Contribution
It asks whether predictive processing (PP) can accommodate mental action, given that “In mental actions there is no motor plant to be controlled, no sensory manifold that could be manipulated by bodily movement” (p. 1). It proposes four “building blocks” towards a working concept: attentional agency as control of precision expectations; cognitive agency as selection from an internal “affordance competition” set up by mind-wandering; the convergence of instrumental and epistemic action for mental action; and epistemic value as the goal. It ends with three open problems and does not answer whether mental actions exist (p. 22).
Key insight
The candidates for mental action are produced by a process that is not an action. “Mind wandering is mental sleep-walking. But unlike embodied somnambulism it actually sets up the inner world which later allows us to autonomously perform mental actions” (p. 12). Thinking is then “controlled mind wandering” (p. 12), as perception is controlled hallucination. Control does not author the content; it selects and sustains one of the offered thoughts, and with M-autonomy can “actively sculpt its own internal affordance landscape” (p. 13).
Assumptions
- Predictive processing, with active inference as the account of bodily action (pp. 1, 9).
- Mental actions are epistemic; reward-directed mental action is set aside, with the suggestion that reward expectation reduces to epistemic value “for the very large majority” (p. 3).
- M-autonomy and veto as in LIT-1941 (p. 4).
- “Roughly … one third of our conscious lifetime” of autonomous self-control, from the 2015 paper’s figure 2 (p. 12).
Key results
- Attentional agency (p. 10): “the ability to control precision expectations”; it is mental because “what is controlled is a property of a representational entity, namely the precision expectation embodied by an internal model”. All volitional attention is “active introspection” (p. 9).
- Effort as collision (p. 10), offered as speculation: “Perhaps we can describe this experiential tension as the way in which active, top-down precision control ‘collides’ with an automatically arising precision prediction error. If so, one might speculate that the temporal curve in which the phenomenal sense of agency unfolds over time exactly reflects this ‘collision’ and its eventual resolution.”
- The cognitive affordance hypothesis (pp. 11–13), stated above; “the mind wandering network sets up this task domain, and sometimes frontal regions of the brain latch onto the best candidate, and create a transient functional integration with the EAM”.
- Three options for the EAM (pp. 13–14): epiphenomenon; “exactly the missing computational element that constitutes epistemic, model-based predictive control”; or “the EAM would not constitute, but represent. It might causally enable relevant second-order functional properties like veto control”. “Currently available empirical data do not help to decide between these three metaphysical options.”
- H1–H4 (pp. 14–15), untested: blocking the EAM’s introspective availability should trigger mind-wandering (H1); re-establishing predictive self-control should induce lucidity (H2) or end an episode (H3); blocking the subject–object loop by TMS should remove the phenomenology of epistemic goal-directedness (H4).
- Model-free and model-based (p. 15): “mind wandering is model-free epistemic self-control, whereas mental agency is model-based epistemic self-control”; after Conant and Ashby, mind-wandering is “self-regulation on the mental level without an explicit model of the ‘epistemic plant’” (p. 17).
- The working concept (p. 17): “the predictive control of effective connectivity aimed at optimizing the epistemic value of attentional and cognitive states embedded in the system’s phenomenal self-model, accompanied by the emergence of a new content layer … the EAM”.
- Agency as enculturated (pp. 20–21): caregivers ascribe veto and “ultimate origination” to children who lack them; the ascription is internalised before autobiographical memory; so the adult’s model of self-causation “could therefore be an enculturated post-hoc confabulation”. A model of initiation exists for cognitive agency (the affordance competition) “but not for AA” (p. 22).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Mental action cannot be active inference through the non-neural body | moderate | conceptual: no sensory sheet in the loop (pp. 1, 16–17) |
| C2 | Attentional agency is control of precision expectations of an internal model | weak; proposal | conceptual, after Hohwy 2013 (pp. 9–10) |
| C3 | Mind-wandering supplies a competition among candidate cognitive actions, from which cognitive agency selects | weak; hypothesis | conceptual, after Clark 2016 and Cisek; DMN and FPN literature cited (pp. 11–13) |
| C4 | Whether the EAM constitutes, enables or merely accompanies mental control is undecided by current data | strong as a report of the state of evidence | the author’s own survey (pp. 13–14) |
| C5 | Felt mental effort tracks a collision of top-down precision control with bottom-up precision prediction error | speculation | none (p. 10) |
| C6 | The sense of origination in agency may be an enculturated post-hoc confabulation | weak; one horn of a dilemma | three-phase sketch; Khalighinejad et al. 2016 cited (pp. 20–21) |
Concepts
- Sensory sheet — “a collection or population of receptors” (p. 2, note 1).
- Mental action / mental behaviour — as in LIT-1941; actions are “typically directed at epistemic goal-states” (p. 3).
- Cognitive affordance — “Every spontaneously occurring ‘task-unrelated’ mental event is a potential task in itself … a dynamic state which has the potential to be selected and transformed from unintentional mental behavior into mental action” (p. 12).
- EAM-space — “a new subregion or partition of the organism’s global self-model, constituted by all those hypotheses exclusively predicting the organism’s epistemic properties” (p. 14).
- Epistemic value — “the expected information gain under predicted outcomes” (Friston et al. 2015, quoted p. 17).
- The bubble in the bubble (p. 18) — modelling mental action as active inference posits a second Markov blanket inside the brain and invites an “epistemic surfer” homunculus.
- Self-deception model of goal selection (p. 19) — “all goal-representations are misrepresentations” from a naturalist third person; organisms solve self-motivation “by hallucinating goals”.
Connections
It builds on LIT-1941 and LIT-1942 (cited as Metzinger 2015 and 2013a), Hohwy 2013 on attention as precision, Clark 2016 and Cisek on affordance competition, Seth 2015b on predictive control, and Friston et al. 2015 on epistemic value. THEORY-882 (Laukkonen, Friston and Chandaria, read in NOTE-1063) is a later PP account in which a hyper-model sets precision at every level and agency is not required for awareness; this chapter makes agency the special case where precision is controlled under a self-model.
Limitations
- No data, no simulation, no equations. The PP terms are used qualitatively; nothing is derived.
- The EAM’s role is left open, by the author’s own statement, so the chapter cannot be used to show that the self-model is a cause of mental control.
- No model of attentional action initiation (p. 22), which is the type M-autonomy’s veto most often concerns.
- “Effective connectivity” is named as what is controlled and never made precise.
Open questions
- H1–H4: does blocking the EAM start mind-wandering, and does restoring control end it or induce lucidity?
- Is felt mental effort a measure of conflict between a bottom-up claim on precision and top-down allocation, and could the two be recorded apart from the feeling?
- Realism or elimination about mental action (Problem 3)?
Bearing on the record’s own argument
Same structure, not shown same mechanism, throughout.
- CLAIM-027 and TERM-023. The cognitive affordance hypothesis has the claim’s shape: several candidate mental processes are available at once, and something other than their content decides which is pursued. “Frontal regions … latch onto the best candidate” (p. 12) is precedence; veto is suspension; sculpting “its own internal affordance landscape” is transformation. “Best” is not defined, so whether selection is by the candidates’ own strength (content) or by maintained relations is exactly what the chapter does not say. The support verdict is ARG-051.
- CLAIM-484. The effort passage (p. 10) is the nearest thing in these readings to the episode the claim needs. Its terms are defined apart from the felt mark: an “automatically arising precision prediction error” is a claim on precision, top-down precision control is the grant, and the sense of effort is the mark, whose “temporal curve” would track their “collision”. If that held, gap and mark could be measured separately at the scale of attention. The passage is speculation, and no study is cited.
- CLAIM-358. Mind-wandering is “model-free epistemic self-control” and mental agency “model-based” (p. 15); “mind wandering is self-regulation on the mental level without an explicit model of the ‘epistemic plant’” (p. 17). That is the claim’s contrast, regulation without a mediator against regulation through a higher model, at the scale of thought. The chapter gives no costs, so it does not bear on the claim’s amortization. The cost of control the claim leans on is THEORY-070’s.
- CLAIM-355 and the record’s split between effective organization and its representation (not this reader’s line for CLAIM-355). The chapter keeps the split open where M-autonomy closed it: the EAM may constitute the control, enable it, or accompany it (pp. 13–14). That is the question CLAIM-355’s A26 answers one way (“self-representations may themselves be parts of the real causal organization they represent”). The reading neither supports nor defeats it. It names the interventions that would decide (H1, H3). QUESTION-367 is filed for them.
- QUESTION-364. The standard by which an adult reads its own acts as self-originated is installed by caregivers’ ascriptions “at a stage where … young children almost certainly do not possess these capacities”, and “the process of amnestic enculturation itself will never be introspectively available” (pp. 20–21). If so, the history of authorization the question asks about begins with a standard the agent did not authorize and cannot recall acquiring. That sharpens the question; it supplies no criterion.
- CLAIM-122. The selection story allows the defeater’s route: if “best candidate” is fixed by the candidates’ epistemic value, a change in that value, not in a governing relation, predicts which thought is pursued. The chapter does not separate them.
- No instruction for machine-learning practice; the PP apparatus is about brains.