What are metacognitive (also called noetic, or epistemic) feelings? Juxtaposing [being metacognitive] and [being a feeling] sounds, at least prima facie, dangerously close to an oxymoron. When Descartes, Locke, and other 17th-century philosophers explored the properties of ideas as being “clear”, “distinct”, “evident”, and “certain” they certainly never took them to be feelings. These notions were taken, rather, to be objective representational properties that the mind, unaided by imagination, is able to detect. David Hume, in contrast, observed in his Treatise that “the vivacity of the idea gives pleasure”, and that “its certainty prevents uneasiness by fixing one particular idea in the mind, and keeping it from wavering in the mind of its objects” (Hume 1739/40, 2007, p. 289). Thus Hume was glad to accept that epistemic feelings exist, and that they vary in their vivacity and in their pleasantness, i.e., in their intensity and in their valence. Following Hume’s lead, let us test how our analysis of FS above fares with the case of noetic feelings. Here, again, is our proposal about the general structure of feelings.
FS Affordancea [Placea=here], [Timea=now/soon], [Valencea=+/-], [Intensitya=.n(comparatively specified on a scale 0 to 1)], [motivation to acta of degreed according to action programa].
What is specific to noetic feelings is that the affordances to which the system resonates are “informational” or “metacognitive” rather than environmental. Hence, the affordance does not relate to the external environment (the “here” slot is often irrelevant, except for perceptual affordances, or place-dependent metacognitive affordances, such as concentrating in a noisy spot). Although a cognitive action does not, in general, consist in physical moves towards or away from an affordance, similar decisions are motivated or inhibited in the domain of mental agency: a high retrieval affordance motivates pursuing the memory search, a low one to quit, etc. Hence our FS analysis also applies to noetic feelings.
As already emphasized, the affordances expressed in feelings do not need to be construed conceptually in order to be detected and assessed through their associated somatic markers. A conceptual construal, however, is suggested by the names given, in the literature and in ordinary language, to M-feelings. The term “feeling of knowing” (in response, for example, to the question: “what is the capital of Australia?”) implicitly presupposes that the emoter has access to the concept of knowledge. Expressing her feeling verbally, indeed, an emoter might say: “I feel that I know the response to this question”. In this sentence, she indeed refers to her disposition to retrieve knowledge and, hence, metarepresents her knowledge disposition.[26] The affordance theory of noetic feelings suggests a different picture. When trying to remember a proper name, a feeling of knowing is a specific experience of having the ability to detect the target, and of predicting its imminent recall. It can be associated with a feeling of tension (Koriat & Levy-Sadot 1999, p. 486). This experience is associated, then, with a graded, intuitive, and affect-like appraisal of a [remembering] affordance. Rhesus monkeys working in experimental labs in comparative psychology show that they can assess their memory affordances (see Beran et al. 2012, Chapter 1).[27] What kind of feedback, then, do monkeys use? A surprising and substantive fact about metacognitive control, first revealed through the pioneering research of Asher Koriat, is that the comparator generating metacognitive feelings (such as a feeling of knowing in a memory task, or a feeling of clearly discriminating in a discrimination task) has no access to the semantic contents stored in memory or made available through perception. In Koriat’s words, M-feelings “are mediated by the implicit application of non-analytic heuristics, relying on a variety of cues.” These cues “pertain to global, structural aspects of the processing of information”, such as ease of processing, time devoted to a task, familiarity, and accessibility (Koriat 2000; Koriat & Levy-Sadot 1999).[28] Therefore, contrary to what epistemologists have always believed, the most common type of epistemic appraisal is not directly based on the content of the thoughts to be evaluated, but on the properties of the underlying informational process.
Neuroscientific research confirms Koriat’s claim. Implicit, associative cues are extracted by the working brain to select, in a cost-efficient way, what there is to learn, to retrieve from memory, to extract from perception, or what is worth storing in memory. These are all to do with the dynamics of information processing: with its onset, with the comparative amount of activity in incompatible neural responses, and with the time needed to converge on a threshold value. Indeed, the neural activity recorded in rats’ OFC when attempting to categorize olfactory stimuli was found to correlate with their predictive behavior (consisting in accepting or rejecting a task trial); similar patterns have been found in other species.[29]
On the FS model, somatic markers have the function of expressing the intensity and valence of the noetic predictions generated from feedback at the neural level. As indicated in section 1, psychophysiological measures (electromyography) provide evidence for the existence of facial markers associated with feelings of fluency and of disfluency (Winkielman & Cacioppo 2001). Increased activity in the smile muscle, the zygomaticus major, produces feelings with a positive valence. A reduction of fluency is correlated with activity in the corrugator supercilii (involved in frowning), which suggests that this additional effort is felt as unpleasant. Intensity of positive or negative confidence, computed implicitly, is expressed by the corresponding intensity of the noetic feeling. A different somatic marker of memory appraisal is the TOT phenomenon. This often occurs when a search in memory for a specific word fails to retrieve that word within the usual time interval. The informational ingredients of FS are conveyed by the intensity of the activity in the tongue muscle, and by the affective quality of TOT. Taken together, these predict the likelihood of successful retrieval. An implicit cue-based heuristic might thus explain why TOTs have the valid predictive value they do (Schwartz et al. 2000).
In our FS single-factor model, M-feelings have an intrinsic intensity and an intrinsic valence. Two-factor theories make a different claim, in ways analogous to Schachter and Singer’s theory of aboutness in affects: M-feelings have an intrinsic arousal level, but their valence depends on the environment. Jacoby and his colleagues were the first to embrace a two-factor view about feelings of fluency. They manipulated participant’s exposure to an item in order to show that enhanced fluency generates an illusory feeling of familiarity. Under conditions of divided attention, reading a list containing both famous and not famous names raised participants’ disposition to wrongly judge as famous some names presented in a second list, merely because these names had already been read in the first list. Schachter and Singer’s idea was that fluency is a generic feeling, that needs to be interpreted on the basis of goals and current cues, in order to deliver a qualitatively different specific feeling:
Inherent in the idea that the subjective experience of familiarity arises from an interpretation of cues is the notion that cues can be interpreted in a variety of ways. As noted above, if ease of identifying an item is obviously being manipulated by the experimenter, the resulting perceptual fluency does not give rise to a feeling of familiarity. Attributions are also affected by one’s goals. In the context of attempts to remember, people may be more likely to interpret ease of generating an item or perceiving it as familiarity. In the context of other tasks, the same cues may be interpreted in other ways. (Kelley & Jacoby 1998, p. 129)
From their viewpoint, the fluency generated by a given name can, according to the task and the information made consciously available to a participant, be experienced as a feeling of familiarity, or as a feeling of recognition of that name as “old” (i.e., presented in a former list). They conclude that a feeling of fluency (generated by a perceived name) will be experienced as a function of the alternative ways of interpreting this feeling, on the basis of the agent’s goals and the additional cues available.[30]
A similar two-factor theory has been defended in the (Whittlesea & Williams 2000; Whittlesea & Williams 2001) model of M-feelings. According to this model, feelings of familiarity result from the perception of a nonspecific discrepancy between the expected and the observed rate of processing of elements in a given context. Valence and the associated action guidance, on the other hand, are based on a conceptual interpretation of what this discrepancy means. For example, you find yourself waiting for the bus next to people you expect to be total strangers. Suddenly, you have an unexpectedly high fluency experience when looking at the face of someone you have already encountered several times—a clerk from the local grocery shop. This unexpectedly high rate of discrepancy-reduction determines an intense feeling of familiarity with a strong motivation to identify the familiar face (see Whittlesea & Williams 2001). Had you seen the clerk in the local grocery store instead, you would have merely had a feeling of recognition when seeing the clerk.
To summarize: the core idea in two-factor accounts is that participants have a primary feeling of fluency, which they interpret in more specific terms as a function of their goals and of the context as they consciously represent it to be. Thus, on this view, a feeling partly relies on background knowledge, and partly on a naïve theory concerning the relation between feelings and mental activity (Schwarz & Clore 2007). The naïve theory is as follows: feelings are about what one is doing, so this feeling must be about this event of trying to perceive, or this attempt at retrieving, etc.
As already observed above, a naïve-theory view is incompatible with monkeys’ and young children’s epistemic evaluations based on fluency. Our FS structure offers an alternative account: cues (associative heuristics) dictate how an affordance is detected, assessed, and exploited in a context, but these cues are not consciously available, and hence do not depend on a naïve theory of the task. The Jacoby and Whitehouse evidence is compatible with a procedural view of engagement in a task through automatic memory processes, and of the feelings of familiarity they generate. A comparator is always activated as a function of a subject having been highly trained in the corresponding first-level cognitive task. Monkeys and humans feel that a memorial or perceptual affordance is present because, if they need to assess whether, for example, an item was seen earlier, the associated comparator produces a feeling of a given intensity and valence indexing the remembering affordance. Thus, it is uncontroversial that a context-dependent factor determines both the task to be performed and the reactive metacognitive feeling about this task.
It does not follow from the context-dependence of a cognitive task, however, that a concept-based interpretation will affect the experienced feeling itself, as maintained by the two-factor theorist. A cue-based, non-analytic heuristic is not inferential in the interpretive, first-person sense. Regrettably, the word “inference” has been loosely used in affective and in metacognitive studies, to refer both to “automatic, non-analytic, largely unconscious and fast associative processes” (Nussinson & Koriat 2008) and to conscious reasoning and theory-building (Schwarz & Clore 2007). These two types of processes (respectively called “automatic” and “controlled”), are now held by many authors to operate independently.[31] While unconscious heuristics rely on implicit associations between cues, inferences comprise deductions from premises to conclusions. Looking back at Jacoby and Kelley’s point above, we see that the authors are referring to unconscious cues being recruited for a task: they are thus referring to unconscious associative heuristics rather than to explicit concept-based reasoning. The memory interactions they are exploring, however, typically involve both automatic and controlled processes, which is a source of confusion. As Jacoby and Kelley are eager to show, implicit associations and explicit reasoning lead to different, incompatible predictions. As a result, the evidence they present shows how automatically-generated feelings can be theorized about in controlled processes. It does not demonstrate, however, that feelings depend upon theorization. A theory of the task, in contrast with automatically generated feelings, offers reasons to attribute to oneself beliefs and motivations to act, and, possibly, to reject the relevance of feelings for any particular task.
Our proposal, then, has several advantages over inferential or theory-based accounts of f-aboutness. First, it explains why a feeling of fluency can be experienced, and why it can motivate agents’ metacognitive responses in species or individuals with no concept-based attributive capacity (i.e., with no capacity for mindreading). Second, our proposal accounts for the difference between a type of M-feeling (a feeling of fluency) and the various ways in which it is experienced across cognitive tasks. Granting that comparative ease of processing can always be computed, and can be used as a reliable indicator of the likelihood of success across a wide range of cognitive activities, it is not surprising that there is a type of feeling based upon it. Fluency can be perceptual, memorial (“retrieval fluency”), or conceptual. It can be used in predictive or retrospective evaluations. If agents are asked to determine which statements are likely to be true or false (presumably a question that only—but not all—humans can understand), felt perceptual fluency will induce a “truth effect”. Agents will evaluate a statement as more likely to be true than another merely because it is easier to read.[32] If agents are asked to detect faces of known people (or of stimuli previously shown), felt fluency will generate a sense of familiarity, which motivates agents to try to identify the target. If people are asked to assess the frequency of a given phenomenon, felt retrieval fluency—that is, what comes immediately to mind—will be used to judge what is more frequent. Felt fluency will also have effects outside of metacognition: if agents are asked which particular face, landscape, or picture they prefer, felt fluency will influence their decision. Several affordances, then, may be associated with the same globally expressive type of feeling (constructed as the set of feelings with the same type of facial markers for ease of processing, for example). The notion of type of feeling is a technical term, which is useful to distinguish the diverse ways in which fluency is used by the brain. But a type of feeling is never experienced; only tokens of the type are. Tokens of feelings of the same type will differ in the specific affordances that are detected, and in the tendencies to act that the feeling motivates. As a consequence, one cannot say that feelings of fluency “feel the same” to an emoter: fluency experienced in an FOK and in an FOR, for example, apply to different segments of processing, assess different things, and motivate a different action program. You may first have an FOK after a question is addressed to you, and then fail to have the associated FOR after having come up with a response. These differences have nothing to do with an interpretation: they are constitutive of what sensitivity to a given affordance amounts to. Take the case of feelings of familiarity. As summarized above, Whittlesea and Williams claim that fluency is the core of the experience, while familiarity is a conceptual interpretation of this core feeling. It is more economical, however, to suppose that familiarity is a different feeling within the general fluency type, and that it is associated with a different affordance.
In summary: engaging in a particular cognitive task (e.g., trying to remember, evaluating retrieval, assessing frequency) does not need, per se, to involve a naïve theory of the task. It only requires having a salient affordance, and an implicit heuristic for metacognitive predictions in that task.
Our proposal also allows us to address in affective terms the issue of incidental versus integral feelings, which, in the literature, is invariably framed in inferential terms (with all the ambiguity relating to this expression). Metacognive feelings are called “incidental” when they are not based on valid cues for the cognitive task at hand, and hence, have no predictive value. They are called “integral” when they actually carry information about cognitive outcome. Granting the universal role of fluency in metacognition, how do people know when a feeling of fluency is relevant to a given task, and which sequence of their cognitive activity needs to be monitored? A frequent answer, in the literature, is that agents believe that fluency applies by default to the present domain of judgment. When, however, agents are led to believe that a feeling of fluency is purely incidental to the task at hand, they will discount it in their decision, on the basis of a theory of the domain of interest (see Schwarz & Clore 2007 and Whittlesea & Williams 2000, 2001). Let us suppose, in what we shall call case (a), that an agent is explicitly told that a given cue, such as the ease of reading a given sentence, is irrelevant to a given task—such as assessing the truth value of the written statement. Or, alternatively, let us suppose—case (b)—that the agent discovers by himself that there is a connection, but with reverse relevance. Perhaps he finds that badly written sentences, involving added processing effort—in a given context—are likely to be true (see Unkelbach 2007 and Unkelbach & Greifeneder 2013). A popular account of these cases is that people will infer respectively, for (a): that the feeling of fluent reading they have had is not about the target task, which entails that reading fluency does not predict truth, or, for (b): that what predicts the truth of a written utterance, in this particular context, is disfluent reading (see Schwarz & Clore 2007, p. 394).
According to this two-factor account, M-feelings are cognitively penetrable. They can be suppressed at will, on the basis of a reinterpretation of their being experienced, or can even be used to predict falsity instead of truth.[33] On the account proposed here, in contrast, M-feelings are never cognitively penetrable. Why, then, do subjects stop trusting their feeling of fluency? Our answer is the following. In the first type of case, subjects do not allow their feelings of fluency to guide their decision because they have received verbal instructions to this effect. In the second type of case, subjects no longer use their feelings of fluency to form an epistemic decision in the proposed task, because they have learned, over time, that these feelings do not predict truth in this task.
In case (a), then, subjects are confronted with a different task. They are no longer asked to express their confidence in the truth of a given sentence (an intuitive, associative task); they are asked to assess the truth of sentences by taking into account the fact that their feelings of fluency are irrelevant. This new task requires the participants to form appraisals based on analytic reasoning. Feelings no longer drive their evaluation and epistemic decision.
In case (b), where bad writing is associated with likely truth, no “theory of the task” needs to be formed, on top of the first-order task, which consists in judging whether a written statement is true or not. A mere change in cue validity can produce, over time, a change in associative heuristics, and, hence, in feelings and in decisions to act. For example, just as our thirsty traveller will eventually learn not to trust an apparent “drinking affordance”, an agent will learn, in certain recurrent contexts, not to trust an apparent “fluency affordance”. Obviously, cue validity can, in humans, be conveyed verbally; this will considerably abridge the revision process of the associated program of action. We then return to case (a): participants will be able to immediately discount an apparently valid cue, to turn to analytic appraisals, and to refrain from acting on their fluent feeling (which, however, is still there). Cue validity, however, can be learnt implicitly as well, which weakens the case for a theory-laden view of feelings.
These observations suggest that feeling-based and analytic appraisal, as hypothesized in this proposal, “tap separate databases representing knowledge in different formats.”[34] A feeling of fluency, as a result, can survive being discounted in decision-making. Another finding points in the same direction. There is evidence that, even when an M-feeling has been explicitly discounted (i.e., shown to agents to unduly bias their epistemic assessment), the initial feeling remains unaffected, and is able to promote further epistemic decisions. In Nussinson & Koriat’s (2008) study, agents exposed to unsolved anagrams and to anagrams accompanied by their solution, were asked to rate the difficulty of these anagrams for naïve participants with no prior access to the solution. The participants’ ratings were influenced by the differential fluency that the anagrams presented for them: the higher fluency of solved anagrams biased their attributions of difficulty. After being informed of the contaminating effect of knowing the solutions, the participants were invited to correct their attributions by re-rating the difficulty of the anagrams, which they did. However, the participants were subjected to a subsequent test, where, under time pressure, they had to predict which of two anagrams would be harder for others to solve. These other-attributions of difficulty presented, again, the same bias for known anagrams. Being under pressure allowed participants’ M-feelings to guide decision. The verbal instruction could shift their controlled responses when re-rating the anagrams, but did not lead the participants to recompute them, as should have been the case if feelings are cognitively penetrable.
In summary: what participants learned (that solved anagrams only look easier to process) did not influence what they felt later (higher fluency is diagnostic of ease of solving).