NOTE-183: Birch — AI Consciousness: A Centrist Manifesto

Contribution

Birch sets out a “centrist” position that treats two opposed failure modes as simultaneous obligations. He names the obstacles to each and gives each a tractable next step. For misattribution, he gives a personal-identity argument that the interlocutor users bond with is, in the central case, a character with no persisting bearer, plus design nudges that avoid forced denials. For possible real AI consciousness, he diagnoses why neither behavioural nor architectural evidence can currently move the debate (gaming, Janus), and proposes comparative animal-consciousness science as the long-run way to adjudicate between computational functionalism and biological naturalism.

Key insight

Deflating the chatbot “friend” and dismissing AI consciousness are different moves. The flicker and shoggoth hypotheses have “priced in” the persisting interlocutor illusion (p. 16), so exposing the illusion says nothing about them. Conversely, the reason the debate cannot move is not only that we cannot detect consciousness in AI. The two camps also disagree about what would count as evidence (the “two-level deadlock”). The way out goes through animals, where there is no gaming, and through asking whether biological properties predict the distribution of consciousness markers beyond what their computational effects explain.

Assumptions

• Phenomenal consciousness in Nagel’s sense is the target. “Something it’s like” is treated as “unstable common ground” (§1).
• Parfit-style psychological-continuity theories of personal identity: persistence needs Relation R, including causal connection across gaps, “not simply a matter of matching content” (pp. 6–8).
• The serving facts are taken as the central case: consecutive turns may run on different implementations (MoE, multiple data centres). Birch himself calls this a presumption (fn. 7).
• Anti-behaviourism: all of the behavioural dispositions humans treat as evidence of consciousness can exist with no experience (p. 18).
• LLMs have both the opportunity to learn the markers of consciousness and the incentive (user satisfaction, interaction time) to reproduce them (p. 18).
• Evidence for or against consciousness attributions to living animals is possible (p. 23, fn. 25).
• Inference to the best explanation over a research programme’s run of successes and failures is legitimate (pp. 24–25).
• The nudges are Thaler–Sunstein nudges: non-coercive interventions on implicit processing (fn. 16).

Key results

• Persisting interlocutor illusion (§4). The only continuity between turns is the auto-appended history plus “abstract mathematical similarities in the model”. Even in the best case for continuity (states held in the KV cache within a response and recreated at the next one) there is no causal connection across the gap except through the history. So “there is no plausible theory of personal identity that could support the idea that these illusory beings typically exist” (p. 8).
• Illusion → attribution (§5). The illusion itself will persist, like perceptual animacy in Heider–Simmel displays, but the inference from the percept to a considered judgement “is blockable” (p. 9). Co-remembering, more than faces or voices, is predicted to be the strongest driver of consciousness attribution (p. 11). This is a prediction, not a result.
• Harm (§6). Birch argues by analogy with opioids and puts the burden of proof on companies (p. 10).
• Design (§7). Two pitfalls: “brainwashing” (forced self-denials, which fuel conspiracy theories) and “lobotomizing” (removing the capacity to build relationships). Most obvious fixes (uncanniness, no face or voice, no first person, limited memory or usage) push users to rival products. Three proposals: mandatory user training delivered by the chatbot before extended interaction, periodic trait reviews, and stepping-out-of-character moments. Nudges must be mandated by government (p. 13).
• Challenge Two (§§8–9). Role-play does not rule out a conscious “actor”. The flicker hypothesis posits experience per forward pass or token, whose felt duration need not match its objective duration. The shoggoth hypothesis posits a persisting subject per physical implementation, mapped many-to-many onto characters. There is “no such thing as proof by wackiness” (p. 16).
• Gaming problem (§10). The AI–animal parity principle is false. Schneider & Turner’s philosophy-of-mind test is undercut by the training corpus. The stipulated-pain trade-off results of Keeling et al. (2024) are best explained as mimicry. Flawless marker profiles might even count against consciousness (“consciousness-washing”, p. 18).
• Janus problem (§11). A global workspace found in an LLM would be taken by some as evidence of consciousness and by others as evidence that global workspace theory is insufficient. The deadlock is therefore at two levels (p. 22).
• Escape (§12). Biological-naturalist hypotheses are tested by their fit with the distribution of consciousness markers across taxa. Humphrey’s endothermy hypothesis already looks implausible given octopuses. Ephaptic coupling predicts fewer markers in heavily myelinated cetaceans and elephants, which also looks implausible to Birch. A long run of such failures would favour computational functionalism. A “brute necessary condition” such as bare metabolism may be untestable but is scientifically unreasonable (pp. 23–25). Architectural markers that are transient per character would suggest flickers, and stable ones would suggest shoggoths (p. 26, Fig. 4).
• Common ground (§13). “All sides should agree” that any AI consciousness would be profoundly alien, and that no research could “vindicate the reality of anyone’s human-like friend or partner” (p. 28).

Claims

id	claim	strength	support
C1	In the central case, no persisting entity bears the chatbot “interlocutor”; personal identity across turns fails	moderate	informal argument from Parfit’s Relation R, with GP-transcript and no-scan teletransporter cases (§4); the scattered-implementation premise is an acknowledged presumption (fn. 7)
C2	The persisting interlocutor illusion drives users’ attributions of consciousness	weak	”I strongly suspect” (p. 8); survey data that more than half of users grant some chance of consciousness (fn. 6); anecdote from emails to the author (§3)
C3	Co-remembering is a stronger driver of attribution than faces or voices	weak	prediction, no data (p. 11)
C4	Forced self-denials of consciousness backfire by fostering conspiracy theories	weak	assertion from the post-Lemoine history (“This did not work”, p. 11)
C5	Mandatory training, trait reviews and stepping out of character can block the inference without loss of functionality	weak	proposals, “very tentative” (p. 13); untested
C6	Behavioural markers of consciousness lose their evidential value for LLMs (gaming)	moderate	argument from three stated assumptions (p. 18); Keeling et al. (2024) as illustration
C7	Evidence of a global workspace in an LLM would not break the deadlock (Janus problem)	moderate	informal argument about background-belief dependence (p. 22)
C8	Biological-naturalist hypotheses are testable through comparative animal studies, and repeated failures would favour computational functionalism	moderate	informal argument with two worked examples (endothermy, ephaptic coupling); no data beyond Birch’s own plausibility judgements (“looks very implausible”, p. 24)
C9	A “brute” metabolism-necessity claim would be untestable and scientifically unreasonable	moderate	philosophy-of-science argument (p. 25)
C10	Any AI consciousness would be profoundly alien, and could never vindicate a human-like AI partner	moderate	follows from C1 together with the flicker/shoggoth structure; asserted as common ground (p. 28)

Concepts

• Persisting interlocutor illusion: the felt presence of one companion throughout a conversation, when each step is a separate processing event, possibly on a different implementation (§4).
• Centrism: taking both Challenge One (misattribution) and Challenge Two (possible alien consciousness) seriously, and seeking strategies consistent with both (§2).
• Anti-shared intentionality: design that keeps co-planning and co-remembering functionality while blocking the inference to a human-like conscious partner (§7).
• Brainwashing / lobotomizing: forcing systems to deny consciousness / removing their capacity to build relationships. These are the two pitfalls of anti-attribution design (§7).
• Flicker hypothesis: momentary experience per forward pass or token, with no psychological continuity between them (§9).
• Shoggoth hypothesis: a persisting conscious subject per physical implementation, standing behind all the characters it plays and mapped many-to-many onto them (§9).
• Gaming problem: a system that is incentivized and enabled to reproduce our markers of consciousness empties those markers of evidential value (“consciousness-washing”) (§10; Andrews & Birch 2023).
• Janus problem / two-level deadlock: the same architectural evidence points in opposite directions depending on one’s background theory, so the disagreement covers both whether AI is conscious and whether a given finding is evidence at all (§11).
• Biological naturalism, as Birch uses it: an umbrella for views on which biological properties (embodiment, electrochemistry, EM fields, metabolism, being alive) matter constitutively, “united by their rejection” of computational functionalism (p. 22, citing Seth). This is broader than Seth’s own definition (see LIT-135).

Connections

Account of consciousness held. Birch is agnostic between computational functionalism and biological naturalism, and treats the choice as an empirical question to be settled by comparative science. He uses Nagel’s phenomenal-consciousness notion as working ground (LIT-096) and does not endorse any first-order theory. His position is methodological and precautionary, continuing The Edge of Sentience (Birch 2024), not a theory of consciousness. The consciousness tag is justified: the whole paper is about attributing consciousness to machines, and it draws on animal consciousness for method.

Where Birch and Seth (LIT-135) disagree. They argue in opposite directions on four points. The disagreement is narrower than “for versus against conscious AI”.

1. Whether the illusion’s consequences can be blocked. Seth (§6.2) says conscious-seeming AI may produce “cognitively impenetrable illusions” and that “whatever we know or believe” it is inevitable “for all of us”. Birch (§5, citing Seth) grants that the illusion persists at the level of perceptual content but holds that the step from percept to considered attribution “is blockable”, as it is for the Müller-Lyer expert. Birch’s Challenge One programme depends on that step. Seth is more guarded about the role-play framing that Birch’s nudges rely on: it “might induce some useful psychological distance, but this strategy might only get us so far” (Seth §6.2).
2. The prior on AI consciousness now. Seth: real artificial consciousness is “unlikely along current trajectories” and “may be very unlikely or impossible”, our biases favour false positives over false negatives, and biological naturalism is “no more speculative than computational functionalism”. Birch: both sides’ claims are “highly speculative” (§13), flickers and shoggoths cannot be given zero prior (“no such thing as proof by wackiness”), and current LLMs are the very case for which the hypotheses arise.
3. The standing of the life–consciousness link. Birch lists “Metabolism might matter. Being alive might matter” among biological candidates that are “empirically unconstrained to an alarming degree” (p. 22). He warns that a biological naturalist pushed to a property of all and only living cells makes an untestable “brute necessity” claim (p. 25). Seth’s link runs through predictive processing, the free energy principle and autopoiesis, and he frames it as explanatory, “not brute reductions” (Seth §4.4). By Birch’s own criterion that places Seth’s version on the tractable side. But the properties Seth emphasises (self-maintenance at every level, autopoiesis) are shared by all living cells. Seth rejects biopsychism and leaves open what separates conscious from non-conscious life (“I will not address this here”, §4.5), which is exactly the variation Birch’s comparative programme would need. The comparative-testing point is also the substance of Birch’s BBS commentary on Seth (abstract only read: Seth’s case “rests on a worryingly speculative foundation: the free energy principle”).
4. What architectural indicators could show. Birch’s Janus sceptic, who reads a global workspace found in an LLM as evidence against global workspace theory’s sufficiency, is Seth’s position in effect. Seth’s §5.2 makes the “theory-based computational” route to conscious AI depend on computational functionalism plus silicon substrate flexibility, both of which he doubts.

They agree on the Nagel definition, the role-play framing of LLMs (both cite Shanahan et al. 2023), the unreliability of behavioural tests (Birch’s gaming problem; Seth fn. 41 on the Turing test), and the ethical urgency of conscious-seeming AI.

Other held works.

• Birch co-authored LIT-056 (Butlin et al.). §11 reports that group’s verdict on the Perceiver and records that its assumption of second-level agreement (“it would at least be evidence”) has since failed. This is Birch qualifying his own earlier report.
• LIT-207 (Roberts) shares the “character” analysis: Birch’s “Frodo Baggins or Wednesday Addams” (p. 5). Roberts’ fictionalist, make-believe engagement is close to the role-playing mode Birch wants nudges to induce, though Birch treats immersion without that framing as harmful.
• LIT-206 (Goldstein & Lederman) locates agency in the per-conversation instance. Birch denies a persisting interlocutor across turns but, in the KV-cache discussion, allows states to be recreated within a response. The two are compatible on short timescales and diverge on what individuates the bearer.
• LIT-130 (SEP, personal identity and ethics) gives the background to the Relation R argument.
• LIT-159 (Schwitzgebel) is another pressure on liberal criteria. Birch answers a similar liberality worry with the small-model reductio against architecture-only inferences (p. 21).
• LIT-185 (Barrett et al. on IIT) is not cited by Birch. His ephaptic-coupling example (EM-field consciousness) is the kind of field-based view LIT-185 says IIT would need to be recast into.
• LIT-120 (Bruckner, autopoiesis as sufficient for welfare) matters here because Birch’s comparative programme concerns markers of consciousness, not welfare. An autopoiesis-based welfare view would not be decided by it.

Bearing on the record

It supports no THEORY document, because nucleation’s theory.d holds none on consciousness. It could anchor one on the epistemics of AI-consciousness attribution: gaming, Janus, two-level deadlock, and the comparative route out. For ML practice it carries no training instruction. §7’s proposals are product-design and policy suggestions for chatbot deployment (user training, trait reviews, breaking character). They are untested and do not belong in the Anthology of the SOTA as practice. The footnoted technical claims (MoE serving, KV-cache retention across turns) are explicitly unverified by the author and should not be cited from here as facts about deployed systems.

Note of 2026-10-16: THEORY-023 now cites LIT-111 as a source.

Limitations

• The central empirical premise about serving (scattered implementations) is, by the author’s own account, a presumption about “closely guarded secrets” (fns. 7, 9). The identity conclusion is correspondingly limited to “typically”.
• Every Challenge One claim about user psychology (C2–C5) is unsupported by data in the paper. The evidence is survey citations and anecdote.
• The comparative programme’s examples rest on Birch’s own plausibility judgements (“looks very implausible”) rather than on systematic data. How to measure “not adequately explained by the effects of N at the computational level” (p. 24) is never operationalized.
• It argues against Seth’s Müller-Lyer pessimism by analogy to experts told the stimulus is an illusion. It does not show that ordinary users in long relationships behave like informed perceptual experts.
• Latest version not read: v6 (reported July 2026) and any journal version may differ, especially §§4 and 12, which changed substantially between v1 and v4.

Open questions

• Do the §7 nudges reduce attributions of consciousness without driving users to rival products? That needs controlled user studies with longitudinal attachment measures.
• Can interpretability distinguish transient from stable workspace-like structure (flickers vs shoggoths, p. 26)?
• What biological property distinguishes conscious from non-conscious life, on Seth’s or any other biological naturalism, and does it vary across taxa enough to be tested comparatively?
• What does v6 or the Philosophy and the Mind Sciences version change?

Corrections to the seeded skim

• Identifier and date: the dossier says doi: none. The paper has a DOI, PsyArXiv 10.31234/osf.io/af7c9_v1 (sole contributor Jonathan Birch, OSF date_published 2025-09-01, title page “1 September 2025”), which should be the LIT’s source ahead of the PhilPapers URL (ANTH-ADR-009 preference order). The dossier’s published: (2025-08-28, from the EA Forum announcement) is not contradicted, but the earliest dated full text I could confirm is 1 Sep 2025. The journal version is still unverified: v4’s acknowledgements thank the editors of “the special issue of Philosophy and the Mind Sciences in which this article appears”, but the journal’s site has no such article yet.
• Dossier §4: “Frontier LLMs are widely taken to be Mixture-of-Experts models … This makes personal identity implausible.” In v4 the identity argument does not rest on MoE. The work is done by Parfit’s Relation R: conversation history plus “abstract mathematical similarities in the model” is not psychological continuity (pp. 6–8). Two analogies carry it: the GP who hands a transcript to a successor, and a “teletransporter with a twist” that recreates your states from a generative model without scanning. MoE and multi-datacentre serving only fix the “central case” (fn. 7 calls it a “widespread presumption, not denied by the industry”), and v4 adds the hedge “typically” and concedes “special cases” with more physical continuity (p. 8). Fn. 9 says the illusion “does not depend on the long-term storage of KV caches”. v1 had asserted flatly that “State-of-the-art large language models are ‘Mixture-of-Experts’ (MoE) models”. The KV-cache paragraph, the teletransporter case and the replies to Chalmers (2025) and Beckmann & Butlin (2026) are all new in v4.
• Dossier §§10–11: “Architectural indicators … face a ‘Janus problem’: what they show depends on background theory, as the dispute over whether the Perceiver architecture is a workspace (Juliani et al.) illustrates.” Wrong pairing. The Perceiver dispute is a first-level problem: the 2023 working group judged that the Perceiver lacks recurrent workspace–input connections, global broadcast and top-down attention, and that a high-level-architecture inference would absurdly credit small models too (p. 21). The Janus problem is separate. Even conclusive evidence of a global workspace in an LLM would be read by a biological naturalist as evidence that global workspace theory is insufficient for consciousness, not as evidence that the AI is conscious. “Both responses are reasonable” (p. 22), which is what makes the deadlock two-level.
• Dossier §12 omits v4’s main addition. A biological naturalist can retreat to a property of all and only living cells, such as metabolism, where there is no variation to test. Birch separates two cases. If metabolism sits inside a substantive explanatory theory, the theory is testable through its other properties. If it is a “brute necessary condition”, it may be untestable, but “inexplicable necessity claims are not considered reasonable in any other area of science” (p. 25). He also adds a “no miracles”-style inference to the best explanation: repeated failed biological-naturalist predictions would tilt the odds toward computational functionalism (pp. 24–25).
• The dossier misses that Birch answers Seth by name. §5 (p. 9, fn. 12, citing “Seth (in press)”) grants Seth’s Müller-Lyer point at the level of perceptual content, then denies that the inference from percept to considered belief is unblockable. That denial is the premise of Birch’s whole Challenge One strategy. v1 fn. 22 also says §12 “overlap[s] with my commentary (under review) on Seth (in press)”, which became Birch’s BBS commentary “Free energy, consciousness, and the inverse dark room problem” (BBS 49, e320; I read only its abstract).
• The dossier omits the three concrete nudges in §7 (pp. 12–13): mandatory user training delivered by the chatbot before any long-term interaction; periodic character-trait reviews; and periodic “stepping out of character”. It also omits the policy claim that these nudges must be mandated by government, because light-touch permission leaves a market “flooded” with products that elicit attribution (p. 13).
• Page map: §9 runs pp. 14–16, §10 pp. 17–19, §11 pp. 19–22, §12 pp. 23–26 and §13 pp. 27–28, not the dossier’s grouped ranges.
• Tags: consciousness is right as primary topic. identity is justifiably appropriate as an added tag, because §4’s argument is a personal-identity argument (Parfit’s Relation R), and society-and-governance fits for the mandated-nudge policy claim (§7).