The theory of constructed emotion (Barrett 2017)
Barrett’s flagship theoretical statement, published in Social Cognitive and Affective Neuroscience alongside the trade book How Emotions Are Made. This is the paper the wiki has been circling for eight ingests: Barrett has appeared here since the first ingest as a critic — of autonomic-specificity, of natural kinds, of locationism — always in someone else’s paper, always in the destructive half of her programme. This is the constructive half, in her own words.
It is also the paper that closes a loop the wiki noticed and could not source. The Seth & Friston (2016) page ends by observing that the VMAs of predictive-coding neuroanatomy are “very close to the core affect network Lindquist’s meta-analysis identifies empirically — a convergence between a predictive-coding neuroanatomical argument and a psychological-constructionist meta-analytic one, made independently and from different starting assumptions.” Here Barrett makes that convergence herself, deliberately, and it is the architecture of the theory rather than a coincidence between two literatures.
The methodological inversion
The article’s organising move is stated in the abstract and is worth taking seriously as a claim about how to do science, independent of whether the resulting theory is right.
The field’s standard approach is inductive: begin with emotion concepts “that are most recognizably English” — anger, sadness, fear, disgust — and search for their biological essences. This “assumes that the emotion categories we experience and perceive as distinct must also be distinct in nature.” Barrett’s objection is not that the search has failed (though she argues it has, at length, in Table 1) but that the premise is a category error, and she reaches for Einstein to make the point: physical concepts “are free creations of the human mind.”
So the paper turns it around and proceeds deductively: start with what a brain is and what it is for, and derive what emotion could be given that architecture. Whether or not one accepts the conclusion, this is why the article spends its first third on allostasis and cortical lamination and does not mention an emotion category until page nine.
What a brain is for: allostasis, and why interoception falls out of it
The pivot of the whole theory, and the part with the most consequence for this wiki.
A brain did not evolve for rationality, happiness or accurate perception.
It evolved to run a body: to “efficiently ensure resources for physiological systems within an animal’s body… so that an animal can grow, survive and reproduce.” That is allostasis — not a condition of the body but a process for how the brain regulates the body according to costs and benefits, where efficiency requires anticipating needs before they arise.
From this single premise three things follow in order, and the order is the argument:
- To regulate a system you must model it (Conant & Ashby’s cybernetic principle, cited in a footnote). So the brain runs an internal model of the body in the world.
- Modelling the world “accurately” in some detached, disembodied manner would be metabolically reckless. The model is built from the perspective of the body’s physiological needs — so it necessarily includes the statistical regularities of the internal milieu. That is interoception, and it is not an add-on: it is “at the core of the brain’s internal model and arises from the process of allostasis.”
- Interoceptive sensations are experienced as lower-dimensional feelings of affect. Valence and arousal are therefore “basic features of consciousness” — and, the sentence the wiki should hold onto, they “are not unique to instances of emotion.”
This is the deepest available answer in the wiki to why the body is in emotion at all. Craig’s answer is anatomical (there is an afferent pathway, and it terminates in the insula). Seth’s answer is computational (the body is one of the hidden causes the brain must infer). Barrett’s is teleological and prior to both: interoception exists because a brain that must budget a body has to model it, and everything experiential is downstream of the budget. See allostasis, where this is now the page’s central claim rather than a Farb-derived footnote.
Predictions are concepts
The second half of the title. Having established a predictive brain, Barrett makes an identification the other predictive-coding sources in this wiki do not:
Predictions are concepts. Completed predictions are categorizations.
A concept, in her usage, is not a stored definition — it is “a group of distributed ‘patterns’ of activity across some population of neurons,” a population of competing predictions each with a prior, assembled on the fly (Barsalou’s ad hoc concepts). Prediction error selects among them. When the winning population is an emotion concept, the resulting categorization is an instance of emotion.
The consequence is the strong claim of the paper: an instance of emotion “is constructed the same way that all other perceptions are constructed, using the same well-validated neuroanatomical principles for information flow within the brain.” There is no emotion mechanism to find, because there is no emotion mechanism — there is one mechanism, doing what it always does, with emotion concepts loaded. See situated-conceptualization.
A thought experiment carries it: you have been happy lying in the sun, finishing a workout, hugging a friend, eating chocolate, winning a competition. Each instance differs from every other. The brain constructs happiness “not in absolute terms, but with reference to a particular goal in the situation.” So “‘happiness’ has a specific meaning, but its specific meaning changes from one instance to the next.” The category is held together by goal, not by feature — which is Darwin’s move on species, applied to emotion.
The anatomical argument, and its relation to Seth & Friston
Barrett’s cytoarchitecture is the same as Seth & Friston’s, sourced to the same tradition (Barbas), and reaches further.
Shared: agranular limbic cortices (ACC, ventral anterior insula) lack a developed layer IV, are structurally unable to receive cortical predictions, and therefore send them — down to hypothalamus, PAG, PBN and the solitary nucleus to control the internal milieu. This is the visceromotor-areas argument, and the two papers agree on it almost exactly.
Barrett’s extension, and the more surprising hypothesis: because the laminar gradient is continuous, it does not stop at the limbic/sensory boundary. Motor cortex is less granular than primary sensory cortices, so it sends them predictions (following Adams, Shipp & Friston). And primary interoceptive cortex (mid-to-posterior dorsal insula) is less granular than primary visual, auditory and somatosensory cortex — so Barrett hypothesizes that interoceptive cortex sends sensory predictions to the exteroceptive senses.
Take that seriously and it is a strong, testable, and somewhat startling claim: your model of your viscera is predicting what you will see. Not “affect colours perception” as a metaphor — a directional anatomical claim about which cortex constrains which. “Accordingly, all action and perception are created with concepts. All concepts contribute to allostasis and represent changes in affect, not just those that construct the events that feel affectively intense.”
The wiki should record that this is a deduction from a gradient, not a measurement. But it is the kind of deduction that could be wrong in an interesting way, which is more than most of the theoretical claims here manage.
The three networks get three jobs
The most concrete set of hypotheses in the paper, and the most checkable.
| network | proposed computational role |
|---|---|
| default mode | hosts the internal model — multimodal summaries from which the prediction cascade issues through the whole cortical sheet |
| salience | issues precision signals: predicts which prediction errors are worth attending, encoding and consolidating; all other error is “noise and safely ignored” |
| frontoparietal control | sustains simulations past the few hundred ms of imminent prediction error; suppresses simulations whose priors are very low |
The salience-network entry is the one that changes an existing wiki page. The wiki has carried the salience network as an ascending detector of salient signals (Craig) versus a predictive comparator (Seth). Barrett gives it a third and more specific job — it doesn’t detect salience, it assigns it, by setting the gain on error-computing neurons before the error arrives. See salience-network.
What this does to the amygdala
Barrett takes the wiki’s most contested structure and makes it not about emotion at all: information from amygdala to cortex “is not ‘emotional’ per se, but signals uncertainty (Whalen, 1998) about the predicted sensory input (via the basolateral complex) and helps to adjust allostasis (via the central nucleus).”
This is a fifth reading for the amygdala page, and note that it is nucleus-specific — which is exactly what LeDoux demands and accuses Barrett’s camp of being unable to deliver. The wiki’s amygdala page currently records LeDoux’s charge that constructionist “salience detector” talk is an artifact of voxel resolution, treating “the amygdala” as one thing. Here Barrett assigns different jobs to different nuclei. That doesn’t answer the resolution objection — she still infers function from imaging plus theory — but it does undercut the specific complaint that she treats the structure as homogeneous.
Emotion perception is event perception
A small line with real consequences for the wiki’s methods pages:
The brain does not process individual stimuli — it processes events across temporal windows. Emotion perception is event perception, not object perception.
Nearly every empirical paper in this wiki shows subjects discrete stimuli and averages over trials — IAPS pictures at 6s, posed faces, single words. If emotion is constructed across temporal windows at multiple timescales, that entire measurement tradition is sampling the wrong unit. Barrett does not press the point, and no source here has tested it.
The self-defence, and what it concedes
The eight “I am not saying” clarifications are unusual enough in a theory paper to be worth reading as evidence about the state of the dispute. Three matter here:
- “I am not saying that emotions are illusions.” Emotion categories are as real as money — “the various objects that have served as currency throughout human history share no physical similarities.” Real, and perceiver-dependent, and not physically constituted. This is the analogy that does the most work, and it is the answer to anyone who reads constructionism as eliminativism.
- “I am not saying that non-human animals are 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’).” Notice she does not claim the fly feels nothing — “it may feel affect.” This puts her closer to LeDoux on can-we-know-animal-feelings than the polemic suggests, and closer to the core-affect page’s observation that core affect is the likeliest site of convergence between frameworks that agree on nothing else.
- “I am not claiming that subcortical regions are irrelevant to emotion.” An instance of emotion “engages the pattern generators for whatever actions are functional in the context.” The decorticate cats of Woodworth & Sherrington (1904) and Cannon & Britton’s (1925) “sham rage” are reinterpreted, not dismissed: pattern generators surviving the removal of the conceptual system, producing actions that look emotional but are no longer in service of survival.
The answer to Volynets that the wiki predicted
Worth flagging because the wiki wrote the answer before it had the source.
The lisa-feldman-barrett page carries a section titled “The challenge she has not answered here,” about Volynets et al. (2020) — 13 emotions with concordant felt body maps across 101 countries, cued by emotion words. The wiki argued that a constructionist has a good answer available: if categories are constructed from core affect plus widely shared concepts, concordant word-cued maps are a prediction of the theory rather than a refutation.
That answer is now sourced, three years earlier, in this paper’s own words: emotion categories “are conceptual because functions are imposed on physically disparate instances by virtue of collective agreement.” Shared concepts, shared maps. The wiki’s inference was correct and can stop being an inference.
This does not resolve cultural-universality-of-emotion — Barrett still has not engaged Nummenmaa’s data, and 2017 precedes 2020 — but it converts “an answer she could give” into “an answer she has given, to a different question, in terms that transfer.”
Where the argument is weakest
Recorded without prejudice to the theory, which is the wiki’s most explanatorily ambitious source.
The tables are curated by the defendant. Table 1 (evidence disconfirming the classical view) and Table 2 (evidence supporting constructed emotion) are the empirical spine of a 23-page argument, and both are assembled by the theory’s author, from a literature she is a principal contributor to. Several Table 2 rows cite her own group. One cites a manuscript “under review.”
The falsifiability problem is the wiki’s recurring one. Degeneracy predicts variability; variability is what the data show; therefore the data support the theory. That inference has the same shape as the one the wiki flags on homeostatic-property-cluster-kinds for Scarantino’s New BET — a framework in which nothing observed can count against it. Barrett half-acknowledges this and reframes it as a virtue: new paradigms “raise more questions than they answer… This is a feature, not a bug.” That is a fair description of early paradigms and not an answer to the objection.
A load-bearing anatomical claim rests on a personal communication. That neurons are multipurpose “even in subcortical regions like the amygdala” — which is what blocks LeDoux’s cell-population objection — is cited to Cerf, personal communication, 30 July 2015, in a footnote.
The pattern-classification argument cuts both ways. Barrett’s claim that classification successes are her strongest evidence rests on non-replication across studies (Kassam 2013; Kragel & LaBar 2015; Saarimäki 2016 all succeed; their patterns differ). But non-replication across studies with different stimuli, samples and methods is also just non-replication. Reading it as positive evidence for degeneracy requires assuming the studies were well-powered tests of the same thing — which, if true, would make the successes harder to explain away, and if false, makes the non-replication uninformative in either direction.