NOTE-1071: Graziano — Illusionism big and small

Contribution

The paper presents illusionism as a framework rather than a single theory, and shows that existing theories can be fitted into it. Its device is a four-item chain from the visual neuroscience of report: Item 1, the thing modelled; Item 2, an automatic model of it; Item 3, higher cognition; Item 4, speech. Applied to the claim “I have a subjective experience”, the question becomes what Item 1 is. Six answers are given: a magical essence, nothing, integrated information, the global workspace, attention, and the deep selective processing that the last three share. The positive proposals are the author’s attention schema theory and a “deep illusionism” that absorbs IIT and the workspace into it.

Key insight

Whatever consciousness is, a person can believe and say they have it only if their brain encodes information that they have it. That encoding is a model, and models in the brain are simplifications. So what people report, an intangible what-it-is-like essence, is a model’s depiction and need not be a property of what is depicted. A real process can be described as an ethereal essence, as a paper map describes New York. Hence “caricature” rather than “illusion” (p. 2).

Assumptions

• Every claim needs an encoding. “if Bob believes he has conscious experience, if he says he has it, then whatever conscious experience is or is not, Bob must have a neural embedding that represents it” (p. 4).
• All brain models are inaccurate. “The brain’s internal models are never fully accurate, nothing is exactly as the brain represents it” (abstract). This is offered as something “all neuroscientists should know” (p. 7).
• Causal closure of report. A feeling can explain a report only if it physically affects neurons that encode it (Arrow B, p. 4). The physics of the brain is taken to be complete without it.
• Models track useful real things. Evolution “shapes the brain to construct models that can monitor and predict real items in the world” (p. 6). This is the argument against hard illusionism.
• The control rationale. “any controller performs better if it has a model of the thing it controls” (p. 8). This motivates a model of the workspace, or of attention.

Key results

There are no new data. By section:

• From vision to speech (pp. 2–3, Figure 1). An apple (Item 1) is encoded in a visual model (Item 2) that informs higher cognition (Item 3) and speech (Item 4). The visual model is automatic, and “necessarily a simplification”: the apple has a reflectance spectrum, not a colour (p. 3).
• From self-model to speech (p. 4, Figure 2). Bob’s “I have a subjective experience of it” needs an Item 2 that “encodes the presence of subjective experience”. Graziano places it at the temporoparietal junction for reasons from his lab and says the location is “not important to the present argument”.
• Hypothesis 1, the magical mind, and Arrow B (pp. 4–6, Figures 3–4). Arrow A is how neurons produce a feeling. Arrow B is how the feeling, once produced, “cause[s] a specific information embedding” (p. 4), which would need neurons that respond to “the emergent feeling itself”. Without Arrow B, “the conscious feeling that emerges from the brain cannot be the consciousness that we say we have” (p. 5). Even with it, what Bob reports “is entirely dependent on the information embedding”, and there is “no reason to suppose that the model is accurate”. So “Even the traditional account, when held to a basic standard of logic, turns into an illusionist account” (p. 6). The blood-flow analogy: the brain has a blood pattern but no model of it, so no one reports it (p. 5).
• Hypothesis 2, hard illusionism (pp. 6–7). No Item 1: the model depicts nothing, perhaps learned culturally. Graziano finds it unlikely, because evolution would not build costly models of nothing and “the sheer scale and consistency of the bamboozlement seems improbable”. Against it, “Subtle illusionism is necessarily true of all models in the brain and must be true of consciousness as well” (p. 7).
• Hypothesis 3, illusionist IIT (p. 7). Item 1 is the brain’s level of integrated information, detected and modelled. Then “consciousness resembles integrated information because consciousness is a distorted representation of integrated information”. The hard problem “is dissolved”, and panpsychism goes, since integration without a model “would have no relationship to consciousness”. The open problems are whether neurons can detect integration and what the model is for.
• Hypothesis 4, illusionist workspace (pp. 7–8). Item 1 is the global workspace, and the brain models its own workspace. Item 3 is also the workspace, so “the global workspace contains simplified information about itself”. The model “depicts me as a magical essence inhabiting my body, capable of subjective, phenomenal experience and of a conscious will” (p. 8). People do not become conscious of the model: “the model contains information” that drives belief and report.
• Hypothesis 5, the attention schema (pp. 8–10). How to identify Item 1: by analogy with the body schema, find something that can affect neurons, resembles the reported consciousness in outline, and “should covary with Bob’s reported consciousness almost all of the time” (p. 9). Attention is offered as meeting all three. It covaries with reports (inattentional blindness), though it can be dissociated from them in the laboratory, which the paper reports as “difficult, though possible” (p. 10, citing blindsight and attention without awareness). AST is called an “illusionized” higher-order theory: the model of attention is a higher-order representation, but automatic (p. 10).
• Hypothesis 6, deep illusionism (pp. 11–12). Attention, the workspace and integration are “different descriptions of the same underlying object—the deep, selective processing of information in the cerebral cortex” (p. 11). Treisman’s feature integration is cited for the link between attention and integration. The theory also explains the field: each school reads off a different feature of the same “schematic, internal model” (p. 11). Where consciousness resides is a “poorly formulated” question: “One cannot point to a part of that framework and say, ‘That’s where the consciousness essence lives’” (pp. 11–12).
• Why it matters (p. 12). “there are now only two broad classes of explanation for consciousness: magic and illusionism”. The framework “could have applied to artificial as well as biological consciousness”.

Claims

id	claim	strength	support
C1	Any belief or report that one is conscious requires a neural encoding of that content	moderate: follows from a physicalist account of report	pp. 3–4
C2	A theory that names a mechanism producing a feeling must also explain how the feeling causes the encoding that drives report (Arrow B), or it does not explain the reports	moderate as an argument; assumes causal closure	pp. 4–6, Figures 3–4
C3	Since brain models are schematic, what people report of consciousness need not match its source (“subtle illusionism”)	moderate as a general point about models; weak as applied, since a schematic description need not be a false one	pp. 6–7
C4	Hard illusionism (a model of nothing) is unlikely	weak: an evolutionary plausibility argument	pp. 6–7
C5	Attention meets the criteria for Item 1: causal access, outline resemblance, covariation with report	weak to moderate: covariation cited from others; the dissociations noted	pp. 9–10
C6	Attention, the global workspace and integrated information are three descriptions of one process, deep selective processing	weak: asserted from overlap, no formal argument	p. 11
C7	The question where consciousness resides is ill-formed	interpretive: a consequence of the framework, not argued independently	pp. 11–12

Concepts

• Item 1–4: the thing modelled, its automatic model, higher cognition, speech output.
• Arrow A / Arrow B: the production of a feeling by neurons; the feeling’s effect on neurons that encodes it for report.
• Magical mind theory: any theory that takes reported phenomenal experience to exist as reported and supplies Arrow A without Arrow B. IIT and the workspace theory in their usual forms are counted as such.
• Hard illusionism: no Item 1; the model depicts nothing.
• Subtle illusionism: an Item 1 exists, and the model depicts it schematically and inaccurately.
• Caricature: Graziano’s preferred word: a real subject distorted, not a fake.
• Attention schema: the brain’s automatic model of its own attention, which depicts it as “an amorphous, mental essence, a mind vividly possessing something in the moment” (p. 10).
• Deep illusionism: Item 1 is the deep selective processing of information, of which attention, workspace and integration are aspects.

Connections

The paper builds on Graziano’s own attention schema theory (2011–2020), on Dennett (1991, 2018, “And then what happens?”) and Frankish (2016) for illusionism, on Nisbett and Wilson and Gazzaniga for confabulated self-report, on Baars, Dehaene and Mashour for the workspace, Tononi for IIT, Treisman and Gelade for feature integration, and on Kentridge and others for attention without awareness. Of these the record holds Frankish (LIT-458), the global neuronal workspace review (LIT-631), IIT 3.0 (LIT-445), feature integration theory (LIT-1082) and Chalmers’s meta-problem (LIT-425). The paper does not cite Chalmers 2018, though Arrow B is close to its challenge.

Bearing on the record

• CLAIM-355 (real binding, constructed owner). The exchange cited the paper “to show that illusionism need not make consciousness wholly fictional”, with an illusionist layer on a workspace or integration account “so that real organization is misdescribed by a simplified self-model”. That matches Hypotheses 3, 4 and 6. Read against the claim:
  • It supports the structure, at the level of one brain. The workspace is real, and the model depicts it as “a magical essence inhabiting my body” (p. 8): a real organization, and an owner that is the model’s content. The paper also says which system holds the representation (Item 2, read by Item 3), which CLAIM-355 requires of an illusion.
  • It goes further than the claim. What it calls the misdescription is not only the single owner but the phenomenal character itself: “The subjective, phenomenal experience that Bob describes is a distorted construct” (p. 7). CLAIM-355 explicitly leaves open whether experience is real, and Graziano’s “small” illusionism closes it on the illusionist side. So the paper is a source for the stronger option CLAIM-355 sets aside, not for CLAIM-355 itself.
  • It complicates CLAIM-355’s test, a misdescription shown. Graziano’s arm analogy (pp. 8–9) is a model that is schematic and mostly right. By that standard, a self-model of attention that presents an owner may be a caricature with true outlines, which is Drayson’s point.
• THEORY-633 (Frankish’s strong illusionism). The exchange read the paper as showing that illusionism need not make consciousness wholly fictional. That is right only if “consciousness” means the process modelled. Graziano’s “small” or subtle illusionism is not Frankish’s weak illusionism: it does not keep real phenomenal properties that are merely misdescribed. It keeps a real Item 1, deep selective processing, and denies that the reported “intangible essence” is a property of it. That is Frankish’s strong illusionism (quasi-phenomenal properties misrepresented as phenomenal) with Item 1 identified. Graziano’s “hard” illusionism, a model of nothing, is more extreme than anything THEORY-633 holds. So “big and small” divides illusionists by whether the model has a target, while Frankish’s strong and weak divide them by whether phenomenal properties survive. The comparison is the record’s; the paper does not draw it. It does supply what THEORY-633 says Frankish lacked, a mechanism for the representation, though as a hypothesis.
• THEORY-577 (Chalmers’s meta-problem challenge). Arrow B is the same constraint, stated in 2013 and here without citing Chalmers: a theory that names a basis of consciousness must say how that basis produces the reports. Graziano draws the conclusion Chalmers resists, that the gap is fatal to “magical” theories. His i-IIT answers Chalmers’s challenge to IIT by making integration the target of a model. Chalmers’s reply, that a zero-Φ duplicate would report alike, does not touch i-IIT, which needs the model and not Φ.
• CLAIM-412 (experience constituted by interpretive organization). The paper cuts against the claim’s bridge. On deep illusionism, the organization (deep selective processing) is real, and what is called experience is its representation (TERM-072 sense 2). No further fact of a phenomenal perspective (sense 3) needs constituting, and the question where it resides is “poorly formulated”. This is the strong-illusionist objection CLAIM-412 lists, now with a mechanism. It also restates the claim’s open sufficiency condition as a question with an answer: an interpretation is reported as conscious when the attention schema models it.
• CLAIM-411 (most interpretation is unconscious). AST supplies a candidate for the further organization the counter-claim asks for, and it is a represented, not organizational, condition. Processing is claimed as experienced when it is attended and modelled as attended. The paper itself reports the dissociations that test this: attention without awareness in blindsight (Kentridge et al. 1999), which it calls “difficult, though possible” to produce (p. 10). That fits the theory if awareness reports track the schema and not attention itself. But it makes the attention schema, not attention, the separating condition. It meets CLAIM-411’s defeater only if being modelled by the attention schema can be stated without appeal to report, which the paper does not do.
• QUESTION-048. Every criterion here is covariation with report, as the method requires: illusionism takes reports as the explanandum. It answers the question by changing it. There is no experience beyond access to explain.
• THEORY-064 (Metzinger’s self-model). Graziano’s workspace model “depicts me as a magical essence”, a self-model content the system cannot see through. The two agree on the self, and Graziano goes on to the phenomenal.
• THEORY-063 (unity). “Everyone knows that consciousness is both rich and at the same time unified” (p. 11) is read as a feature of the model, the source of IIT’s appeal. It gives no account of unity.
• THEORY warranted. One is filed, THEORY-873: deep illusionism, with Arrow B as its argument and attention as its proposed target.
• No instruction for machine-learning practice. The remarks on artificial consciousness are speculative and explicitly set aside (p. 12).

Limitations

• An essay. The empirical support is cited, and only for attention and report covarying.
• The argument slides between two senses of illusion. “Nothing is exactly as the brain represents it” makes any model inaccurate, which is trivially true and says nothing about phenomenal properties. The conclusion, that the intangible essence is a distortion, needs the stronger claim that the depicted property is absent from the target. That is assumed, on causal-closure grounds.
• Arrow B assumes the feeling is distinct from, and must act on, the neural process. A view on which the feeling is the process, or is the model’s activity, needs no extra receptor. The paper treats every realist theory as positing a further essence.
• Deep illusionism’s unification of attention, workspace and integration is asserted from their overlap. Treisman’s feature integration is cited for the link to integration, which is a different notion from IIT’s.
• The tone is polemical (“a Trojan horse for the entry of overt spiritualism”, p. 6), and the alternatives are named “magic”.
• The temporoparietal location of the model is set aside as unimportant.

Open questions

• What would separate subtle illusionism from realism about a schematic but true self-description? Graziano’s arm schema is mostly right, so a mostly-right attention schema would make “illusion” a misnomer.
• Can being modelled by the attention schema be stated without report, so that blindsight-type dissociations test it rather than define it?
• Does the paper’s claim that integration and attention are one process survive the formal difference between Treisman’s integration and IIT’s?