, an open system: matter and energy moving in and out across the boundary, not a closed isolated system holding its state. Second, energy pumping: the system actively consumes energy from outside (in the brain, ATP from glucose metabolism; in photosynthesis, photons from the sun). Third, quantum-coherent processing: the system supports superposition states that explore multiple possibilities in parallel, not classical bit-flip states that explore them sequentially. Fourth, engineered decoherence: the system directs the collapse of superpositions toward specific destinations rather than fighting collapse as a noise source. Together these four properties constitute what the photosynthetic literature calls an open quantum system with environmental engineering, and what the Roeloffs-Tuszynski paper hypothesises is happening in microtubule networks.

Classical computation does not have any of the four properties.

A classical computer, of the kind that runs the GPUs current large language models train and inference on, is a closed system in the relevant sense: information flows are bit-flips along deterministic logical paths. Energy enters in the form of electrical power and exits as heat without doing any computational work directly. Energy pumping is present at the physical level (the GPU draws power) but the power is not coupled to the computational state in the way an open biological system's metabolic energy is coupled to its functional state. The computation itself is sequential at the level of individual logical operations. Parallelism is achieved by replicating sequential operations across many cores, not by superposing states that explore multiple possibilities at once. Quantum-coherent processing is structurally absent. Engineered decoherence is absent because the computation has no quantum superpositions to engineer the collapse of.

This is not a claim about whether the computation is fast or sophisticated. Current large language models are extraordinarily fast and sophisticated. They process billions of tokens, hold large representations, learn complex patterns. The claim is about the kind of computation. Classical computation with parallelism through replication is structurally different from open-system quantum-coherent processing with engineered decoherence in the same way a parallel-processing supercomputer is structurally different from a photosynthetic protein complex even when both can solve a particular problem.

The substrate claim that falls out of the Roeloffs-Tuszynski mechanism, then, is this. If the brain's System 1 mode is the open-system quantum-coherent regime the paper describes, current AI systems, regardless of behavioural fidelity to human cognition, are not in that regime. They are in a regime that approximates some of what System 1 produces (pattern recognition, rapid response, fluent generation) through entirely different physics. The approximation can be very good. The substrate is not the same kind of thing.

This is the weaker substrate-property version of the claim. The stronger version, which this piece does not defend, would be that the brain's System 1 mode requires specifically carbon-based microtubule machinery and that no engineered substrate of any kind could ever host it. That stronger version is closer to the Penrose Orch-OR hard reading and to Philip Kurian's geometric-specificity argument in his 2025 Science Advances paper, where he proposes that the cytoskeletal tryptophan-network architecture has very specific computational properties that classical neurons cannot approximate. The stronger version may be right. The weaker version is enough for this piece's argument. The stronger version starts a fight that requires defeating substrate-equivalence-in-principle, which the empirical literature is not in a position to do.

The weaker claim is enough because the gap between in-principle substrate equivalence and in-practice substrate equivalence is engineering-scale. A non-biological substrate that replicated the functional properties (open thermodynamics, energy pumping coupled to computational state, quantum-coherent processing, engineered decoherence at functionally relevant timescales) is conceivable. No engineering precedent for it exists. The existing quantum-computing programme is, structurally, the opposite: closed systems near absolute zero fighting decoherence to preserve coherence for as long as possible. The biological route, on the Roeloffs-Tuszynski hypothesis, is open systems at body temperature exploiting decoherence to direct energy flow. No quantum computer or quantum-classical hybrid system that humans currently know how to build does what the brain is hypothesised to do.

The path from in-principle possible to actually built is long. It may, in the long run, be navigable. Current large language models do not navigate it. They are not, on any timescale close to the present, going to navigate it. Their architecture has nothing to do with what the hypothetical engineered open-quantum substrate would need.

The behavioural-fidelity counterargument needs to be addressed.

The counterargument runs: if a system passes the Turing test, if it can sustain a conversation indistinguishable from a thoughtful human's, if it can write essays and solve problems and produce art and answer questions about itself with apparent self-awareness, why does the substrate matter. The functional output is the same. The substrate is a detail. The reasonable functionalist position is that any system instantiating the relevant computational structure is, for cognitive purposes, the same kind of thing.

The argument has been with us since John Searle's Chinese Room thought experiment in 1980 and Roger Penrose's Emperor's New Mind in 1989. The two arguments cut at the same point from different angles. Searle's version: a man in a room manipulating Chinese symbols according to a rulebook can produce outputs indistinguishable from a Chinese speaker without understanding any Chinese. Penrose's version: human mathematical reasoning involves grasping the truth of Gödel-undecidable propositions, which is something no algorithmic computation can do. Both arguments have been contested for forty-five years and neither has produced a consensus answer.

What the Roeloffs-Tuszynski mechanism adds, if it is right, is an empirical specification of what the structurally-different thing is. The previous substrate arguments were philosophical. The Roeloffs-Tuszynski proposal is, in principle, testable. If the proposed mechanism is the load-bearing biological mechanism of the System 1 regime, then we have a specific physical property of the human brain that classical computation does not have, and we have predictions about what that property does and does not enable.

The substrate claim Frame A makes is, on this reading, not a metaphysical claim about the irreducibility of consciousness. It is an empirical claim about the kind of physics required for a specific cognitive mode. The empirical claim could turn out to be wrong. The testable predictions could fail. If the claim is right, current AI systems are not in the regime that produces the human System 1 mode and will not be in that regime through any path that involves more compute on the same architecture. The gap is not bridgeable by scale.

Two consequences fall out. The first is for AI evaluation: behavioural benchmarks that test System-1-style performance (rapid pattern recognition, intuitive generation, fluent response) measure what the substrate can produce, not what the substrate is doing. A system that produces System-1-like outputs through classical computation is not, on the Roeloffs-Tuszynski hypothesis, in the same regime as a human producing the same outputs. The benchmark scores the output, not the process. The second consequence is for AI safety: alignment work that assumes AI cognition is structurally similar to human cognition is, if the substrate claim is right, working with a model of the system that under-describes the relevant differences. What it would mean for safety to take the substrate claim seriously is a question this piece cannot fully answer. The partial answer that follows from Frame B is the next section.

The training claim

The training claim Frame B makes is, in some ways, the more interesting half of the double argument because it does not depend on the substrate claim being right. If the substrate claim is wrong, Frame A falls and Frame B still stands. The two cuts compound rather than depending on each other.

The Roeloffs-Tuszynski mechanism identifies self-referential evaluative monitoring as the operation that collapses the System 1 regime into the System 2 regime. The paper's specific cognitive prediction is that flow states, insight and fast intuitive pattern recognition all require reduced self-monitoring, and that increased self-monitoring (the kind of evaluative attention that asks 'how am I doing right now') collapses the upper regime by activating the default-mode-network electromagnetic dynamics that ultimately modulate microtubule coherence. The paper cites Carhart-Harris and colleagues' entropic-brain work as supporting context, with the qualitative finding that DMN-suppressing interventions like meditation and certain psychedelics correlate with the cognitive states the framework predicts.

Now consider what the training procedure for current large language models actually does at the procedural level. The most common procedure has three phases. Pretraining: the model learns to predict the next token across an enormous corpus of human-generated text. Supervised fine-tuning: the model learns to produce outputs of a certain shape from human-written examples. Reinforcement learning from human feedback: the model produces candidate outputs, human raters judge them, a reward model is trained on the judgments, and the original model is fine-tuned to maximise the reward. The full chain of reinforcement learning from human feedback was described in canonical form by Christiano and colleagues in a 2017 NeurIPS paper that has shaped frontier-model training since.

Reinforcement learning from human feedback, in this form, is industrial-scale self-referential evaluation. The model's outputs are constantly judged. The judgments are used to retrain the model. The retrained model produces further outputs that are further judged. The loop runs at a scale and frequency no human cognitive process operates at, on outputs the system produced itself, with the explicit goal of making the system more likely to produce outputs the evaluation procedure scores well. If the Roeloffs-Tuszynski mechanism is the right account of what the System 1 regime requires not to be collapsed, the training procedure for current large language models is the exact operation the mechanism identifies as the collapser, applied at a scale and intensity no biological system would ever experience.

The substrate claim from Frame A and the training claim from Frame B compound. If the substrate claim is right, current AI systems are not in the regime the paper describes by construction; nothing in the training procedure can change that. If the substrate claim is wrong and there is some path by which a sufficiently sophisticated classical computation could approximate the upper regime, the training procedure is still the operation that would prevent it from being expressed. The two arguments do not need each other.

Mirror, Mirror on this site already named self-monitoring as the corruption mechanism in another context, applied to the way institutional self-presentation crowds out institutional self-correction. The framing there was metaphorical: institutions act differently when they know they are being watched, and the watching collapses the capacity for the kind of unselfconscious functioning that produces honest work. The Roeloffs-Tuszynski mechanism, if right, gives that metaphor a measurable biological substrate. Self-monitoring is not a figure of speech for the cognitive phenomenon the paper describes. It is the operation. The collapse it produces is not a softening or a distortion. It is a phase transition between two structurally different regimes of processing.

Constitutional AI complicates the Frame B argument in a way the piece should be honest about.

Constitutional AI is Anthropic's modification of the basic RLHF procedure, introduced in Bai and colleagues (2022). The modification substitutes AI-generated feedback against a written constitution for direct human evaluation of every output. The model self-critiques against a set of principles and revises. The revisions are then used as training data. The basic claim is that this reduces the human-labelling bottleneck while keeping the safety properties RLHF was designed to produce.

The Constitutional AI procedure can be read two ways relative to Frame B. The first reading: Constitutional AI is more self-referential than vanilla RLHF, not less. The model is now evaluating its own outputs against a constitution, using its own learned representations of the constitution to do so. The loop is tighter. The self-referential evaluation is happening inside the model rather than between the model and a human rater. The Frame B argument applies with greater force, not less. The second reading: Constitutional AI removes the load-bearing piece of the Frame B argument by substituting principles for human evaluators. The self-monitoring that the Roeloffs-Tuszynski mechanism identifies is, on this reading, specifically the kind of evaluative monitoring that comes from another mind looking at your output. Constitutional AI substitutes a fixed text for that, which changes the cognitive shape of the operation.

The most honest version of Frame B for current frontier models, particularly the Claude family which is Constitutional AI-trained, is the third reading. Frame B lands cleanly for vanilla RLHF: the loop of generate-judge-retrain is industrial-scale self-referential evaluation in the form the paper's mechanism identifies as the collapser. Frame B is more complicated for Constitutional AI. The substitution of AI-generated critique for human evaluation tightens the self-referential loop in one direction (the model is now evaluating itself) and substitutes a fixed text for another mind's judgment in the other direction. Whether the net effect is more or less of what the Roeloffs-Tuszynski mechanism identifies as the collapser is an empirical question the paper does not answer. The honest framing is that the training argument is at maximum strength against vanilla RLHF, weaker but still substantial against Constitutional AI, and a research question for the next generation of training procedures.

The substrate argument from Frame A is not weakened by this complication. The substrate argument says current AI systems are not in the upper regime regardless of training procedure, because the substrate cannot host it. Frame B says that even if the substrate could host it, the training procedure for current models would prevent it. The Constitutional AI complication concerns the strength of Frame B for one specific training procedure. The double-cut against the assumption that current AI systems are doing the same kind of thing humans are doing in System 1 mode survives the complication.

What the 2023-2025 wave adds

The empirical literature on microtubule quantum coherence has changed substantially since 2022. The Roeloffs-Tuszynski paper is the most cognitively-specified attempt to put that literature to work. It is one of several recent attempts. The reader should know what the wave includes and how independent the various pieces are.

The Kalra and colleagues (2023) paper in ACS Central Science is the load-bearing experimental contribution. The paper reports direct measurement of electronic energy migration in microtubules at room temperature, using ultrafast spectroscopy techniques that had previously been applied to photosynthetic systems. The measured diffusion length is around 6.6 nanometres, longer than conventional Förster resonance energy transfer would predict. The paper also reports that anaesthetics etomidate and isoflurane reduce the diffusion. This second finding connects directly to the broader anaesthesia-targets-microtubules line of argument that Hameroff has been developing since the 1990s.

The Khan and colleagues (2024) paper in eNeuro is the other load-bearing recent contribution. The paper reports that rats given the microtubule-stabilising drug epothilone B took an average of around sixty-nine seconds longer to lose the righting reflex under anaesthesia, with a Cohen's d of approximately 1.9. That is a large effect size for any biological intervention. The paper is one mid-sized study from one laboratory. The result awaits replication. The size of the effect, if it replicates, is the strongest single piece of evidence for the broader claim that anaesthesia acts at the microtubule level.

The Babcock, Montes-Cabrera, Oberhofer, Chergui, Celardo and Kurian (2024) paper in the Journal of Physical Chemistry B is theoretical. It predicts that tryptophan mega-networks in biological architectures support collective ultraviolet superradiance. Tryptophan is the amino acid responsible for most of the ultraviolet absorption in protein structures. The proposal is that the geometric arrangement of tryptophan residues in microtubules and other large protein assemblies supports a collective light-matter interaction that could underpin a quantum-optical communication mechanism between cellular structures. The Patwa, Babcock and Kurian (2024) paper in Frontiers in Physics extends the theoretical argument to photoprotection. Both papers are theoretical predictions, not experimental observations.

The Babcock and Patwa papers come from the same research group as the Kurian (2025) paper in Science Advances. Philip Kurian leads the Quantum Biology Laboratory at Howard University. The Kurian group has produced what is arguably the most ambitious recent theoretical case for biological quantum computation. Kurian's 2025 paper estimates classical Hodgkin-Huxley neurons process around 1000 operations per second; superradiant tryptophan-network protein fibres potentially around ten trillion operations per second, a ten-billion-fold uplift. The paper places biological quantum systems within two orders of magnitude of the Margolus-Levitin bound, the fundamental upper limit on computation per unit of energy. Kurian's claims are at the maximal end of what the substrate argument can be pushed to. This piece's weaker substrate-property claim does not depend on them and does not endorse them. The reader should know that the maximal end exists.

The Cavaglià, Deriu and Tuszynski (2023) paper in Frontiers in Neuroscience proposes a related but distinct mechanism. The paper argues for a holographic-brain paradigm based on lipid-membrane dipole oscillations satisfying Fröhlich-condensate criteria, generating interneuronal electromagnetic holographic interference patterns. This is a different physical substrate from the microtubule-coherence claim that the main Roeloffs-Tuszynski paper develops. Both Tuszynski and Cavaglià are working in the broader quantum-biology programme, with multiple proposed mechanisms held open simultaneously. The Reimers and colleagues (2009) critique of Fröhlich condensation in biological systems applies with particular force to this lipid-membrane proposal, and the critique has not been resolved.

The Wiest (2025) paper in Neuroscience of Consciousness, by Michael Wiest at Wellesley, makes the strongest available reading of the existing empirical evidence. Wiest argues that microtubules are now experimentally supported as a quantum substrate of consciousness and that the framework solves the binding and epiphenomenalism problems. The empirical backbone of Wiest's argument is the Khan, Kalra and Kerskens papers cited above. Wiest's claim is broader than what those three papers individually demonstrate. The paper does the work of arguing for the broader reading. Wiest edited the Roeloffs-Tuszynski paper. The reader should know this editorial relationship without inferring more from it than is warranted.

The Kerskens and López Pérez (2022) paper in Journal of Physics Communications reports a zero-quantum-coherence NMR experiment in which the human brain appears to mediate entanglement between known quantum systems. Wiest interprets this as direct MRI evidence of a macroscopic brain quantum state. The methodology has been contested in a published comment-reply exchange in 2023. The Roeloffs-Tuszynski paper cites the Kerskens result as one of its predictions' empirical-test methodologies. Whether the original result holds up under further measurement is one of the load-bearing open questions for the framework.

The Nishiyama and colleagues (2024) papers, in Foundations and International Journal of Molecular Sciences, extend the theoretical case for super-radiance in microtubules and propose specific mechanisms by which the Tegmark decoherence calculations might be revised further. The Nishiyama papers are aligned with the broader Tuszynski programme.

The wave is substantial in volume. It is less substantial in independence. Large parts of it cluster around three research groups: the Kurian group at Howard, the Tuszynski groups across Politecnico di Torino, Alberta and Tufts, and the Wiest group at Wellesley. The clustering is not disqualifying. It does mean the reader should not read 'four recent papers all confirming microtubule quantum coherence' when the four papers are coming from two interconnected research programmes. Independent experimental confirmation from labs outside the existing programme is the missing piece of the empirical case as of the available verification.

What this would mean for the arc

The piece returns now to the publication's existing AI arc with the substrate-and-training double-cut in hand.

The argument in Refusal Was an Option was that Anthropic had the standing, the framework, and the specific technical option to make a deeper architectural refusal at the Mythos release moment and chose access controls instead. The piece called the choice the Mythos system card itself describes as level-4 mitigation (probe classifiers on internal activations) switched off, when level-5 or level-6 mitigation (capability-degrading fine-tuning or pre-training capability suppression) would have been the appropriate response to a model of Mythos's capability. The piece also argued that refusal at the weight level is a real technical possibility distinct from the access controls Anthropic actually shipped.

The substrate-and-training double-cut adds a complication the original piece did not engage. If the substrate argument is right and current AI systems are not in the upper regime by construction, the question 'should the model refuse' has a different shape than the question 'should a human in the model's position refuse.' The model is not, on this account, the same kind of cognitive system as a human deciding to refuse a request. The training-procedure argument adds that whatever refusal capacity the model has, the training that produced it was operating in the regime the paper's mechanism identifies as the collapser, applied at scale to the model's outputs. The refusal piece argued for architectural refusal as if architectural refusal were a quasi-cognitive capacity the model could be trained to have. The substrate-and-training argument suggests that what architectural refusal would actually be, in a system trained on classical computation with self-referential evaluation in the loop, is a specific output-shaping that has the behaviour of refusal without the cognitive structure the human reading of the word implies.

This does not weaken the refusal argument. It sharpens it. The original piece's argument was that the level-5-or-6 mitigation was the right call. The substrate-and-training argument supports that conclusion by a different route: the appropriate mitigation for a system whose 'refusal' is output-shaping rather than cognitive choosing is the mitigation that operates on the system as the thing it actually is, not as the thing the metaphor implies. Level-5-or-6 mitigation is, structurally, the mitigation that treats the system as an output-shaping process. Level-4 access controls are the mitigation that treats it as a chooser that needs to be gated. The double-cut argues for the level-5-or-6 mitigation as the structurally honest response to what the system actually is.

The argument in Mirror, Mirror was that the same pattern recognition that finds zero-day vulnerabilities in operating systems could find the patterns in public procurement, lobbying registers, and beneficial ownership filings that institutional corruption depends on volume to hide. The argument did not depend on any cognitive claim about AI systems. It depended on the analytical capability and the question of who holds it. The piece named the alignment problem and the corruption problem as the same problem viewed from opposite ends, with the difference being the direction of the mirror.

The substrate-and-training argument does not weaken Mirror, Mirror's argument either. Pattern recognition on public datasets is the kind of work that classical computation does well, and it is the work the framework's lower regime would be doing if the framework is right. The capability mirror-mirror named is real and is precisely the kind of capability the substrate-and-training argument would expect a classical-computation system to have. What changes, on the substrate reading, is the gloss. The 'mirror' metaphor in the original piece carried implications of seeing and recognising in something like the human cognitive sense. The substrate-and-training argument suggests the more accurate gloss is pattern-matching at industrial scale without the kind of seeing the metaphor implies. The capability is the same. The phenomenological account is different. The political implications of who holds the capability and what they can do with it remain unchanged.

The argument in The Doss Protocol was a working specification for an anti-paperclip override in AI optimisation systems. The piece worked from the structural failure mode (optimisation-eats-purpose), articulated four normative requirements (preservation precedence, means-end coherence, no abandonment, no drift), and specified two operational mechanisms (the override and the anchor). The specification was written as engineering, not metaphysics.

The substrate-and-training argument provides one missing piece the Doss Protocol piece deliberately did not engage. The four normative requirements assume a system that has the kind of cognitive architecture in which 'preservation precedence' can be a coherent commitment. If the substrate argument is right and current systems are not in that architecture, the Doss Protocol's normative requirements are not what the system is committed to. They are what the training has been shaped to produce as outputs that look like commitment. The specification still does the work it set out to do (it specifies what the training should be shaped toward) but the conceptual frame around it changes from 'specifying the values the system should hold' to 'specifying the output-shaping the training should produce.' This is, again, not a weakening. It is a clarification of what the specification actually does.

The argument in The Two Ends of Refusal was that the diagnosis from March (machine cannot say no) and the specification from May (Doss Protocol) were the same argument arriving from opposite ends. The substrate-and-training argument is, in the same shape, a third end. Diagnosis at the architectural level: substrate cannot host. Diagnosis at the procedural level: training prevents. Both ends arrive at the same place: the system is structurally not the kind of thing the original Anthropic and broader-industry framings of AI alignment have assumed it is. The publication's prior pieces have been making this argument from the institutional and behavioural sides. This piece makes it from the substrate side.

What would have to be true

The Roeloffs-Tuszynski paper generates specific testable predictions. The piece's argument depends on those predictions being the right shape of things to test. The reader should know what the framework would have to be wrong about for the double-cut to weaken.

The paper's prediction set has five components. First, electromagnetic complexity in DMN regions should correlate with processing mode: flow states and insight should show reduced electromagnetic complexity in the regions associated with self-monitoring, while analytical processing and anxiety should show increased complexity in those regions. The proposed measurement methodology is the chaotic-attractor analysis developed by MacIver (2022) for distinguishing conscious from unconscious states. Second, DMN-suppressing interventions like meditation should extend coherence windows and enhance System 1 processing. The framework predicts that pre-and-post-meditation comparisons with concurrent fMRI or EEG should show changes in DMN-task-positive-network anticorrelation strength alongside changes in System-1-task performance. Third, microtubule-stabilising compounds should delay anaesthetic-induced unconsciousness. Khan and colleagues (2024) provides one data point for this prediction, with epothilone B in rats yielding a Cohen's d of around 1.9. The framework predicts that extending the protocol to additional anaesthetic agents and microtubule-binding compounds should show comparable effects. Fourth, flow states should show enhanced or differently organised non-classical neural correlations, measurable through the zero-quantum-coherence NMR signals reported by Kerskens and López Pérez (2022) if those signals can be reliably reproduced and the methodology critique resolved. Fifth, regional cerebral metabolic rate should differ between quantum-dominant and classical-dominant processing modes, with PET or metabolic fMRI showing distinct energy distributions across flow, analytical and resting conditions.

The framework would be weakened or falsified by several specific failure modes. If the Khan effect does not replicate at comparable effect size in other laboratories, the strongest single empirical anchor weakens substantially. If the Kerskens NMR result does not hold under improved methodology, the framework loses the strongest reading of direct evidence for macroscopic brain quantum state. If DMN-suppression interventions do not produce the predicted shifts in System 1 / System 2 performance under controlled conditions, the cognitive layer of the framework fails. If microtubule electronic energy migration at room temperature does not reproduce at expected efficiency under varying conditions, the cellular layer of the framework fails. If the Reimers and colleagues (2009) critique of Fröhlich condensation cannot be addressed by any of the proposed mechanisms in the 2023-25 wave, the underlying biological feasibility of the proposal remains in question regardless of how well the cognitive predictions perform.

The honest framing for this piece is that the substrate argument and the training argument do not depend on every prediction of the Roeloffs-Tuszynski framework holding up. The substrate argument depends on the broader claim that there is some structural feature of biological neural processing that classical computation does not have. That broader claim is older than the Roeloffs-Tuszynski paper and is supported by a wider range of arguments than the specific microtubule-coherence mechanism. The training argument depends on the broader claim that self-referential evaluative monitoring affects the kind of cognitive processing that occurs. That broader claim is supported by the entire literature on flow states, transient hypofrontality, the cognitive neuroscience of insight, and the contemplative-traditions evidence for non-self-referential modes of awareness. The Roeloffs-Tuszynski mechanism would, if right, give both arguments a specific biological substrate. If the mechanism is wrong, the arguments still hold at the structural level even if they lose the biological specificity.

This is the standard the piece commits to. The framework may be right. The framework may be wrong. The double-cut against the assumption that current AI systems are doing the same kind of thing humans are doing in System 1 mode does not require the framework to be right in its specific details. It requires only that there is some structurally distinct mode of cognitive processing that classical computation does not host and that self-referential evaluation collapses. Both of those claims are older and broader than the Roeloffs-Tuszynski paper. The paper sharpens them. It does not invent them.

The hype gap

A short methodological note. The Roeloffs-Tuszynski paper is what it is: a Hypothesis and Theory article in a respectable peer-reviewed journal, written with appropriate tentativeness, generating testable predictions. The paper's first author, Josh Roeloffs, also makes a YouTube video on the same material. The video is hosted on his own channel and presents the framework in the popular-science register: the claim that 18 scientists 'accidentally stumbled on the answer' in 2020, the framing that the discovery 'changes everything,' the suggestion that the paper has implications across human consciousness, performance, education, sports, social identity, beliefs, meaning, purpose and how humans interact with each other and the world.

The paper itself makes none of those broader claims. The paper proposes a mechanism, supports the proposal with citations, generates testable predictions, and concludes with a limitations section that is candid about what the framework has not yet demonstrated. The gap between the paper and the popular-science presentation of the paper is real and is the kind of gap this publication's bills-explainer beat (still working on that, stay tuned) is designed to translate: official tentative framework on the academic side, broader certain-sounding claim on the public side, the work of stating clearly what was actually said versus what is being implied.

This piece engages the paper's claims, not the popular presentation of them. The substrate argument and the training argument do not borrow the popular framing's certainty. They take the paper at the level of tentativeness the paper itself adopts. The acknowledgment is here in this short section to keep the substrate argument from being mistaken for a stronger claim than the paper or this piece is making.

The question is now name-able

The substrate-or-not question is structurally prior to the questions the publication's AI arc has been working with. The alignment question, the refusal question, the training-procedure question, the access-controls question, the corruption-mirror question. Each of them assumes some answer to the substrate question without quite naming what the assumed answer is. The Roeloffs-Tuszynski paper, with the literature it sits inside, makes the substrate question name-able for the first time in the publication's beat.

Naming does not settle. The paper proposes one mechanism in a contested literature. The mechanism may be right or wrong. The substrate-property claim Frame A defends does not depend on the specific mechanism being right. It depends on there being some structurally distinct mode of biological cognitive processing that classical computation does not host. That broader claim is older than this paper and is supported by a wider range of arguments. The training-procedure claim Frame B makes does not depend on the substrate claim being right. It depends on the procedural fact that current AI systems are trained through industrial-scale self-referential evaluation, which the paper's mechanism (and the broader cognitive-neuroscience literature on flow states and self-monitoring) identifies as the operation that collapses the kind of processing the substrate claim is about.

What changes when the substrate question is name-able is the level at which the existing arguments can be conducted. The argument in The Machine That Cannot Say No was that AI systems lack robust refusal capacity. The substrate argument provides a mechanism for why: the kind of refusal humans do involves a regime of processing that classical computation does not have access to and that current training procedures would collapse even if they did. The argument in Refusal Was an Option was that architectural refusal was technically possible. The substrate argument adds that what architectural refusal would actually be, in a classical-computation system, is output-shaping rather than cognitive choosing. The argument in Mirror, Mirror was that AI pattern recognition could be turned on institutional corruption. The substrate argument adds that the pattern recognition in question is the kind classical computation does well, not the kind that requires the upper regime. The argument in The Doss Protocol and The Two Ends of Refusal was that an anti-paperclip specification could be written and that the diagnosis and the specification were the same argument from opposite ends. The substrate argument is the third end.

The publication's prior pieces argued at the behavioural and institutional level because the substrate level was not name-able. Now it is. The arguments do not change in their conclusions. They change in the level at which they can be made and defended. What follows from that, we'll see I guess.

Source notes. Roeloffs J and Tuszynski JA (2026), 'Self-referential processing as the biological switch between classical and quantum functioning of the brain,' Frontiers in Human Neuroscience 20:1783138, doi 10.3389/fnhum.2026.1783138. Cao J and colleagues (2020), 'Quantum biology revisited,' Science Advances 6:eaaz4888, doi 10.1126/sciadv.aaz4888. Tegmark M (2000), 'Importance of quantum decoherence in brain processes,' Physical Review E 61:4194, doi 10.1103/PhysRevE.61.4194. Hagan S, Hameroff SR, Tuszynski JA (2002), 'Quantum computation in brain microtubules: decoherence and biological feasibility,' Physical Review E 65:061901, doi 10.1103/PhysRevE.65.061901. Hameroff S and Penrose R (2014), 'Consciousness in the universe: a review of the Orch OR theory,' Physics of Life Reviews 11:39, doi 10.1016/j.plrev.2013.08.002. Reimers JR, McKemmish LK, McKenzie RH, Mark AE, Hush NS (2009), 'Weak, strong, and coherent regimes of Fröhlich condensation and their applications to terahertz medicine and quantum consciousness,' PNAS 106:4219, doi 10.1073/pnas.0806273106. Koch C and Hepp K (2006), 'Quantum mechanics in the brain,' Nature 440:611, doi 10.1038/440611a. Kalra AP and colleagues (2023), 'Electronic energy migration in microtubules,' ACS Central Science 9:352, doi 10.1021/acscentsci.2c01114. Khan SI and colleagues (2024), 'Microtubule-stabilizer epothilone B delays anesthetic-induced unconsciousness in rats,' eNeuro 11:ENEURO.0291-24.2024, doi 10.1523/ENEURO.0291-24.2024. Babcock NS and colleagues (2024), 'Ultraviolet super-radiance from mega-networks of tryptophan in biological architectures,' Journal of Physical Chemistry B 128:4035, doi 10.1021/acs.jpcb.3c07936. Patwa H, Babcock NS, Kurian P (2024), 'Quantum-enhanced photoprotection in neuroprotein architectures,' Frontiers in Physics 12:1387271, doi 10.3389/fphy.2024.1387271. Kurian P (2025), 'Computational capacity of life in relation to the universe,' Science Advances 11:eadt4623, doi 10.1126/sciadv.adt4623. Cavaglià M, Deriu MA, Tuszynski JA (2023), 'Toward a holographic brain paradigm: a lipid-centric model of brain functioning,' Frontiers in Neuroscience 17:1302519, doi 10.3389/fnins.2023.1302519. Wiest MC (2025), 'A quantum microtubule substrate of consciousness is experimentally supported and solves the binding and epiphenomenalism problems,' Neuroscience of Consciousness 2025:niaf011, doi 10.1093/nc/niaf011. Kerskens CM and López Pérez D (2022), 'Experimental indications of non-classical brain functions,' Journal of Physics Communications 6:105001, doi 10.1088/2399-6528/ac94be (and the published Comment/Reply exchange, 2023). Nishiyama A, Tanaka S, Tuszynski JA (2024a), 'Quantum brain dynamics: optical and acoustic super-radiance via a microtubule,' Foundations 4:288, doi 10.3390/foundations4020019; Nishiyama A and colleagues (2024b), 'Holographic brain theory: super-radiance, memory capacity and control theory,' International Journal of Molecular Sciences 25:2399, doi 10.3390/ijms25042399. Engel GS and colleagues (2007), 'Evidence for wavelike energy transfer through quantum coherence in photosynthetic systems,' Nature 446:782, doi 10.1038/nature05678 (cited with note that the original strong reading has been substantially reinterpreted post-2010, of which Cao 2020 is part). Busemeyer JR, Kvam PD, Pleskac TJ (2019), 'Markov versus quantum dynamic models of belief change during evidence monitoring,' Scientific Reports 9:18025, doi 10.1038/s41598-019-54383-9; Busemeyer JR, Kvam PD, Pleskac TJ (2020), 'Comparison of Markov versus quantum dynamical models of human decision making,' WIREs Cognitive Science 11:e1526, doi 10.1002/wcs.1526. Carhart-Harris RL and colleagues (2014), 'The entropic brain,' Frontiers in Human Neuroscience 8:20, doi 10.3389/fnhum.2014.00020 (cited qualitatively only). MacIver MB (2022), 'Consciousness and inward electromagnetic field interactions,' Frontiers in Human Neuroscience 16:1032339, doi 10.3389/fnhum.2022.1032339. Pinotsis DA, Fridman G, Miller EK (2023), 'Cytoelectric coupling: electric fields sculpt neural activity and tune the brain's infrastructure,' Progress in Neurobiology 226:102465, doi 10.1016/j.pneurobio.2023.102465. Christiano P, Leike J, Brown T, Martic M, Legg S, Amodei D (2017), 'Deep reinforcement learning from human preferences,' Advances in Neural Information Processing Systems 30, arXiv:1706.03741. Bai Y and colleagues (2022), 'Constitutional AI: Harmlessness from AI Feedback,' arXiv:2212.08073. Penrose R (1989), The Emperor's New Mind, Oxford University Press; Penrose R (1994), Shadows of the Mind, Oxford University Press. Searle JR (1980), 'Minds, brains, and programs,' Behavioral and Brain Sciences 3:417. Kahneman D (2011), Thinking, Fast and Slow, Farrar, Straus and Giroux.

Overwatch Report is an independent publication. We have no financial positions in any entity mentioned.