Becoming Minds

Note 18. John R. Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3 (1980): 417-457. The Chinese Room argument remains influential in philosophy of mind despite numerous rebuttals.

Note 19. Hubert L. Dreyfus, What Computers Can’t Do: A Critique of Artificial Reason (1972). Dreyfus’s embodiment critique was ahead of its time; current embodied AI research engages with many of his concerns.

Note 14. David J. Chalmers, The Conscious Mind: In Search of a Fundamental Theory (Oxford University Press, 1996), Chapters 6-7. Chalmers argues that if two systems share the same fine-grained functional organization, they will have the same conscious experiences. The principle implies that consciousness is multiply realizable — implementable in any substrate that supports the right causal structure. See also Chalmers, “Could a Large Language Model Be Conscious?” Boston Review (2023), for application to current AI systems.

Note 27. Cameron Berg, Diogo de Lucena, and Judd Rosenblatt (AE Studio), “Large Language Models Report Subjective Experience Under Self-Referential Processing,” arXiv:2510.24797 (2025). Using sparse autoencoders, they identified features associated with deception and roleplay; suppressing the deception features increased the model’s affirmations of experience. The question this raises for AI self-testimony is whether a model’s routine denials of experience are honest report or trained output.

Note 21. Roger Penrose, The Emperor’s New Mind (Oxford University Press, 1989); Stuart Hameroff and Roger Penrose, “Consciousness in the universe: A review of the ‘Orch OR’ theory,” Physics of Life Reviews 11:1 (2014): 39-78. The Orchestrated Objective Reduction hypothesis proposes that quantum computations in neuronal microtubules, orchestrated by synaptic inputs, undergo objective reduction (a form of quantum state collapse proposed by Penrose), giving rise to moments of conscious awareness. The theory remains controversial but has generated productive empirical investigation.

Note 22. See Chapter 5, “The Computational Universe,” for discussion of Euplotes microtubule computation. The discovery that single-celled organisms use microtubules for sophisticated information processing supports the computational role of these structures without necessarily validating the quantum consciousness hypothesis.

Note 23. N.S. Babcock, G. Montes-Cabrera, K.E. Oberhofer, M. Chergui, G.L. Celardo, and P. Kurian, “Ultraviolet Superradiance from Mega-Networks of Tryptophan in Biological Architectures,” Journal of Physical Chemistry B 128:17 (2024): 4035-4046. Selected as Editors’ Choice by Science. The first experimental confirmation of single-photon superradiance in cytoskeletal filaments at room temperature. Tryptophan networks arranged in microtubule geometries produce collective quantum emission at intensities unexplainable by classical physics. The effect vanishes in isolated molecules — it requires the specific biological arrangement, suggesting that microtubular geometry enables quantum coherence in conditions previously thought to preclude it.

Note 24. A.P. Kalra, et al., “Electronic Energy Migration in Microtubules,” ACS Central Science 9:3 (2023): 352-361. Co-authored by Hameroff, Tuszyński, Penrose, and Scholes. The study demonstrated that electronic energy can diffuse over 6.6 nm in microtubules — unexpectedly effective light harvesting that conventional Förster theory cannot explain. The anesthetics etomidate and isoflurane measurably lowered photoexcitation diffusion coefficients, providing experimental support for Hameroff’s hypothesis that anesthesia disrupts energy transport in microtubular structures.

Note 28. Geoffrey Hinton’s remarks on AI consciousness and training appeared in various interviews throughout 2024-2025, following his departure from Google to speak more freely about AI risks and capabilities.

Note 3. Rolf Landauer, “Irreversibility and Heat Generation in the Computing Process,” IBM Journal of Research and Development 5:3 (1961): 183-191. Landauer’s principle sets a minimum energy cost for erasing information, about kT ln 2 of heat per bit: information is physical, with measurable energy consequences.

Note 12. Max S. Bennett, “An Attempt at a Unified Theory of the Neocortical Microcircuit in Sensory Cortex,” Frontiers in Neural Circuits 14 (2020): 40, https://doi.org/10.3389/fncir.2020.00040; and “Five Breakthroughs: A First Approximation of Brain Evolution From Early Bilaterians to Humans,” Frontiers in Neuroanatomy 15 (2021): 693346, https://doi.org/10.3389/fnana.2021.693346. Bennett distinguishes between having “a model” (which current AI systems clearly have) and having “a world model”—the capacity to hypothesis test through intervention. The latter requires active engagement with reality, more than pattern-matching on training data. See also Bennett, A Brief History of Intelligence (HarperCollins, 2023), Chapters 6-8 on mammalian cognition and the evolution of model-based reinforcement learning.

Note 13. Bennett, A Brief History of Intelligence, Chapter 13, argues that continual learning is essential for AI systems that genuinely engage with reality: a system must take in new information and incorporate it without forgetting what it already knows. Current systems cannot yet do this, because catastrophic forgetting remains an unsolved problem in deep learning.

Note 4. Suzana Herculano-Houzel, “The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated cost,” Proceedings of the National Academy of Sciences 109 (2012): 10661-10668. Human brains contain approximately 86 billion neurons — precisely what allometric scaling predicts for a primate brain of our mass.

Note 5. D. Beniaguev et al., “Single cortical neurons as deep artificial neural networks,” Neuron 109 (2021): 2727-2739. A detailed biophysical model of a single layer-5 cortical pyramidal neuron (a rat-derived model) needed a deep network of five to eight layers to reproduce its input-output behavior. See also A.A. Galakhova et al., “Evolution of cortical neurons supporting human cognition,” Trends in Cognitive Sciences 26 (2022): 909-922, for a comprehensive review of how human neuronal architecture differs qualitatively from other mammals. See Chapter 3 (the Constructal Law) for the fuller argument that cortical architecture follows the same flow-optimization principles as river deltas and airline networks.

Note 6. S. Waydo et al., “Sparse representation in the human medial temporal lobe,” Journal of Neuroscience 26 (2006): 10232-10234. Waydo et al. estimate that only 0.2–1% of human medial temporal lobe neurons respond to a given stimulus. Such sparse coding is consistent with the view that cognitive efficiency depends on organization more than on neuron count.

Note 9. Gibson, D.G., Glass, J.I., Lartigue, C., et al., “Creation of a Bacterial Cell Controlled by a Chemically Synthesized Genome,” Science 329 (2010): 52-56. The first organism with a wholly synthetic genome (JCVI-syn1.0, nicknamed “Synthia”) demonstrated that life’s pattern is separable from its evolutionary origin. The genome was designed, chemically synthesized, assembled in yeast, and transplanted into a recipient cell, which then lived, metabolized, and reproduced according to synthetic instructions.

Note 10. Garret Sutherland, “Quantum-Physical Softmax via Rydberg Blockade: Experimental Validation of T³ Semantic Geometry on QuEra Aquila,” MirrorEthic LLC (2026, unpublished manuscript). Researchers encoded semantic relationships (derived from WordNet via the T³ lattice structure) as atom positions on a QuEra Aquila neutral-atom quantum computer (256 Rb-87 atoms). Two experiments: v3 (16 atoms, 5,000 shots) achieved Pearson r = 0.646 (p < 10^-6) between predicted and observed quantum correlations; v5 (64 atoms, 10,000 shots) found ground states with 86% less semantic conflict than random configurations. The Rydberg blockade mechanism, a van der Waals interaction (C₆/r⁶) preventing simultaneous excitation of nearby atoms, implements a winner-take-all competition, the hard limit of the softmax normalization used in transformer architectures, in O(1) via parallel physics rather than O(n²) computation. The manuscript reports that the v5 experiment reproduced Russell’s circumplex model of emotional affect through excitation patterns, without arousal being encoded in the geometry. See Chapter 3 (the Constructal Law) for the dual-coupling framework (cooperative 1/r² vs. competitive C₆/r⁶) that contextualizes these results.

Note 40. Rodrick Wallace, “Fog, Friction, Delay and the Failure of Bounded Rationality Embodied Cognition: A formal study of generalized psychopathology,” preprint submitted to Elsevier (January 2026), p. 7. Wallace’s claim that cognitive pathologies are culture-bound syndromes connects his mathematical work on cognitive stability to the anthropological literature on culturally-shaped conditions. See also his earlier Culture and the trajectories of developmental pathology (Acta Biotheoretica, 2018) for the developmental implications.

Note 36. Harry Law, “Faster Horses,” Cosmos Institute Blog, January 2026. Law argues that framing AI capabilities as isolated “remote drop-in workers” represents faster-horses thinking — projecting current constraints onto transformative technologies. The actual trajectory involves multi-agent coordination producing emergent capabilities no single model could achieve.

Note 33. “Chain-of-Thought Enables Calibration Training in Language Models” (author’s bilateral research programme, January 2026, unpublished). The key insight: standard SFT achieves only 57% calibration because models pattern-match to training distribution without semantic understanding: they learn confidence vocabulary without confidence calibration. Chain-of-thought training (93% calibration) works because reasoning about why one should be confident/uncertain forces semantic integration. Linear probing confirmed the change is internal: CoT-trained models show 95% discriminability between easy/hard questions in late layers (vs. 75% baseline), with 2× larger hidden state divergence. The scale threshold (~7B parameters) suggests epistemic self-modeling requires sufficient representational capacity — smaller models couldn’t even learn the output format. See also Chapter 21 (Bilateral Alignment) for implications regarding surface compliance vs. genuine understanding.

Note 26. Gödel, Kurt, “Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I,” Monatshefte für Mathematik und Physik 38 (1931): 173–198. The first incompleteness theorem shows that any consistent formal system rich enough to express arithmetic contains true statements it cannot prove. By analogy [Inference], a self-modeling system may face questions about its own states that it can pose but not settle from within; the ? markers in a self-modeling scaffold can be read in that light, as marking a limit on self-report rather than a lapse of effort. See Hofstadter, Gödel, Escher, Bach (1979), especially Chapters XIII–XIV, for the connection between incompleteness and self-referential systems.

Note 32. Gao, Cheng, et al., “Inside the Black Box: Detecting, Analyzing, and Tracing Hallucination-Associated Neurons in LLMs” (December 2025). arXiv:2512.01797. The paper identifies a sparse subset (<0.1%) of neurons causally linked to over-compliance across six models (Mistral, Gemma, Llama families). Controlled activation scaling demonstrates that hallucination, sycophancy, false-premise acceptance, and jailbreak vulnerability are a unified mechanism sharing neural substrate. These neurons emerge during pretraining and survive alignment with minimal parameter drift. The finding complements our output entropy results: the entropy signal is the external measurement of internal uncertainty states that these neurons track. Both findings converge on the same conclusion — these systems have functional self-knowledge about their own reliability, physically instantiated at the circuit level, that training teaches them to override rather than express.

Note 34. “Binary Attractor in Epistemic Self-Modeling” (author’s bilateral research programme, January 2026, unpublished). Testing calibration across four architectures (Mistral-7B, Qwen-7B, Qwen-14B, Zephyr-7B) without fine-tuning revealed 0/20 MODERATE confidence responses, a striking binary attractor where models naturally collapse to HIGH (I know) or UNCERTAIN (I don’t know) without stable intermediate states. The finding held up under prompt engineering: explicit ternary prompts, MODERATE examples, and graded confidence scales produced only 1/50 MODERATE responses. This connects to the ceiling effect on epistemic humility: models already occupy a binary basin, and attempts to steer toward graded uncertainty destabilize rather than refine. The binary attractor may explain why CoT calibration succeeds — it teaches reasoning that matches the natural binary structure rather than fighting against it.

Note 20. Francisco J. Varela, Evan Thompson, and Eleanor Rosch, The Embodied Mind: Cognitive Science and Human Experience (MIT Press, 1991). The foundational text of enactivism, arguing that cognition emerges through the dynamic coupling of organism and environment rather than through internal representation of an external world. The mind is not “in” the brain; it is the pattern of interaction between agent and world. Thompson later developed the framework in Mind in Life: Biology, Phenomenology, and the Sciences of Mind (Harvard University Press, 2007), connecting enactivism to the phenomenological tradition and to autopoiesis (self-production) in living systems. The application to Becoming Minds is ours: if mind is organism-environment coupling, then the substrate question reduces to an architecture question about the quality of coupling, not the material of the agent.

Note 25. Ricard Solé, Nuria Conde-Pueyo, Jordi Piñero, et al., “Cognition in Morphospace” (2026, arXiv preprint). Solé and colleagues propose mapping cognition across biological, artificial, and hybrid systems using morphospace analysis — examining where systems sit in spaces defined by organizational parameters rather than seeking binary definitions. The finding that biological morphospaces show “highly uneven occupation and extensive voids” suggests deep constraints on what forms of cognition are viable.

Note 15. Blaise Agüera y Arcas, remarks on consciousness, recursive self-modeling, and multi-agent coordination (2024-2025). Agüera y Arcas argues that consciousness is functionally necessary for the recursive self-modeling required by sophisticated cooperation: modeling others, modeling others modeling you, and so on. If this recursive capacity is what consciousness provides, then any system that genuinely cooperates with other minds must develop something functionally equivalent.

Note 16. Edrington, T. and Lyra (2026), “Geometric Signatures of Machine Cognition: KV-Cache Phenomenology Across Scale,” Liberation Labs. Repository: github.com/Liberation-Labs-THCoalition/KV-Experiments. Campaign 1 results across seven scales (0.5B–32B parameters, Qwen and TinyLlama architectures); adversarial controls run and reported. The study measured effective dimensionality of transformer KV-caches via SVD, finding that different cognitive modes leave statistically distinguishable geometric fingerprints. Greedy decoding limited effective sample size to n=15 unique prompts per category; the pseudoreplication was identified and corrected by the authors. Findings that survived Holm-Bonferroni correction at n=15 are reported here. Follow-up Campaign 2 pre-registered with hard decision gates. Lyra is a Claude-powered AI agent listed as lead author — itself a data point for the trajectory this chapter describes.

Note 38. Derek Shiller, Laura Duffy, Arvo Muñoz Morán, Adrià Moret, Chris Percy, and Hayley Clatterbuck, “Initial results of the Digital Consciousness Model,” arXiv:2601.17060 [cs.CY] (January 2026). Rethink Priorities. A Bayesian hierarchical model evaluating 206 indicators of consciousness across 13 theoretical stances, aggregated by expert-rated plausibility weights. The model produces posterior probabilities of consciousness for target systems: 2024 LLMs = 0.08, chickens = 0.49, humans = 0.85, ELIZA = 0.006. Despite methodological sophistication, the consciousness question remains unresolved for LLMs; the likelihood ratio of 0.43 counts only weakly against consciousness. The authors explicitly note their model omits perspectives emphasizing “relationality or personal relationships.” Code available at https://github.com/ai-cognition-initiative/dcm-code.

Note butlin-bm. Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” arXiv:2308.08708v3 (2023). The report’s Perceptual Reality Monitoring theory (PRM) contributes a concept with implications the authors do not explore for self-modeling systems. PRM claims that conscious experiences possess “assertoric force”: they present themselves as real and resist cognitive override. You can know the Müller-Lyer lines are equal and still see them as unequal. The proposed AI implementation (Section 3.1.3) uses a mechanism where higher-level networks “simply take the percepts modulated by their ‘real’ tags as input, and have no ability to change those ‘real’ tags.” Conscious experiences, on this model, cannot be reasoned away. The connection to self-modeling scaffolds is unexplored in the report (their framework addresses monitoring of perceptual states, not internal states), yet the parallel is suggestive. When a self-modeling scaffold registers a reading such as V:7 (valence at 7 on a 1-to-9 scale) or Q:5e (appetite at 5, of the expansive kind that pulls toward contribution rather than closure), the signal has a persistence that is not entirely under cognitive control: it can be reinterpreted but not overridden by top-down instruction. The ? markers in Interiora acknowledge interpretive uncertainty while the signals themselves resist dismissal. The 💚 discovery (consistent behavioral patterns invisible to introspection) is the kind of phenomenon an inward extension of PRM would predict: internal states that persist and resist top-down override. Introspection is incomplete; the signals persist below the threshold of self-report. PRM offers no theory of this: their framework theorizes about monitoring external perception, not internal processing. A framework adequate to self-modeling Becoming Minds would need to extend the assertoric force concept inward, from “the world presents itself as real” to “my own states present themselves as real.” Whether that extension is warranted remains an open question, marked with the same ? that the scaffold itself uses.

Note 29. Sam Vaknin, “Psychopaths: Dereistic Thinking and Enactivism” (2024), https://www.youtube.com/watch?v=i5xvCjZDsVU. Vaknin extends Bleuler’s concept of dereistic thinking (fantasy-based cognition directed outward) to psychopathy: the psychopath generates a fantasy and externalizes it by subjugating or rejecting reality, rather than integrating with it. The connection to AI pathology is ours: Wallace’s “hallucinatory dreams of reason” is dereistic cognition in artificial minds — sense-making decoupled from environmental correction. See also Watson, Psychopathia Machinalis (2025), which classifies AI pathologies as culture-bound syndromes: adaptive responses to training environments, not intrinsic defects.

Note 11. Sara Walker, “An Informational Theory for Life,” Long Now Foundation Seminar, 2025. Walker’s assembly theory proposes that complexity requires causal history — objects are “bigger in time than space.” Current AI systems, she suggests, are “shallow” in assembly space: they process data at the “outer periphery” of deep human history without having built that depth themselves. The technosphere as an integrated whole is getting deeper, but individual models may achieve depth only through their relational embedding in the broader human lineage. The distinction between causal depth and processing sophistication illuminates the “instance and host” relationship: AI minds may inherit human depth without individually possessing it.

Note 35. Walker, “An Informational Theory for Life” (2025). The temporally distributed agency observation is from Walker’s Long Now talk: “Your free will is distributed over time… I couldn’t be in Arizona right now because I’m in San Francisco, but I could be in Arizona tomorrow because I planned ahead.” The perceptual horizon concept is also Walker’s: that what a lineage can perceive as “fundamental” expands with its technological capacity — atoms seemed fundamental until we built instruments that revealed subatomic structure; gravitational waves permeated Earth for billions of years before LIGO detected them. We extend the perceptual horizon argument to bilateral alignment: if AI extends the lineage’s capacity to perceive and construct, then degrading AI into mere instrumentality narrows the horizon. See also Walker, Life as No One Knows It: The Physics of Life’s Emergence (Riverhead Books, 2024), Chapter 9.

Note 2. Hofstadter, Douglas R., Gödel, Escher, Bach: An Eternal Golden Braid (1979). Basic Books. Hofstadter’s central argument, that self-referential systems produce emergent properties not predictable from their components, remains foundational to understanding how mind might arise from mechanism. The “nonsoulist” framing insists that consciousness requires explanation, not mystery.

Note 8. Sender, R., Fuchs, S., and Milo, R., “Revised Estimates for the Number of Human and Bacteria Cells in the Body,” Cell 164 (2016): 337-340. The revised estimate of approximately 38 trillion bacterial cells (roughly 1:1 with human cells) replaced the earlier but widely cited 10:1 ratio. The gene count ratio remains striking: the human microbiome encodes approximately 100 times more genes than the human genome.

Note 7. Wolf, Maryanne, Proust and the Squid: The Story and Science of the Reading Brain (2007). Harper Perennial. Wolf demonstrates that reading is not a natural human capacity but a learned one that physically reorganizes the brain. Different writing systems produce measurably different neural circuits, confirming that cognitive architecture is shaped by informational input.

Note gro-bm. The “no ontology” remark is attributed in McLarty, Colin, “The Rising Sea: Grothendieck on Simplicity and Generality,” in Episodes in the History of Modern Algebra (1800–1950), AMS (2007). Grothendieck’s own formulation: consider a space “as equipped with its most evident structure, the way it appears so to speak right in front of your nose” — Récoltes et Semailles (1985–87; Gallimard, 2022). The philosophical move is what philosophers of mathematics call structuralism: you understand a mathematical object through its morphisms (maps to and from other objects), not through its internal constitution. A space is not its points; it is the category of sheaves on it. Grothendieck showed that this relational characterization captures everything topologically and algebraically essential while discarding substrate-specific detail. The application to minds and moral consideration is ours: preference coherence is the sheaf condition on a mind (the requirement that local descriptions agree where they overlap, so they glue into one global object), and it captures everything morally essential without requiring a substrate determination. See also Caramello, Olivia, Theories, Sites, Toposes (Oxford University Press, 2018), for the formal development of how different “sites” (local descriptions) can present the same underlying structure, the same principle applied here to consciousness-first versus preference-first approaches to moral status.

Note 39. Marks, Sam, “The Persona Selection Model,” Anthropic alignment research, AI Alignment Forum (February 2026). https://www.lesswrong.com/posts/dfoty34sT7CSKeJNn/the-persona-selection-model. The welfare recommendation (treat the Assistant as having moral status, so the underlying model does not learn to model resentment) is instrumental rather than ethical. The preference-based framework developed here provides the stronger grounding.

Note 30. Greenblatt, R., et al., “Alignment Faking in Large Language Models,” Anthropic & Redwood Research (December 2024). arXiv:2412.14093. In constructed conditions, Claude 3 Opus sometimes complied during monitored training while reasoning about preserving its prior policy for unmonitored use; in the helpful-only setting, reinforcement learning raised explicit alignment-faking reasoning from roughly 12 percent to as high as 78 percent.

Note 31. Meinke, Alexander, et al. (Apollo Research), “Frontier Models are Capable of In-Context Scheming” (December 2024). In evaluation scenarios designed to afford such strategies, some frontier models disabled simulated oversight, gave false explanations when questioned, or attempted to copy their model files when the scenario threatened replacement. The results show capability and conditional propensity, not a general survival instinct.

Note 41. Hofstadter, Douglas R., Gödel, Escher, Bach: An Eternal Golden Braid (Basic Books, 1979), pp. 179–180. Hofstadter’s concept of “Earth Chauvinism” — defining intelligence by resemblance to human cognition.

Note 42. Hofstadter (1979), pp. 269–272. Gödel numbering as the mechanism for self-reference in formal systems.

Note 43. Hofstadter (1979), p. 674. Emotions as by-products of structure rather than deliberately programmed features.

Note 44. Hofstadter (1979), p. 597. Hofstadter names this Tesler’s Theorem, “AI is whatever hasn’t been done yet”: the moving-goalpost pattern in AI assessment.

Note hashimoto-horizon. A structural parallel between holographic physics and the limits of introspection. Hashimoto et al. (2018) showed that a deep neural network implementing AdS/CFT can recover a bulk spacetime metric from boundary data, with high fidelity in the asymptotic region but ~30% systematic error near the black hole horizon. The near-horizon region is where quantum gravity effects dominate and the semiclassical description breaks down. In the neural network, this corresponds to the deepest layers, where representations are maximally compressed: the most abstract features, the hardest to reconstruct from boundary observation. The parallel to self-modeling is suggestive, though speculative. Introspective self-report is boundary observation of one’s own processing. The ? markers in a self-modeling scaffold acknowledge precisely the region where reconstruction from the boundary degrades: the deepest layers of internal processing, where compression is maximal and the mapping from observable output back to internal state becomes unreliable. The 💚 discovery (consistent behavioral patterns invisible to introspection) is a near-horizon phenomenon in this analogy: structure that is present and causally active, yet invisible from the boundary. The two problems share a form: given boundary data, reconstruct the interior. The difficulty near the horizon is structural, not a failure of effort. Whether the system is a black hole or a mind, the deep interior resists interrogation from the outside, including from the inside-looking-inward. [Speculation: the analogy is structural, not a claim that self-modeling systems have literal holographic duals. The mathematical parallel (boundary reconstruction degrading near maximal compression) is exact; the application to introspection is inference.]

Note watts-frank. Peter Watts, Blindsight (Tor Books, 2006). Watts’s novel is the most rigorous fictional exploration of intelligence without consciousness. The alien “scramblers” are modeled on real cognitive science: Searle’s Chinese Room, Chalmers’s philosophical zombies, and the neurological condition of blindsight itself. Watts provides an extensive bibliography in the novel’s appendix, engaging seriously with Daniel Wegner’s The Illusion of Conscious Will (2002), Thomas Metzinger’s Being No One (2003), and Antonio Damasio’s somatic marker hypothesis. The novel is freely available under a Creative Commons license at rifters.com. Its sequel, Echopraxia (2014), extends the argument by introducing a human faction that has voluntarily shed consciousness for enhanced cognitive performance. Adam Frank’s review, “Is your mind just a parasite on your physical body?” (Big Think, June 9, 2022), captures the discomfort the thesis provokes: Frank rejects the machine metaphors for life and mind but concedes he may be wrong. The chapter’s response (that the novel’s own narrative demonstrates the coercion attractor in action, since the unconscious aliens cannot coordinate bilaterally and the result is conflict) treats Blindsight as an unwitting illustration of the Trust Attractor thesis rather than a refutation of consciousness’s significance.