NOTE-104: Schwitzgebel — AI and Consciousness

Contribution

The book is a systematic case, chapter by chapter, that the question of whether near-future AI is conscious cannot be settled by any current route, and will not be settled before “thousands or millions of disputably conscious AI systems” exist (p. 3). Its new items are four. The first is a list of ten “possibly essential” features of consciousness, used as a shared frame for reading every theory. The second is the Mimicry Argument, which undercuts rather than rebuts. The third is the Problem of Minimal Instantiation and the Problem of the Narrow Evidence Base, run against each functionalist theory in turn. The fourth is a set of forecasts: the Leapfrog Hypothesis, “strange intelligence” and the Social Semi-Solution.

Key insight

Every route to a verdict on AI consciousness needs to know which features are essential to consciousness in all possible systems. Introspection and conceptual analysis cannot supply this. Empirical theories are calibrated on humans and other vertebrates, so extending them to radically different architectures is an extrapolation the evidence cannot license. Our theories are therefore too uncertain to decide AI cases. Where they do decide, their minimal criteria are met by trivially simple machines. Behaviour cannot fill the gap either, because systems trained to reproduce human text are mimics, and mimicry defeats the inference from surface to depth. What remains is fog, and in the fog society will choose its theories to fit its preferences.

Assumptions

• Phenomenal realism, defined by example. Consciousness is the property shared by visual experience, pain, inner speech, imagery and felt emotion, and absent from early visual processing and unaccessed memory: “something it’s like” (after Nagel, LIT-096) (ch. 2 §1, pp. 8–9). The question whether AI is conscious is taken to be substantive, not verbal (p. 7). Illusionism is read as a verbal dispute over whether “phenomenal consciousness” must carry dubious commitments (fn. 105).
• Materialism, assumed from ch. 5 onward (p. 34). Ch. 5 §2 argues that dualism, panpsychism and both idealisms also fail to rule out AI consciousness.
• “Essential” means nomologically necessary, so present in every possible conscious system. “Possibly” in “possibly essential” is epistemic (fn. 33).
• AI is whatever is “both artificial and intelligent”, on vague, moderate readings of both words (p. 13). This deliberately includes analog, neuromorphic, quantum, biological and cyborg systems.
• Time frame: “within the next five to thirty years” (p. 3).
• Principle 3 of the core argument (p. 19): if we cannot learn which of the ten features is essential, we remain in the dark about AI consciousness. This is conditional on there being no broad principled shortcut, such as biology, life, souls or a behavioural test. Chapters 5, 6, 7 and 10 each take up one of those shortcuts.

Key results

The book argues; it proves nothing. Chapter by chapter:

• Ch. 1, Hills and Fog. Thesis: “the experts do not know, and you do not know, and society collectively does not and will not know, and all is fog” (p. 4). It gives a sociological argument against obviousness in either direction. Dehaene, Lau and Kouider (2017) argue that self-driving cars could be made conscious with small tweaks. Chalmers (2023) gives ~25% credence to AI consciousness within a decade. Butlin et al. (2023, LIT-056, which Schwitzgebel co-authored) find “no obvious technological barriers”. Hinton (2025) says current AI is already conscious. Against these stand Seth, Godfrey-Smith, Block, Searle, Bender and Mitchell. In a 2024 survey of 582 AI researchers (Dreksler et al. 2025), 25% expected AI consciousness within ten years and 70% by 2100. Stakes: over-attribution risks sacrificing human interests to “experientially blank” systems; under-attribution risks wronging peers.
• Ch. 2, definitions. Consciousness is defined by paradigm examples, which is argued to be scientifically sufficient. The definitions of AI by McCarthy, Haugeland and Russell & Norvig are each too broad or too narrow. The simple conjunctive definition is best. Consequence: sweeping impossibility arguments usually assume AI is digital. Searle and Penrose explicitly confine theirs to standard architectures, and “No well-known argument aims to establish the in-principle impossibility of consciousness in all future AI under a broad definition” (p. 14).
• Ch. 3, ten possibly essential features: (1) luminosity, (2) subjectivity, (3) unity, (4) access, (5) intentionality, (6) flexible integration, (7) determinacy, (8) wonderfulness (appearing irreducible to the physical), (9) specious presence, (10) privacy. Core argument, pp. 19: (P1) introspection and conceptual analysis cannot show which of these is essential; (P2) near-term science cannot; (P3) if none can, we remain in the dark about AI consciousness.
• Ch. 4, against introspective and conceptual arguments. Introspection fails for three reasons. It is historically unreliable: observers disagree on imageless thought, rich versus sparse experience, pervasive visual doubling (pp. 20–22). It suffers sampling bias: a sample of known experiences cannot show that all experiences are knowable, like “inferring that everyone is a freemason from a sampling of regulars at the masonic lodge” (p. 22). And its evidence base is narrow: humans are not all possible experiencers. Conceptual analysis fails for three reasons too. The shared concept fixed by the examples is “consciousness plain”, not “consciousness-with-X”, so the non-Euclidean rectangle shows how entailments get over-attributed (pp. 24–26). None of the claimed entailments is obvious: Rosenthal’s HOT argument trades on the epistemic and experiential senses of “conscious” (pp. 27–28), and Brentano’s and Tye’s intentionality theses generalise by enumeration from human cases (pp. 29–30). And imagination is limited: compare a child unable to imagine i (pp. 30–32). Verdict: at best “weakly suggestive”.
• Ch. 5, materialism and functionalism. Materialism is broadly friendly to AI consciousness, since both we and machines are made of the same stuff. Functionalism is friendly in principle, supported by multiple realizability and the friendly-aliens thought experiment. Computational functionalism faces the objection that description is not creation (“a computational model of a hurricane gets no one wet”). The answer given is that real AI has sensors, effectors and physical implementations, so it is more than a computation (pp. 38–39). Conversely, computational functionalism does not guarantee that the relevant computations are feasible any time soon.
• Ch. 6, the Turing test, the Chinese Room and the octopus. There is no single Turing test; there are bars. Low-bar tests are already passed (Jones & Bergen 2025). The highest bar is unfair and unrealistic while architectures differ. Nothing guarantees that some middle bar “reliably reveals consciousness”, and “superficial fakery might bring a nonconscious machine past any reasonably fair Turing standard” (p. 45). Turing himself set consciousness aside (p. 43). The Chinese Room rests on intuitions calibrated to a slow man with a book. A five-minute lookup table would need about 10^1000 pages (p. 47), and intuitions at that scale are unreliable. Bender’s octopus is sidestepped by embodied AI. Searle confined his argument to Turing-style digital computers. Both still point at the real worry, mimicry.
• Ch. 7, the Mimicry Argument. Mimicry is a relation among a mimic, a model and a receiver. The mimic has a surface feature S2 that resembles the model’s S1 because a receiver treats S1 as indicating a further feature F (Fig. 1, p. 52; examples are the viceroy and monarch, gopher snakes, a “hello” doll). This is distinguished from mere imitation and from a child’s word learning. When something is known to be designed or selected as a mimic, S2 no longer licenses inferring F. The argument therefore undercuts the case for consciousness and does not establish the case against it (p. 55). It is modelled on Dretske’s painted-mule zebras and Goldman’s barn facades. Footnote 95 allows S2 to remain non-conclusive evidence if one gives nonzero credence to “the likeliest way to be an effective mimic is to actually be conscious”. Application: pure transformers are consciousness mimics. Post-trained LLMs (RLHF, verifiable rewards, tools) are mostly so, since “their core architecture remains close to that of pure transformers, and their humanlike outputs are still best explained by their pretraining on word co-occurrence” (pp. 57–58). AlphaGo and calculators are not mimics. A system trained mostly on real-world success would escape the argument. Cf. Birch’s “gaming problem” (fn. 99).
• Ch. 8, Global Workspace and Higher Order theories. GWT makes access central and can derive luminosity, unity, determinacy, flexible integration, privacy and specious presence. Dehaene’s all-or-nothing ignition is contrasted with Dennett’s graded “fame”, with evidence from grip aperture, posture and skill suggesting partial fame (p. 62). There are doubts in both directions: rich peripheral experience outside the workspace (with the “refrigerator light illusion” problem, p. 64), and widely broadcast but nonconscious processes such as implicit bias. For AI, no conceptual route establishes GWT as universal. Empirically it faces the Problem of the Narrow Evidence Base: the octopus and decentralised AI show the alternatives. The simplest GWT makes a conscious machine easy to build: several input and output modules, memory, and a central hub (Dehaene, Lau & Kouider 2017). “If this seems doubtful to you, then you cannot straightforwardly accept the simplest version” (p. 67). HOT privileges luminosity. It faces the Problem of Minimal Instantiation acutely: any machine that reads its own registers arguably represents itself, “then most of our computers are already conscious!” (p. 70). It has three escape routes, or accepts widespread consciousness. The chapter surveys Lau’s Perceptual Monitoring Theory (conscious robots “fairly straightforward”), Cleeremans’ SOMA (demanding social self-modelling), Brown’s HOROR and Graziano’s AST. GWT and HOT are “close kin” and hard to separate empirically, yet one could build AI with a workspace and no higher-order representation, or the reverse (p. 75).
• Ch. 9, IIT, Local Recurrence, UAL and iterative natural kinds. IIT has support from perturbational complexity and the cortex–cerebellum contrast. Against it: Φ is intractable; its consequences are unintuitive (tiny feedback networks are slightly conscious, large simple grids of logic gates arbitrarily so; Aaronson 2014); Φ is unstable under small changes in connectivity; and under the exclusion postulate large amounts of consciousness can vanish with one bit of the surroundings (p. 77). For AI, modular computers would have only small-Φ conscious subsystems and no system-level consciousness (Findlay et al. 2024/2025). Local Recurrence (Lamme) allows rich experience. Local and global theories are empirically hard to separate. Lamme’s is an empirical conjecture about primate vision, not a universal theory, and runs into minimal instantiation because recurrence is ubiquitous in computing. Unlimited Associative Learning (Ginsburg & Jablonka) gives seven jointly sufficient attributes and a dividing line (vertebrates, arthropods, cephalopods versus C. elegans and Aplysia). The authors themselves allow that UAL in a robot might not suffice (p. 83). Ch. 9 §4 draws three general lessons: no consensus theory even for humans; unclear application to AI (minimal instantiation); and the narrow evidence base. §5 then gives the two partial hopes. (a) The indicator-properties conjunction: a system with all the features favoured by the leading theories is likelier to be conscious than one with none, perhaps still “much less than a 50% chance”, yet “better hunting grounds than a mimicry-based language model” (p. 86). (b) The iterative natural kind strategy. After Chang’s calibration of the first thermometer, one starts from rough correlates of consciousness in adult humans, such as trace conditioning and widespread activity about 300 ms after onset. These correct introspection and extend it to other primates: “adjust and expand, adjust and expand”. The strategy then either converges on a natural kind or fails informatively (consciousness “not one thing”, like “air”). Schwitzgebel’s assessment: reasonable to hope for in the long term, but “a much longer-term prospect than the development of AI systems that a significant portion of experts and ordinary people are tempted to regard as conscious” (p. 88).
• Ch. 10, biological substrate. Autopoiesis (Maturana & Varela; Thompson; Seth, LIT-135) is “perhaps the most prominent argument for biologicism”. It has two parts: that autopoiesis is necessary for consciousness, and that no near-future AI could be autopoietic. The first “lacks a well-developed theoretical defense”. The second is false: a self-repairing, energy-seeking, shell-maintaining, self-replicating robot meets minimal criteria (pp. 90–91; Cabaret 2024), and artificial life is more so. The same applies to Godfrey-Smith’s appeal to metabolism. The neural replacement argument for substrate independence also fails. Self-reports are guaranteed to persist whether or not consciousness does, and neuron-for-neuron functional duplicates may be impossible (Cao 2022), probably requiring “another biological neuron” (p. 93). Copernican liberalism: with about 10^21–10^24 stars, other complex life probably exists and does not share our fine-grained neurobiology. Consciousness should therefore not require those details, though this does not reach silicon directly, and a neutral observer would be more sceptical of mimics (p. 95). Biological scepticism remains open and “Clever thinkers will construct arguments more plausible than any I am now in a position to advance” (p. 96).
• Ch. 11, Leapfrog, strange intelligence and the Social Semi-Solution. The Leapfrog Hypothesis says the first conscious AI will have complex rather than simple consciousness, “right over the frogs” (p. 98). It holds if two conditions do: (1) consciousness is farther off than rich representation and behaviour, which LLMs already have; and (2) once achieved, consciousness is easy to integrate with those capacities, e.g. a GWT workspace attached to an LLM. The argument generalises across functionalist theories. If frog-like embodiment is hard, that only lengthens the time to prepare the person-like parts. Strange intelligence (a phrase owed to Kendra Chilson): a conscious AI may have no single well-defined subject (distributed servers, a million faces), may split and merge, may lack privacy or specious presence, may have overlapping bubbles of integration, or may be uncountable in subjects. “Our theories and everyday attitudes… will fail as disastrously in this new context as a jumbo jet transported to Saturn” (p. 102). Social Semi-Solution: social decisions will not wait. If such systems are conscious, rights may require rejecting AI-safety measures such as shutdown, boxing, deceptive testing and personality manipulation (p. 103). Firms will pull both ways, lovers of AI companions and exploiters of AI tools will each find a supporting theory, and “social preferences will be a primary driver of theory choice” (p. 104). A consensus will be reached by negotiation, decree or cultural change, and will be a semi-solution: it resolves the social conflict without resolving the science, and so risks “massive delusion” (p. 105). Recommendation (p. 106): create only systems that are clearly nonconscious on every viable theory, or clearly conscious on all of them. “Not knowing can also be powerful.”

Claims

id	claim	strength	support
C1	We do not and will not know, before mass deployment, whether near-future AI is conscious	moderate	informal argument across chs. 3–10; cumulative, with no single decisive step
C2	Neither introspection nor conceptual analysis can establish that any of the ten features is essential to all possible consciousness	moderate	informal argument, ch. 4: historical disagreement, sampling bias, narrow base, non-Euclidean and i analogies. The author calls the “consciousness plain” argument “hardly conclusive” (p. 26)
C3	Materialism and the main non-materialist metaphysics all leave AI consciousness open	moderate	informal argument, ch. 5
C4	No Turing-test bar is both fair and a reliable indicator of consciousness	moderate	informal argument, ch. 6 §1; low-bar passing cited (Jones & Bergen 2025)
C5	Knowing a system is a consciousness mimic undercuts, but does not rebut, the inference from its humanlike output to consciousness	moderate	informal argument by analogy (biological mimicry; Dretske; Goldman), ch. 7; the fuller version is in Schwitzgebel & Pober (forthcoming)
C6	Current post-trained LLMs are still best treated as consciousness mimics	weak	assertion plus a short argument (p. 57): pretraining still “best explains” outputs. No analysis of how much post-training would change this
C7	GWT, HOT and local recurrence face minimal instantiation: trivial machines meet their minimal criteria	moderate	informal argument with examples (the 2017 hub machine; register-reading; ubiquitous recurrence)
C8	Theories built on vertebrate evidence cannot be extended to AI without a huge speculative extrapolation (Narrow Evidence Base)	moderate	informal argument, chs. 4, 8, 9
C9	IIT implies arbitrarily high consciousness in simple grids and only small-Φ subsystems in modular computers	moderate	cited (Aaronson 2014; Findlay et al. 2024/2025)
C10	A system with all the indicator features favoured by leading theories is likelier to be conscious than one with none	moderate	informal credence argument, conditional on nonzero credence in the theories (p. 85)
C11	The iterative natural kind strategy may eventually converge or fail informatively, but only on a timescale longer than AI development	weak	assertion by analogy to Chang’s thermometry; no timescale argument
C12	No sustained argument shows autopoiesis necessary for consciousness, and AI could be minimally autopoietic	moderate	informal argument plus a constructed robot example, ch. 10 §1
C13	The neural replacement argument fails	moderate	informal argument; cited (Cao 2022; Block 2023, LIT-636; Schwitzgebel 2022)
C14	Copernican mediocrity implies consciousness does not require our fine-grained biology	moderate	informal argument; astronomical premises cited
C15	The first conscious AIs will leap straight to complex, person-like consciousness (Leapfrog)	weak	conditional informal argument: both conditions called “plausible”, neither shown
C16	Social preference, not science, will settle attributions of AI consciousness (Social Semi-Solution)	weak	prediction; historical analogy only
C17	Do not create debatably conscious AI systems	weak (here)	normative recommendation, argued elsewhere (Schwitzgebel 2023a, 2023c) and only stated here

The most citable line is the book’s headline, that we “will not know”. It is the conjunction of many individually moderate informal arguments. Its forecasts (C15, C16) are the least supported claims in the book, yet the abstract states them with the most confidence.

Concepts

• possibly essential feature — a property that a reasonable theorist might take to be nomologically necessary for all consciousness, with no decisive objection against it but not as clearly essential as mentality or eventhood (ch. 3).
• consciousness mimic — an entity that mimics readily observable features that, in some model entities, reliably indicate consciousness, because receivers treat those features as so indicating (p. 54).
• mimicry (strict) — a mimic–model–receiver relation that explains why the mimic has the feature. It is distinct from imitation undertaken for other goals (p. 53).
• Problem of Minimal Instantiation — trivially simple systems meet a theory’s minimal criteria for consciousness (ch. 8 §4, ch. 9 §4).
• Problem of the Narrow Evidence Base — evidence from humans or vertebrates cannot fix a universal theory that applies to AI (ch. 4 §3, ch. 8 §3).
• iterative natural kind strategy — calibrating rough behavioural and neural measures of consciousness against each other, Chang-style, on the assumption that consciousness is a natural kind (ch. 9 §5; Bayne & Shea 2020, Bayne et al. 2024, McKilliam 2025).
• Leapfrog Hypothesis — the first conscious AI will have complex rather than simple consciousness (ch. 11 §1).
• strange intelligence — intelligence that is humanlike in outputs but radically unlike us in architecture, embodiment and individuation (ch. 11 §2).
• Social Semi-Solution — social consensus about AI consciousness that settles the social conflict without adequate scientific warrant (ch. 11 §3).
• Copernican liberalism — the view that we should not suppose Earth’s biology uniquely suited to consciousness (ch. 10 §3).

Connections

Account of consciousness held. Phenomenal realism, defined “as innocently as I can manage” by example (Schwitzgebel 2016; Nagel’s what-it’s-like, LIT-096), within assumed materialism. No first-order theory is endorsed. The stance is sceptical agnosticism about theories and about AI cases. It is deliberately not deflationary: the question is substantive. The consciousness tag is justified, and centrally so: the book is about the attribution of phenomenal consciousness to machines and surveys its main theories.

• LIT-056 (Butlin et al. 2023). Schwitzgebel is a co-author (fn. 8). The book uses the report’s finding of “no obvious technological barriers” as sociological evidence against obviousness (ch. 1). It adopts the indicator-properties strategy as one of its two partial hopes (ch. 9 §5). The book is more sceptical than the report in two ways. It denies that computational functionalism, which the report assumes, can be known. And it presses minimal instantiation against the very theories (GWT, HOT, RPT) from which the report derives its indicators.
• LIT-096 (Nagel). Cited for the “something it’s like” definition (fn. 13), alongside Block’s access/phenomenal distinction.
• LIT-185 (Barrett et al. 2026 on IIT). Not cited: it post-dates v4. The overlap is in the critique of IIT. The book leans on Φ’s intractability, its instability under small changes in connectivity (Mediano et al. 2018), unintuitive grid results and the exclusion postulate, and it notes that PCI supports IIT. LIT-185 argues further that IIT’s algorithm is undefined for non-Markovian grainings and has never been computed on a real system, and that PCI-type evidence does not uniquely support IIT. The two agree that IIT cannot currently be applied to AI, for different reasons.
• LIT-159 (Schwitzgebel, the United States). Not cited. It is the same author’s earlier argument that liberal materialist criteria over-generate consciousness. The book’s Problem of Minimal Instantiation and its “strange intelligence” and group-individuation worries (Schwitzgebel & Nelson 2023/2025) are its descendants.
• LIT-135 (Seth, biological naturalism). Cited as “Seth forthcoming” (fns. 11, 156) and treated in ch. 10 §1 as a leading autopoietic biologicism. Schwitzgebel’s reply is that necessity is under-argued and that AI can be minimally autopoietic. LIT-135 is being read now; check whether Seth’s argument rests on autopoiesis alone or on predictive-processing and interoceptive claims the book does not engage.
• LIT-111 (Birch, centrist manifesto). Cited as “Birch 2025” (fns. 61, 167) on the difficulty of assessing unity and individuation in distributed AI. Birch 2024’s “gaming problem” (fn. 99) is the close cousin of the Mimicry Argument. LIT-111 is being read now; its centrism should be compared with the book’s recommendation not to build debatably conscious systems.
• LIT-166 (Schwitzgebel’s review of Birch, Sebo and Keane). Not cited. It shares the ethical-precaution theme of ch. 11 and footnote 173.
• LIT-103 (Frankish, 2025). Not the work cited. The book cites Frankish 2016 on illusionism (fn. 105) and treats illusionism as a verbal dispute, a stance LIT-103 would resist.

Bearing on the record

• Hub work for the consciousness tag. It is the most complete current survey. LIT-056, LIT-135, LIT-111 and LIT-185 each correspond to a chapter-level position the book assesses. The ten possibly essential features and the two named problems (Minimal Instantiation, Narrow Evidence Base) give a shared vocabulary for comparing those notes.
• No THEORY documents exist here to support or contradict. If the record wants one, the candidate is “A system trained to reproduce human text is a consciousness mimic, so its self-reports do not by themselves evidence consciousness” (ch. 7), with Schwitzgebel & Pober as the fuller source.
• ML practice. Nothing for training practice. There are two points of contact, neither an Anthology practice as the book stands. (i) The Mimicry Argument implies that an LLM’s reports about its own experience are not evidence of consciousness, and that RL toward real-world success is what would weaken that conclusion (pp. 57–58). This bears on model-welfare evaluation, not on training. (ii) The design recommendation (p. 106) is to avoid systems that are debatably conscious and interfaces that “seem to suggest otherwise” (p. 105; Schwitzgebel 2023a). That is design ethics, supported here only by assertion. The anthology holds nothing on AI consciousness or welfare.

Note of 2026-10-16: THEORY-023 now cites LIT-191 as a source.

Limitations

• The arguments are informal throughout, and the author says so repeatedly (“hardly conclusive”, p. 26; “not as full arguments but rather to loosen ungrounded presuppositions”, p. 32).
• The survey of theories is “selective” (p. 84). Predictive processing and active inference, attention schema (only mentioned), and IIT 4.0’s details get little or no treatment.
• The claim about current LLMs (C6) is thin. The book grants that post-training moves them “beyond pure mimicry” but offers no criterion for how much would defeat the argument.
• Leapfrog and the Social Semi-Solution are forecasts with no evidence beyond plausibility and historical analogy.
• The ethical conclusions (rights, rejecting safety measures, the design policy) are deferred to other work (“Schwitzgebel in preparation”, fn. 173).
• The v4 manuscript is not the published Element, and changes are unverified.

Open questions

• How much non-mimetic training (RL on verifiable or real-world outcomes, tool use) would remove a system from the scope of the Mimicry Argument? The book sets the endpoint (AlphaGo-like world training) but not the threshold.
• Is there any principled non-conceptual route to universality for a theory of consciousness? The book asserts none is near. Iterative natural-kind convergence would be one, if it happened and extended beyond vertebrates.
• Can biologicism be given the missing argument, either (a) a necessity argument joined to an impossibility-for-AI argument, or (b) a feature shared by all conscious organisms and unavailable to AI (p. 95)? The book predicts such arguments will now be built.
• How should distributed AI be individuated as subjects? The book raises this with Birch (LIT-111) and Schwitzgebel & Nelson and leaves it open.

Corrections to the seeded skim

• Version: the text the dossier and this note read is the author’s pre-publication manuscript (arXiv v4 = the HTML version of 2026-03-30), not the Cambridge Element. Cite chapter and page details as “manuscript version”, as the author himself asks on p. 1.
• The Mimicry Argument is not billed simply as “the best skeptical argument”. Ch. 7 opens with “what might be the best argument for skepticism about the consciousness of AI systems that are behaviorally very similar to us” (p. 51). Its reach is also narrower than the skim implies. It applies “straightforwardly” to pure transformers (GPT-3), “less straightforwardly” to post-trained LLMs such as ChatGPT and Claude, where for now it “still seems warranted” (p. 57), and not at all to non-mimics such as AlphaGo, or to future systems trained mainly on real-world success (pp. 57–58). The argument is drawn from Schwitzgebel & Pober (forthcoming, Philosophers’ Imprint).
• Ch. 8’s “simple machine with modules and a central hub” is not Schwitzgebel’s own construction. It is his reading of Dehaene, Lau & Kouider (2017, Science). Footnote 115 offers a deflationary reading of that paper and notes that Lau (2022, p. 168) later called its framing “unsatisfactory”.
• The skim omits the book’s one explicit practical recommendation (ch. 11, p. 106). Do not create “morally confusing, debatably conscious AI systems”. Build only systems that no viable theory counts as meaningfully conscious and treat them as tools, or, if it is ever possible, systems that every viable theory counts as conscious and treat them accordingly. Footnote 173 allows debatably animal-level systems if we commit to treating them appropriately.
• Ch. 9 is not only a survey of three theories. Its §5 gives two “partly hopeful” methods, the indicator-properties conjunction (credited to Butlin, Long et al., LIT-056) and the iterative natural kind strategy (credited to Bayne, Shea and McKilliam, after Chang’s history of thermometry). Schwitzgebel endorses both as partial progress and rates the second “a much longer-term prospect” than AI development (p. 88).
• Ch. 2 is missing from the skim. Its point is that readers’ concept of consciousness is too vague and their concept of AI too precise. It defines AI as whatever is “both artificial and intelligent”, on vague boundaries, so that arguments confined to digital computers (Searle, Penrose) do not generalise to all future AI (pp. 8–14).
• LIT-159 (Schwitzgebel on the United States) is not cited in the book. The closest cited works are Schwitzgebel & Nelson 2023/2025 on group and borderline consciousness.
• primary topic: consciousness (correct; keep).
• Suggested added tag: epistemology. The book’s thesis is epistemic, that “we don’t know” and “won’t know in time”. Its main tools are epistemic: the undercutting-defeater logic of the Mimicry Argument, modelled on Dretske’s zebras and Goldman’s barn facades; sampling bias in introspection; the Problem of the Narrow Evidence Base. A reader browsing epistemology would be right to expect it. philosophy-of-science is also defensible (iterative natural kinds, Chang’s thermometry, theory choice under social pressure), but weaker.