Attention Schema Theory

Question 1 asks for a causal generative mechanism for WHY phenomenal experience arises in a physical system — engagement with the hard problem. AST, by its proponents' own explicit commitment, declines this: Graziano states the theory "emphatically does not explain how we have a subjective experience. It explains how a machine claims to have a subjective experience, and how it is that the machine cannot tell the difference." AST is a meta-explanation/dissolution of the hard problem (illusionism): it explains why a system REPORTS phenomenality (it introspects a simplified, mechanistically incomplete attention schema that depicts awareness as non-physical), not why there is something it is like to be that system. Under the strict standard, explaining the report while presupposing there is no extra phenomenal ingredient to generate is precisely a NO — it identifies the information state described as experience without giving a generative mechanism for experience itself.

Key evidence: Graziano's explicit statement (Graziano & Webb 2015; Graziano 2013) that AST "emphatically does not explain how we have a subjective experience. It explains how a machine claims to have a subjective experience, and how it is that the machine cannot tell the difference."

The question asks WHY the experiencer and the agent are the same entity (the unity of subject and initiator, distinguishing 'I' from 'my body'). AST's own commitments, by its proponents' description, only stipulate that the same self-model (S) figures both in the awareness representation (S+A+V) and in motor/cognitive control, so experiencer and agent "are modeled as the same entity." This is a NO under the strict standard: positing a shared self-model is exactly the move the yes-bar excludes — it asserts that one representation does double duty without explaining why that representation IS a unified subject-agent rather than two co-located models, and it offers no account distinguishing the 'I' from 'my body'. AST itself concedes it gives no "deep metaphysical argument for why these must coincide" and "treats their coincidence as a feature" — i.e., it assumes rather than explains the identity. Identifying where/how a self-model is reused is a WHERE/HOW-it-is-implemented answer, not the asked-for WHY of experiencer-agent unity.

Key evidence: Per the theory's stated position (Graziano & Webb 2015; Graziano 2013), AST "does not offer a deep metaphysical argument for why these must coincide; it treats their coincidence as a feature of integrating self-, attention-, and action-models in one representational system."

AST identifies consciousness with the attention schema — a physical, information-bearing internal model the brain builds and uses to monitor and control attention. This gives consciousness an explicit, non-epiphenomenal causal role: like a body schema improving limb control, the attention schema improves top-down (endogenous) control of attention because a controller performs better when it models what it controls. AST distinguishes this from mere correlation via a concrete mechanism (model-based control) plus a falsifiable dissociation prediction — remove an adequate schema and attention control degrades (gets "stuck," capture is harder to suppress). Crucially, on AST consciousness IS that causally efficacious model, so under the theory's own commitments there is no separate inert phenomenal essence; the causation is genuine and mechanistic, not a promissory note.

Key evidence: Wilterson & Graziano (PNAS 2021, e2102421118) demonstrated that adding an internal model of attention to a deep RL agent improved control of visuospatial attention, operationalizing AST's claim that the attention schema (= awareness) plays a real control function rather than being epiphenomenal.

AST's core commitment is an account of what awareness IS (the brain's attention schema, the S+A+V model), not of why tonic arousal varies. By its proponents' own framing, the theory is "thin on graded arousal/state changes" and explicitly delegates global tonic arousal to "underlying brainstem/thalamocortical mechanisms it does not model in detail." Its only implied account of level — that awareness reports cease when the S+A+V model is not being constructed/updated (deep sleep, anesthesia) — is a correlation/gating observation, not an intrinsic control mechanism explaining HOW the transitions between wakefulness, non-REM sleep, and anesthesia are generated. In the proposed "standard model," level/access aspects are handed off to Global Workspace Theory, confirming AST itself supplies no internal state-control mechanism. This fails the yes-bar's demand for an intrinsic internal mechanism over mere external correlation.

Key evidence: Graziano & Webb (2015) and the standard-model paper (Graziano et al. 2020) frame AST as a theory of subjective awareness as a model of attention, with level/arousal left to brainstem/thalamocortical systems and access/level functions delegated to GWT — i.e., AST offers no intrinsic mechanism for graded states.

Question 5 asks only which cognitive/psychological functions are enabled by consciousness, and AST gives a specific, positive answer rather than a vague "it's useful" gesture. On AST, awareness IS the attention schema, and Graziano/Webb explicitly tie that model to named functions: (1) improved endogenous control of one's own attention (the schema as a control model), (2) social cognition / theory of mind via attributing attention and awareness to other agents, and (3) self-attribution of consciousness underwriting introspective report and certainty about having experience, plus binding/integration of information into a coherent reportable representation. These are concrete functional roles, and the control-of-attention claim is mechanistically operationalized and empirically tested in Wilterson & Graziano (2021), where a neural-network agent uses a descriptive attention-schema model to control visuospatial attention. The hard-problem objection in the standard targets the "why is there something it is like" question, which is a different question; it does not undercut AST's explicit, specific account of function.

Key evidence: Wilterson & Graziano (2021, PNAS 118(33)) show a neural-network agent using a descriptive attention-schema model to improve control of visuospatial attention, directly instantiating the named function "consciousness/awareness enables endogenous attention control."

Question 6 asks specifically about the diversity of conscious contents, their semantic integration, and the role of attention/intentionality in gating what enters consciousness — and AST is fundamentally a theory of exactly this. Its core commitment is that awareness IS the brain's model of attention, so awareness inherits attention's full range of targets (vision, audition, color, motion, space, plus internal thoughts, emotions, memories); content diversity is therefore mechanistically derived from the diversity of what attention selects, not left unexplained. The attention-determines-content claim directly supplies the required account of how attention/intentionality gates entry into consciousness, and the dissociation framework (attention without awareness, and rivalry as competitive attentional selection determining which percept is modeled) addresses the binocular-rivalry switching the yes-bar names. The integrated self-attention-stimulus (S-A-V) representation also speaks to the semantic-integration requirement. The residual gap — why modeled content feels like anything — is the hard-problem residue, which Question 6 does not ask about.

Key evidence: Graziano & Webb (2015) and Graziano et al. (2020): awareness can take the same range of contents as attention because awareness is the model of attention, with awareness and attention dissociable but normally tightly coupled — directly tying content diversity and the attentional gating of consciousness (including rivalry-type selection) to a single mechanism.

AST defines awareness as the content of an information-processing model (the attention schema) — a self-model S, a modeled attention relation A, and a represented stimulus V — none of which is tied to biological neurons or specific neuroanatomy. Graziano explicitly frames AST as substrate-independent: his 2017 paper is titled "The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness," arguing any system that builds a rich model of its own attention would, when reporting, claim awareness and "cannot tell the difference." The 2021 PNAS neural-network agent is a concrete artificial instantiation showing an attention schema implemented outside biology. The Q7 yes-bar requires only substrate-independence and in-principle applicability to artifacts, not a solution to the hard problem, so AST clears it.

Key evidence: Graziano, M. S. A. (2017). The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness. Frontiers in Robotics and AI, 4:60 — explicitly presents AST as substrate-independent and applicable to machines, supported by the artificial attention-schema agent in Wilterson & Graziano (2021, PNAS).