Contribution
The paper proposes that the psychology of processing fluency can be restated in active inference. Fluency is read as the brain’s recognition of the precision of its own predictions. What is first registered is not fluency but a change in it relative to expectation. The paper’s own addition is the pairing of Whittlesea’s discrepancy-attribution hypothesis with precision-weighting, and a two-level vocabulary for it: unfelt fluency below, felt fluency above.
Key insight
The feeling of ease is an inference about one’s own inference. Its cause is the precision of internal processing, not anything in the world, and introspection does not reveal that cause. So when expected and experienced ease come apart, the surprise is attributed to the most salient candidate, usually the stimulus, and shows up as familiarity, liking, nearness or a sense of error.
Assumptions
- The predictive-brain premise. The brain “predicts the content of the sensorium and crucially the precision of … its own predictions”. It is taken as accepted, not argued.
- Precision is attention. “Psychologically, this kind of precision weighting is often equated with attentional set”; physiologically it is synaptic gain under neuromodulation (dopamine, noradrenaline).
- Fluency is a feeling by definition. Fluency is “the subjective (metacognitive or qualitative) experience of the level of ease with which our own cognitive processes proceed” (after Alter and Oppenheimer 2009). “Unfelt fluency” is therefore a new, extended use.
- Bayesian beliefs are subpersonal: “implicit probability distributions encoded by neuronal activity and synaptic gain”, not propositional attitudes.
- Helmholtz’s unconscious inference: “any mental state is realised by unconscious inference”, whence “the causes of the subjective (conscious) inference must have unconscious perceptual precursors”.
Key results
The paper reports no simulation and no experiment of its own. Its results are a review and a proposal.
- Absolute and relative fluency (pp. 8–9). After Whittlesea and Leboe (2003): absolute fluency is the degree of ease, as in priming; relative fluency is ease against expectation, as when something expected at once appears after a delay and “will seem more fluid than if it had appeared immediately”.
- The hension effect (pp. 9–10), as reported. In Whittlesea and Williams (1998, Exp. 3), new orthographically regular non-words (“hension”) drew more false recognitions than new real words or irregular non-words, although real words were read fastest. Read as surprise: easy to pronounce against hard to identify. “Participants cannot attribute surprise to its source, so they attribute the fluency felt to the most obvious source, the stimulus”.
- The same pattern, extended (pp. 10–13), as reported.
- Conceptual against perceptual fluency (Brouillet et al. 2023): the discrepancy acts “both anteriorly (forward) and posteriorly (backward)”.
- Motor fluency of an unrelated preceding gesture (Brouillet et al. 2017) and of a colour associated with a fluent gesture (Brouillet, Rousset and Perrin 2022) raises “old” responses to new items.
- Typing (Wilbert and Haider 2012, Exp. 4): the sense of having made an error comes from “a discrepancy between expected fluency and experienced fluency”, resolved by searching for a cause.
- Conclusion drawn: “the phenomenological perception of a gap directs attention, and … the fluency experienced is merely the consequence of a non-felt fluency or an expected fluency.”
- Unfelt fluency (pp. 14–16). Surprise “is not a state (i.e., an emotional state of mind …)” but self-information. It drives a stimulus-bound change in attentional set. “These subpersonal changes can then be recognised as unfelt fluency; namely non-conscient recognising a change in attentional set”. When precision is optimal, it “constitutes unfelt fluency”, likened to exogenous attention.
- Felt fluency (pp. 15–16). Unfelt fluency is “phenomenologically transparent (surprise) until it is rendered opaque by the mental action that attends its recognition, felt fluency”. Mental action is “top-down or endogenous attention”. In a hierarchy, felt fluency is “a hierarchically deeper representation of fluency”: it infers fluent processing below and sends priors on precision down. It is compared to Solms’s felt uncertainty, mediated by ascending neuromodulation.
- The summary claim (pp. 16–18). Felt fluency “supervenes on—or is an attributional inference about—the resulting optimisation of active inference”, experienced as “a precise grip on the world of affordances”. It is both cause (a prior on precision) and consequence (a “prediction error of prediction errors”) of fluent processing. “Changes in unfelt fluency (i.e., precision) induce felt fluency.”
- Why misattribution follows (p. 18). Because felt fluency “arises from internal inferences about the precision of predictive processing, not the sensorium”, it “cannot be attributed to its subpersonal cause—and is therefore plausibly attributed to the sensorium”.
- Affect, in passing (p. 8). Hesp et al. (2021) are cited for formalising emotional valence as a state of self inferred from changes in confidence in one’s generative model, and for proposing “that it would be changes in processing fluency across different domains that … would be at the root of affective states”. Brielmann and Dayan (2022) are cited for a model in which fluency, as likelihood, predicts aesthetic pleasure.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Fluency judgements depend on discrepancy from expected fluency, not on fluency alone | moderate | the reviewed experiments, reported second-hand, pp. 9–13 |
| C2 | Fluency is the brain’s recognition of the precision of its predictions | weak | proposed by analogy between fluency findings and precision-weighting; no model |
| C3 | Unfelt fluency is a subpersonal change in precision that orients attention | weak | a definition offered within the framework, pp. 14–16 |
| C4 | Felt fluency is a hierarchically deeper inference about unfelt fluency, reached by endogenous attention | weak | proposal; cites Sandved-Smith et al. (2021) as formally consistent |
| C5 | Felt fluency is misattributed to the stimulus because its cause is internal | weak | an inference from C4, matched to the attribution findings |
Concepts
- Fluency. The subjective experience of the ease of one’s own processing (the field’s definition, which the paper adopts).
- Absolute / relative fluency. Degree of ease / ease relative to expectation (Whittlesea and Leboe). Westerman’s different sense of “relative” is noted and set aside (footnote 4).
- Unfelt (non-felt) fluency. A non-conscious recognition of a change in attentional set, i.e. in precision. Treated as like exogenous attention.
- Felt fluency. The opaque, reportable recognition of unfelt fluency, made by mental action (endogenous attention); like felt uncertainty in Solms.
- Transparent / opaque. Metzinger’s sense: a transparent state is not represented as a representation, so its source is not given.
- Surprise. Self-information of sensations, explicitly not an emotion.
Connections
The paper joins Whittlesea’s SCAPE model and discrepancy-attribution hypothesis to active inference as in Feldman and Friston (2010), Parr and Friston, and Kanai et al. (2015). Its nearest formal neighbours are Hesp et al. (2021, “Deeply felt affect”) for valence and Sandved-Smith et al. (2021) for meta-awareness, neither held in the record. It shares the predictive-coding scheme that Barrett’s constructed-emotion paper uses (LIT-542, NOTE-434), not the Markov-blanket theorem of LIT-526. James’s “fringe of consciousness” is invoked for unfelt fluency.
Bearing on the record
- QUESTION-059 (Deferred). The LIT was filed as separating “felt from
unfelt fluency inside one active-inference framework”, and as “a
candidate mechanism for what feeling adds to a functional state”. The
reading weakens both.
- Fluency is not affect. The paper treats fluency as a metacognitive feeling of ease. It touches valence only by citing Hesp et al. and the hedonic-fluency literature.
- Felt versus unfelt is access, not phenomenality. “Felt” fluency is the state that endogenous attention makes opaque and reportable (“I lost a fluent grip on this task because my attention was distracted”); unfelt fluency is likened to exogenous attention. That is a difference of functional position in a hierarchy, of exactly the kind QUESTION-059 asks whether a felt state can be reduced to. The paper simply identifies feeling with the higher position and gives no argument that the lower one lacks experience. It even speaks of “a phenomenal sense of non-felt fluency” (p. 11), so its own usage does not keep the two apart. So it is, at most, an instance of the identity side, as THEORY-101 is: feeling is the occupant of a functional role. It is not a mechanism that could tell the two sides apart, and nothing in it would show a regulatory role met without feeling.
- THEORY-045. Barrett’s felt affect is a prediction about the body; Brouillet and Friston’s felt fluency is an inference about the precision of one’s own processing. If both are right, some felt states are about the brain’s own processing and not about the body, which is the dissent from Minsky that THEORY-045 already records. The paper does not claim that fluency is affect, so it does not contradict THEORY-045.
- THEORY-093. Damasio’s feelings are “hybrid”, made partly of the body state. Felt fluency on this account is wholly an inference within the cortical hierarchy, with no body in the loop. Again not a contradiction, since fluency is not offered as a homeostatic feeling.
- Misattribution. Its claim that a feeling with an internal cause is attributed to the most salient external source is the same structure as Barrett’s categorization of affect by context (THEORY-057). The paper does not make the connection.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- No model. Nothing is simulated or derived; “unfelt” and “felt” fluency are not given variables in a generative model. The formal claims rest on cited simulations of attention and meta-awareness.
- No new data. The empirical case is a review, much of it the first author’s own experiments, reported briefly and without effect sizes.
- Terms shift. Fluency is defined as a subjective experience, then extended to an unfelt case; unfelt fluency is “phenomenologically transparent”, yet the gap between it and felt fluency is “phenomenal”.
- Neuromodulation is assigned, not shown. The link of felt fluency to monoaminergic precision control is by citation to Moran et al. and Solms.
- What would refute it is not stated.
Open questions
- Can unfelt and felt fluency be written as two levels of a generative model, with a prediction that the discrepancy-attribution findings do not already fix? A simulation that reproduced the hension effect and predicted a new reversal would make the proposal testable.
- Is “felt” here more than “reportable after endogenous attention”? A case where fluency is attended and reported yet not felt, or felt without being attended, would separate them.