Paul Griffiths has claimed that emotions do not constitute a natural kind, in the sense that they do not form “a category about which we can make inductive scientific discoveries” (2004, pp. 901–911). One can agree with latter claim, however, without concluding that feelings do not constitute a natural kind. First, feelings are not only affective ingredients in emotional awareness. Some feelings, such as feeling cold or sick, or feeling that one is acting, have nothing to do with affective episodes. Second, there are evolutionary reasons to distinguish, within emotions, two classes of subjective appraisals. Emotion theorists usually contrast feelings expressed in primary emotions—fear, anger, happiness, sadness, surprise, and disgust—with various appraisals cum conative dispositions associated with higher cognitive emotions, such as envy, guilt, pride, shame, loyalty, vengefulness, and regret. The first are phylogenetically and ontogenetically prior to cognitions. They belong to the ancient limbic system, which is present in some form in most animals. A quick route from the retinal image to the amygdala through the thalamus allows affective information to control behavior (see LeDoux 1996). Primary feelings are thus triggered independently of concept possession and motivate specific responses. Secondary affective experiences, in contrast, might have evolved on the basis of social constraints in relation to cooperative action among humans. Indeed (with the possible exception of pride and shame) they are not present in nonhuman primates.[4] They activate newer brain structures; they require concept possession, depend on background beliefs, and do not generate characteristic behaviors. Finally, primary feelings are clearly embodied, while secondary emotions seem to have no proprietary somatic markers. An interesting idea, suggested by Jesse Prinz (2004, p. 95), is that the facial or somatic correlate of secondary emotions, when they have one, involves a blend of the somatic markers for primary feelings.
In summary: emotions differ, among other things, because of the unequal role that feelings have in the two classes of emotions just discussed. The wider scope of feelings, when understood as “reactive, subjective experiences with a distinctive embodied phenomenal quality”, seems to be more unified than emotions, and making feelings seem like plausible candidates for a natural kind.
We need, however, to turn this tentative definition into a general functional characterization that presumably holds for all feelings (beyond affective ones) and only for them. Here is a proposal: feelings constitute the sensitive part of predictive and retrospective processes of non-conceptual evaluation of one’s own and others’ well-being and actions. Being essentially evaluative, feelings are always the output of a comparator: in other terms, they are crucial monitoring ingredients in self-regulated adaptive control systems. In such systems, the specific function of a feeling consists in detecting how much a current observed value of a parameter deviates from its expected value, on one or several dimensions relevant to survival (see Carver & Scheier 2001). Their formal object, when they have one,[5] (such as being afraid of the bear in front of me) cannot be analyzed independently of the monitoring function they serve within a specialized control loop.[6] Relevance to well-being, however, extends to bodily condition, goal achievement, and availability of preferred goods of all kinds (food, partner, social status). The relevant dimensions of variation that feelings track may accordingly be of a sensory, proprioceptive kind (feeling thirsty, cold, etc.), social-affective (feeling angry), or agentive (goal-related). Goal achievement, however, involves either epistemic or instrumental success, respectively generating epistemic feelings (feeling interested, bored, epistemically uncertain) and agentive feelings (feeling of happiness, of agentive confidence, of ownership of one’s action, etc.). Feelings, in summary, are the outcomes of comparators in a control loop; they carry non-conceptual information about how much one’s present condition deviates from the expected condition. From a functional viewpoint, they form a natural kind insofar as their function is to indicate a comparative outcome through a dedicated embodied experience.
Note, however, that there are comparators that trigger no feelings at all: these non-sensitive comparators may either work outside consciousness (for example, error signals driving immediate correction[7], not to mention comparators that work at the cell level), or they can take concepts as their input, rather than reacting to percepts or situations (for example comparators of currency or of educational value).
As far as feelings are concerned, they are directly related to a presently-perceived context (or an imagined or remembered context, but in a “present-like”, indexical mode): one can feel too hot, too cold, or too tired (or feel “OK”, which usually means a tolerable deviation from the expected value). One can feel the fright one has had, even after the frightening event has ended. The outcome of a feeling-based appraisal, from a functional viewpoint, has to consist in some disposition to act that is adaptive, relative to the input to which the feeling is a reaction. Granting that feelings, as sensitive comparators in a control system, form a natural kind, there should be common properties cutting across the various types listed above. In fact we find three types of functional relations between feelings of a given kind and the associated disposition to act. First, feelings, according to their embodied valence, typically determine actions of approach or of avoidance. Some dictate caution, others boldness. Some encourage self-restraint, others self-assertion. Fear promotes a flight tendency, hunger a tendency to approach food. Second, they have a specific orientation in time: some feelings have a predictive function, and thus induce a behavior that is based on contingencies to be further displayed in the present context. For example, fear, when directed at a possible danger, increases the readiness to flee in case the danger concretizes. Others have a retrospective function, and induce corrections to the commands one has previously used, or to one’s previous preferences. For example, feeling nauseous after food ingestion induces food avoidance, i.e., a change in the agent’s preferences. In contrast, feeling disgust at the sight of some food may prevent the agent from approaching it. A subset of feelings, such as feeling happy, have both temporal orientations. Third, according to their embodied dynamics and intensity (which is called their “level of arousal”), feelings can provoke an elevation in the energy available to the system: they provoke excitement, agitation, power in the coming response; or, on the contrary, they may have a soothing effect and diminish the tendency to act.
One major functional property of feelings, from the viewpoint of information extraction and use, is that they can very rapidly extract and synthesize multiple cues from perception. This rapidity is a consequence of the automatic and encapsulated character of the control mechanism whose output they express. Feelings are automatically triggered by a specific type of input (which is the definition of informational encapsulation).[8] Automaticity is associated with feelings being inescapable, at least for those feelings that have been allowed to develop within a culture, granting normal development.[9] The mechanism that generates somatic, noetic, or affective feelings from inputs (perceptual, imaginative, or memorial) does not require one to have specific beliefs or intentions.[10] Informational encapsulation explains why transitive feelings persist when the agent finds out that the situation is different from what she thought to be the case. Just as an optical illusion such as the Müller-Lyer effect does not immediately dissipate when it turns out that the segments are equal, a feeling of anger does not disappear as soon as the agent realizes that its formal object is not exemplified.
Automaticity and informational encapsulation seem also to characterize agentive feelings (see Pacherie 2008). Feelings generated in the course of a physical action come in two varieties: generalized or specialized. Some, such as feelings of agency, of initiation of action, of ownership and of motor control, are indicators monitoring action in progress: they concern “who” is performing the action, and “how” the action is being conducted (see Proust 2000). Others concern the evaluation of an action in one’s own repertoire: a professional carpenter or an experienced musician, for example, have feelings telling them if an action sequence (whether their own or another agent’s) in this repertoire sounds or looks right, even before they identify why they have this feeling. These feelings are also the outputs of a comparison between motor anticipations and observed properties of the action (a “forward model of action” supposedly stores the expected values of crucial parameters; Wolpert et al. 2001). They can predict the likelihood of completing an action (when the question arises, in difficult or non-routine cases), or evaluate—on-line or in retrospect—how swiftly, effortlessly, or unhesitatingly an action was performed. Agentive feelings thus have an essential role in regulating the fundamental properties of physical actions, such as the quality of the outcome,[11] and the ownership of the action.[12]
Noetic feelings, finally, are functionally similar to somatic, affective, and agentive feelings—although their evolutionary pattern seems to be different from the other three kinds. While most organisms have proprioceptive, affective, and motor control, and hence, presumably, somatic, affective, and agentive feelings, few are able to control their cognitive decisions through metacognitive feelings (see Beran et al. 2012 and Proust 2013). The latter are generated when trying to perceive, to remember, or to plan a cognitive task (in particular, when trying to plan how long to study material in order to master it).[13] They are also relied upon when trying to reason or to solve a problem; when conversing, feelings of effort, and of informativeness, are monitored by speakers and hearers in order to maintain a common level of relevance. Like other feelings, they have two distinctive temporal orientations. Some have a predictive function. A feeling of knowing (FOK) may arise when trying to remember an item—for example a proper name—that one has not yet retrieved: having a strong FOK reliably predicts that one will finally retrieve the searched content (Koriat & Levy-Sadot 2001). A feeling of having a name on the tip of one’s tongue (TOT) both signals the fact that a word is not presently available, and, according to its onset, valence, and intensity, whether it is worth or not worth pursuing one’s effort to retrieve it (see Brown 1991 and Schwartz et al. 2000). Feelings of fluency are the sense of ease of processing one may feel or fail to feel when attempting to perceptually discriminate objects with a given property, or to retrieve items from episodic or semantic memory. A feeling of familiarity is particularly salient, in human adults, when no further fact about the target can be retrieved. It offers useful information about the epistemic status of the target: that it is not new, but nevertheless not fully recognized. A feeling of familiarity, then, motivates, among others, an attempt to recognize what or who a target is. Other metacognitive feelings have a retrospective function. When a name is retrieved, a feeling of rightness (FOR) motivates the agent to consider her response the expected one.[14] Various feelings of uncertainty, based on fluency, coherence, plausibility, informativeness, or relevance, also have retrospective functions: their valence and intensity tell the agent whether she should accept or reject a cognitive outcome. These parameters are expressed through specialized somatic markers, such as increased activity in the facial muscle involved in smiling, the zygomaticus major—for positive valence—or the corrugator supercilli (involved in frowning)—for negative valence (Winkielman & Cacioppo 2001).
Taken together, these considerations are compatible with the view that somatic, agentive, metacognitive, and “primary” affective feelings, even if they differ in their formal objects, form a natural kind. Our attempt above at a functional characterization focused on the general relations of feelings to inputs, outputs, and mediating evaluative mechanisms. From this characterization, it emerges that feelings are gradients in comparators that are felt subjectively, rather than being propositional states describable in analytic, objective terms. These observations, however, suggest that, in order to express a specialized and fine-tuned reactivity to one or several formal objects, and to motivate adapted behaviors, in order to be remembered and conveyed to others feelings must have their own representational format. We now turn to the following question: What is the structure of the information that is extracted and expressed in a feeling?