Source
Laukkonen, Friston & Chandaria (2025), LIT-1363, read in full in NOTE-1063 from the PsyArXiv preprint, version 3: sections 2–5 and 7 (pp. 3–20), Table 1 (p. 13) and the discussion (pp. 26–30). Page numbers are the preprint’s.
What was actually shown
Nothing was tested. The paper argues by construction and redescription.
- The conditions. A world model is “a necessary condition for consciousness because it defines what can become conscious” (p. 3). Entry is a “precision-weighted competition between possible explanations for the causes of sensory data” (p. 7), in which “coherence and boundedness are a central criterion” (p. 6). The global model is “necessary but not sufficient for conscious experience. It is only when this global posterior is reflected back through the underlying hierarchy that the conditions for consciousness are met” (p. 9).
- The formalism. Table 1 conditions each level of a hierarchical generative model on a precision hyper-parameter φ⁽ˡ⁾ under a common prior p(Φ), and minimises a free energy over all levels and Φ. The text says “It is this recursive aspect that equips belief updating with epistemic depth” (p. 11). No simulation is run.
- The redescriptions. Inattentional blindness, subliminal priming, lucid and non-lucid dreaming, lucid dreamless sleep, minimal phenomenal experience, cessation and psychedelic states are each placed by their epistemic depth, abstraction and precision (pp. 19–26). Cessation is the coherence condition failing: “the world model is not contentless but incoherent” (p. 25).
- What could have come out otherwise. As stated, little. The paper says “The theory is testable and falsifiable at the level of computational modeling, but also in terms of neural implementation” (p. 30), and names as its test that if the conditions are met “we ought to see evidence of awareness or deep and flexible epistemicity, as well as success on any Turing-type tests” (p. 30). It names no observation that would count against it.
What this does not say
- It does not show that the hyper-model knows anything. A hierarchy with learned layer-wise precision is a standard model; the paper itself names the hierarchical Gaussian filter (p. 12). That such a model “contains the knowledge that it exists” is asserted, not derived.
- It does not guarantee experience. Note 11: “Our position does not assert that satisfying these conditions guarantees phenomenal experience in the subjective sense” (p. 29), though the main text concludes that an AI meeting the conditions should be judged conscious.
- Epistemic depth is defined by awareness. It is “a capacity or continuum (i.e., deepening) of knowing, or awareness” (p. 9), so the third condition is not stated independently of what it explains.
- The contemplative evidence is report. MPE and cessation enter as reported states. Whether meditative “contentless” states lack content is disputed by a work the record holds unread (LIT-1351).
- It is not a higher-order theory. The paper separates its recursion from metacognition (p. 20). It is a claim about global, recursive precision control.
Connections
- QUESTION-254. Answers it in part with a recursive, higher-order candidate, on the organizational side. It fails the question’s test of saying what a nonexperiential interpreter would lack, because an HGF has Table 1’s structure.
- QUESTION-255. The coherence condition takes the side of unity as consistency, and makes incoherence the loss of awareness. It does not meet the felt-conflict case.
- TERM-072. It denies that a first-person perspective, in the represented or phenomenal sense, is required, while its world model is an organizational perspective.
- THEORY-063. Shares Metzinger’s ordering of world before self and adds a mechanism, without evidence.
- CLAIM-414. Epistemic depth is endogenous allocation of efficacy made global. That claim shows behaviour cannot reveal it, so the paper’s behavioural tests for AI cannot either.
- THEORY-883 (Whyte et al.) is the access criterion this paper calls complementary. THEORY-876 (IWMT) shares the coherent world model and differs on embodiment, which this paper does not require.
Note of 2026-10-30: what holds it back
No simulation in the record shows a property of the hyper-model that a hierarchical Gaussian filter lacks, and no measure of global precision control taken apart from reports of awareness is in the record. The paper has neither (NOTE-1063).