Theory of constructed emotion
Stated in full in Barrett (2017); the renamed successor to the conceptual act theory. See lisa-feldman-barrett.
The wiki met this theory backwards. Barrett has been present since the first ingest as the person who says emotions are not natural kinds, that autonomic-specificity is not there, that no brain region is consistent-and-specific for a category — always the negative claim, always in a paper written by someone else or in a meta-analysis (lindquist-2012-brain-basis-of-emotion) framed as a test of the rival. This page is what the theory says when it is not busy denying something.
The derivation, in order
The theory is presented as a deduction, and it is easier to hold if kept in sequence. Each step is supposed to follow from the one above.
- A brain is for allostasis — anticipating the body’s resource needs and meeting them before they arise. Not for rationality, happiness or accurate perception.
- To regulate a system, you must model it. So the brain runs an internal model of the body in the world.
- The model is built from the body’s perspective, because modelling the world in a detached, disembodied way “would be metabolically reckless.” So the model necessarily includes the internal milieu: interoception is at the core of the internal model, not an addition to it.
- The model is predictive, not reactive. Predictions cascade down; prediction errors ascend and update (predictive-coding, active-inference).
- Predictions are concepts. Completed predictions are categorizations.
- The consequences for allostasis are made available to consciousness as affect — valence and arousal, which are “basic features of consciousness” and not unique to emotion. See core-affect.
- When the concept doing the categorizing is an emotion concept, the categorization is an instance of emotion.
Step 7 is the whole theory. Everything distinctive follows from the fact that it is step seven — emotion arrives last, as a special case of a general-purpose process, rather than first, as a thing with its own machinery.
What it denies, and what it does not
Barrett devotes a numbered list of eight disclaimers to this, which is itself evidence about how the theory gets read. The three that matter most for this wiki:
Emotions are not illusions. The load-bearing analogy is money: the objects that have served as currency throughout human history “share no physical similarities,” yet money is entirely real. Emotion categories are real in that way — real, perceiver-dependent, not physically constituted. Anyone reading constructionism as eliminativism has the wrong theory in view. (Compare Scarantino’s diagnosis that folk emotion categories were coined “to negotiate social transactions among language users” — the two are closer here than their dispute suggests.)
Animals are not emotionless. “‘Is the fly fearful?’ is not a scientific question, but ‘Does a human perceive fear in the fly?’ and ‘Does the fly feel fear?’ can be answered scientifically (and the answers are ‘yes’ and ‘no’).” And explicitly: the fly may still feel affect. See can-we-know-animal-feelings.
Subcortical regions are not irrelevant. An instance of emotion “engages the pattern generators for whatever actions are functional in the context.” The theory needs pattern generators; it denies only that they are organized by emotion category.
Why no emotion category has a signature
Because of degeneracy: dissimilar neural populations give rise to instances of the same category, so “it is unlikely that all instances of an emotion category share a set of core features (i.e. a single facial expression, autonomic pattern or set of neurons).” The summary representation of a category is an abstraction that “need not exist in nature.”
This is Darwin’s move, and Barrett says so: a species is a population of variable individuals grouped by function, not a type with shared features. Emotion categories are conceptual categories in exactly that sense.
The consequence for measurement is the sharpest thing the theory says about method: a pattern that diagnoses sadness across a sample is not the brain state for sadness. It is a statistical summary of highly variable instances, and “the voxels that make up a pattern for a category need not [be] observed in every (or even any) single instance.” The average US family has 3.14 people; no family does.
Relation to the other predictive accounts in this wiki
The theory is not in competition with Seth’s interoceptive inference — it is the same machinery with a different explanatory target and a stronger claim about scope.
| Seth / Seth & Friston | Barrett | |
|---|---|---|
| what interoception is for | inferring the hidden causes of visceral signals; regulating them by active inference | budgeting the body — interoception arises from allostasis |
| the agranular cortex argument | AIC/ACC/SGC/OFC send predictions (visceromotor-areas) | same, extended: the gradient is continuous, so interoceptive cortex predicts exteroceptive cortex too |
| what emotion is | an interoceptive inference | a categorization, of which interoceptive inference is one ingredient |
| scope | emotion and the embodied self | all perception and action, of which emotion is a special case |
They agree on more than either says. Neither cites the other’s central claim as a rival. The real difference is that Seth explains emotion by interoception, while Barrett explains both by allostasis, and then denies that emotion is a separate thing to be explained.
The unresolved cost
Recorded here rather than buried: nothing in the imaging or psychophysiological record can now disconfirm this theory, because degeneracy predicts that instances will not share features, and that is what the data show. This is structurally identical to the objection the wiki records on homeostatic-property-cluster-kinds against Scarantino’s New BET — from the opposite side of the same dispute. Both of the wiki’s most sophisticated positions on emotion categories have made variability a prediction, which means the data that were supposed to adjudicate between them can no longer do it.
Barrett’s response is to reframe the complaint: a new paradigm “barely gets started before it is criticized for not providing all the answers,” and raising more questions than it answers is “a feature, not a bug.” That is true of early paradigms in general and is not an answer about this one. What would count against the theory of constructed emotion is not stated in the source.
The theory extended to social motivation: guilt as a category (Theriault et al. 2021)
Theriault, Young & Barrett (2021) apply the theory to social pressure, and the application sharpens what constructed emotion says about a discrete category. They model the sense-of-should — a felt obligation to conform to others’ expectations — and show it is formally equivalent to the economic guilt-aversion model of Chang et al. (2011): both make an agent minimize the discrepancy between behaviour and the inferred expectation of another. Where they diverge is on what “guilt” is.
On the constructed-emotion reading, “guilt” is not a state with a signature but a category of heterogeneous instances — the physiological change, the behaviour, the affect all vary by situation — unified only by the sense of should, i.e. by the motivation to conform to others’ expectations. Guilt is therefore one instance under the sense of should, not synonymous with it (the more general motive shows up in non-guilt instances too — a servile lackey, a parent caving to a child). This is degeneracy and population-thinking applied to a social emotion: same function (conform to expectation), physically dissimilar instances, no core feature. It is the wiki’s first case of the theory doing work on a specific named emotion in the social domain, and it reproduces the theory’s structure exactly — the category is held together by function and collective agreement, not by a shared bodily state.
The theory extended down the lifespan: change the ingredient, change the emotion (MacCormack et al. 2021)
MacCormack, Henry, Davis, Oosterwijk & Lindquist (2021), from Lindquist’s lab, run the theory developmentally, and the move is a clean test of step 3–5 of the derivation above. If interoception is one ingredient the brain categorizes to build an emotion, then a relative shift in the amount or quality of afferent bodily information should alter emotional experience — not by breaking the machinery but by feeding it a quieter body. Aging supplies exactly that manipulation (peripheral-nerve demyelination, reduced autonomic reactivity), so the theory predicts emotion should become less embodied and more conceptual/situational with age. That is what they find: the emotion–body association weakens (for high-arousal sensations specifically) while the emotion–situation association does not, and in daily life interoception decouples from situations but not behaviors. See maturational-dualism.
Two things this contributes to the theory as stated on this page. First, it puts the constructionist claim that emotion concepts are dynamic — updated by experience across the lifespan, not fixed after childhood (Doyle & Lindquist 2018) — to work: the interoceptive content of an emotion category shifts with age. Second, it is a rare case where constructionism makes a directional prediction that could have failed — if aging were general cognitive decline, situational and behavioral emotion content should have faded too, and they did not. The interoceptive-specificity is what a “change the bodily ingredient” account predicts and a “general aging” account does not. It does not escape the falsifiability worry above (the result is also consistent with several non-constructionist stories, and is self-report, not physiology), but it is the theory generating a specific, modality-selective prediction rather than absorbing whatever the data show.
The theory tested by individual differences, and the ingredients do not come apart (Ventura-Bort et al. 2021)
Ventura-Bort, Wendt & Weymar (2021) attempt the obvious individual-differences test of steps 3–6 above: if core-affect and conceptualization are separate ingredients, then measures nominated as correlates of each should separate, and each should moderate emotional experience in its own way. Interoceptive-sensibility questionnaires stand in for core affect; emotion-clarity and alexithymia questionnaires stand in for conceptualization.
The ingredients did not separate. A PCA over eight instruments split them into two orthogonal factors that each contain both bodily and emotional questionnaires: Sensibility (believed accuracy about one’s internal states, whether visceral or affective) and Monitoring (how much attention is deployed inward). Body and emotion instruments were interleaved throughout. The authors state it plainly: “the factor scores did not differentiate between interoceptive sensibility and emotional conceptualization scales.”
The theory absorbs it, and that is the point worth recording. Their resolution is that Sensibility indexes conceptualization and Monitoring indexes attention — which is legitimately one of TCE’s four components (Barrett 2017), so nothing has been refuted. But look at the structure of the move: a result that failed to recover two of the theory’s ingredients was accommodated by reassigning the factors to a different pair of its ingredients, post hoc. The theory has four components and the data had two factors; a mapping was available.
This is the falsifiability worry recorded two sections up, arriving in a new form. The earlier version was about neural data — degeneracy predicts variability, so variability cannot count against the theory. This version is about psychometrics: the theory names ingredients but not the instruments that index them, so any factor structure can be mapped onto some subset of the four. Barrett’s own reply (“raising more questions than it answers is a feature, not a bug”) does not touch it.
Two things the study does supply that are on the theory’s side. Its positive findings are constructionist in shape: believed accuracy about internal states predicted finer granularity for negative emotions, which is what “better conceptual grip on internal states yields finer categories” (situated-conceptualization; Lindquist & Barrett 2014) predicts. And the whole result is a demonstration that emotional and bodily self-knowledge are not psychometrically separable at trait level — a prediction constructionism makes and a discrete-emotion account has no reason to make.
Held with the study’s own limits: no interoceptive accuracy measured, no physiology, a forced two-factor solution, and its laboratory task (emotion-differentiation-task) found nothing at all.
The theory adopted as clinical consensus (Khalsa et al. 2018)
The wiki met this theory as Barrett’s own; the Khalsa et al. (2018) roadmap shows it absorbed into the field’s clinical mainstream. The roadmap’s psychopathology figure reproduces the EPIC model (Barrett & Simmons 2015) — the predictive-coding implementation of constructed emotion at the level of insular cytoarchitecture — and states its illness hypothesis: interoceptive input (posteriors) “becomes increasingly decoupled from interoceptive predictions issued by the agranular visceromotor cortex (priors), leading to increased interoceptive prediction error signals,” which “may present in the brain as ‘noisy afferent interoceptive inputs’” (Paulus & Stein 2010). Two things are worth recording. First, the constructed-emotion machinery here is doing clinical work — a prediction/precision failure as a mechanism of anxiety and depression — which is a use Barrett’s own paper gestures at but does not develop; see computational-psychiatry, interoceptive-psychopathology. Second, it shows the theory’s anatomy (agranular predicts, granular errors) treated as consensus infrastructure by a 24-author committee, even as the falsifiability worry above remains unanswered — the framework is being built on before it has been decisively tested, which the roadmap concedes (“indirect so far”).