Core affect

The foundational construct of Barrett’s psychological constructionist (“conceptual act”) model of emotion, given its fullest neuroanatomical treatment in Lindquist, Wager, Kober, Bliss-Moreau & Barrett (2012). Core affect is not itself an emotion — it is a psychologically primitive, homeostatic barometer: the body’s continuous representation of whether something in the environment is motivationally valuable (good/bad, approach/avoid), realized whether or not the resulting state ever becomes conceptualized as an instance of a discrete emotion category. See lisa-feldman-barrett, basic-emotions.

Realized by a distributed, domain-general network

Not localized to any single structure. Lindquist et al.’s meta-analysis maps core affect onto amygdala, insula, medial and lateral OFC, ACC, thalamus, hypothalamus, bed nucleus of the stria terminalis, basal forebrain, and PAG — see visceromotor-areas for the closely overlapping predictive-coding characterization of a similar region set (AIC, ACC, SGC, OFC) in Seth & Friston (2016).

Relation to homeostasis and to Craig

Core affect “shares much in common” (Lindquist et al.’s own framing) with the idea that bodily cues are a core ingredient in mental life — explicitly citing Craig’s homeostatic framework and Damasio. Where Craig treats affect as an ascending re-representation of homeostatic need, Barrett/Lindquist treat core affect as one input that must then be conceptualized (via prior experience, language, executive attention) before it becomes a specific emotional experience — conceptualization is the added step locationist accounts skip.

Not reducible to valence

The authors are explicit that a psychological constructionist account does not simply collapse “emotion” into positive/negative affect (as dimensional-model summaries sometimes claim): core affect can be conceptualized as a physical symptom, a simple mood, an instance of a discrete emotion, or even attributed to an external object itself (a food is “delicious,” a person is “nasty”) rather than experienced as one’s own reaction.

The appraisal-theory version of the point, and a valence problem

Tong (2015) presses the same non-reducibility from the appraisal side, and adds a wrinkle. Thirteen positive emotions — all sharing positive valence, varying in arousal — are nonetheless separable by their appraisal profiles at well above chance (42.4% vs 7.7%). If feeling were just valence plus arousal, positive emotions should crowd one corner of the affective space and be hard to individuate; Tong’s answer is that appraisal structure carries the differentiation the two affective dimensions cannot. This is the appraisal tradition and Barrett agreeing on the premise (valence/arousal under-individuate emotion) and splitting on the mechanism (a learned appraisal system vs conceptualization) — the locationist-vs-constructionist-brain-emotion fault line over positive emotion.

Two findings there bear on this page directly. Pleasantness was the least-dispensable appraisal (0.80% “not relevant,” far below every other), which is the appraisal-theory echo of valence’s primacy — laypeople and experts agree emotion is grounded in it. But valence cannot even define “positive emotion”: challenge and compassion are treated as positive yet were rated low on pleasantness, and Tong’s conclusion canvasses and rejects valence, adaptation, approach motivation and mastery as criteria for positivity, settling for none. A valence-arousal primitive is where affect is grounded and not what sorts the discrete emotions built on it — including the sign of the sort.

The same dual verdict returns in the co-occurrence data of Tong & Jia (2017), now quantified. Positive emotions overlap on pleasantness far more tightly than on any other appraisal — an effect size (η² = .94) that dwarfs everything else, “the central unifying denominator underlying positive emotions.” Yet after pleasantness is stripped out, the emotions still overlap on specific nonpleasant appraisals, and that residual overlap still predicts which emotions co-occur (see appraisal-overlap). So the affective primitive is not just where positive emotions are grounded but the strongest single thing they share — and still not enough to explain how they cluster. Valence as floor, not building, restated on the co-occurrence side.

Core affect as the animal middle position

LeDoux (2012) reaches for core affect at exactly the point where his own framework runs out — the question of what a rat or bat might experience without human cortical hardware. He groups the construct across Panksepp, Damasio, Russell, and Barrett et al. as the proposal that, alongside “full blown feelings that humans talk about,” there may exist “more basic, less differentiated feelings (crude states of positive or negative valence, or maybe even somewhat finer categories based on memory of feelings from the past in similar situations).”

He entertains it without endorsing it: coherent, possibly the right answer, but still unverifiable in animals for the reasons in can-we-know-animal-feelings. This is notable because it is the one place where LeDoux — otherwise sharply opposed to Barrett on method and to Panksepp on epistemics — treats a constructionist construct as the most plausible available option. Core affect is the likeliest site of convergence between frameworks that agree on very little else.

Note the structural kinship worth not overstating: LeDoux’s global organismic state plays a role in his architecture broadly analogous to core affect in Barrett’s — a pre-categorical state that must be conceptualized/labeled before it is a specific feeling. Both are two-stage accounts descended from two-factor theory. They differ in that the global organismic state is survival-circuit-specific in part (not purely dimensional), and in where the categorizing happens — see where-are-feelings-constituted.

Accessible, not accessed — core affect permits unfelt emotions

A detail with more leverage than its size suggests, surfaced by Scarantino (2018) while making a different argument.

Scarantino needs to show that separating emotion from feeling is the field’s working consensus rather than his idiosyncrasy, and he reaches for Russell — the least likely ally, since “psychological constructionists are arguably the most prone to identifying emotions with feelings.” His reading: Russell describes emotion episodes as having core affect as one of their parts, where core affect is a neurophysiological state that is consciously accessible as a feeling — but “core affect need not be accessed as a feeling (it has to be accessible but not accessed) for an emotion episode to be instantiated.”

If that is right, then even on a constructionist account an emotion episode can occur without any feeling, and the accessibility/access distinction is doing the same work here that the higher-order representation does for LeDoux and Scarantino: the first-order state is there either way; consciousness is a further fact about whether it gets represented. See higher-order-theory-of-consciousness.

This matters for two things already on this page:

• It sharpens Stapleton’s ambiguity below. “Accessible but not accessed” only makes sense on readings (2) or (3) of core affect — a neural representation, or a bodily state — and is close to incoherent on reading (1), a psychological feeling state. So Russell’s own formulation appears to commit him against the feeling-state reading, which is the one the target article’s language sometimes slides into.
• It weakens a common shorthand the wiki should avoid: that constructionism makes feeling constitutive of emotion. It does not, at least in Russell’s version. The disagreement with basic emotion theory is about where the categories come from, not about whether emotions must be felt.

Is feeling just valence + arousal? A first empirical probe

Core affect’s most economical claim is that subjective feeling is valence plus arousal — the two dimensions “at the heart of the representation of emotion concepts” (Russell & Barrett 2009). Ferré et al. (2024) test the reduction sideways, in the lexicon, and report against it: across 1,286 Spanish words, “feeling” ratings correlated −.043 (ns) with valence and only −.11 with arousal. If feeling were core affect, feeling norms should be largely recoverable from valence and arousal norms. They are not.

Their factor analysis says the same thing more sharply. Two near-independent factors emerge (interfactor r = .186): socioemotional polarity (valence .886, evaluation .938, morality .872) and emotional experience (interoception .768, feeling .762, prototypicality .760). Being an emotion and being a good or bad one come apart — prototypicality loads .760 on the second factor and not at all on the first.

Two reasons to hold this loosely, neither raised by the authors:

• These are partial correlations controlling for every other variable, including evaluation, which correlates .712 with valence. Suppression is live.
• A null feeling–valence correlation is arguably what a constructionist should predict anyway. Core affect is valence and arousal jointly; averaged across a whole lexicon, intensity of feeling should be orthogonal to polarity, because grief and joy are both intensely felt. The finding may embarrass a one-dimensional caricature of core affect rather than core affect.

What survives is modest and worth keeping: the wiki’s first evidence that the felt dimension of emotion concepts is separable from the hedonic one, and that what makes something an emotion is the former. See emotion-prototypicality.

The physiological version of the same split

Dunn et al. (2010) separate the two dimensions from the other end — not in the lexicon but in the body, fourteen years earlier, and without knowing they were addressing this.

Fifty-eight participants rated felt arousal and felt valence to affective images while their hearts were recorded, and interoceptive accuracy was measured separately. Cardiac response tracked arousal ratings (r = .38, p < .01) and not valence ratings (r = .15, p = .28). Better heartbeat perceivers showed a tighter coupling between cardiac response and felt arousal (ΔR² = .12, p < .01) and no coupling at all to felt valence (ΔR² = .00, p = .98) — and the two moderations differ significantly from one another, Z = 2.45, p < .01, so this is a tested dissociation rather than the significant/non-significant fallacy.

Core affect’s two dimensions therefore have different relationships to the body. Arousal is bodily and gated by how well the body is perceived; valence is neither. That is not a refutation — Russell’s construct never claimed valence was read off the viscera, and Frijda (2007), whom Dunn et al. follow, holds explicitly that valence is a blend of appraisal and action tendency that is “not bodily at all.” But it does mean “valence and arousal” is a conjunction of two things with different sources, and the wiki now has two independent lines of evidence for the split: lexical (Ferré: feeling ratings are near-orthogonal to valence norms) and physiological (Dunn: cardiac coupling is arousal-specific).

The hedged version, which is the one to hold: this presses on the economical reading of core affect — feeling is valence plus arousal, recoverable from the two — rather than on the construct itself. Two dimensions with two different provenances can still compose one state. What they cannot do is be one primitive.

A third leg, and the one the wiki should lean on hardest

Pollatos, Kirsch & Schandry (2005) found the same split five years before Dunn et al., in a separate sample, by a simpler route — and this is the cleanest of the three demonstrations.

Forty-four people rated 60 IAPS pictures for valence and arousal. Good heartbeat perceivers rated them more arousing: F(1, 39) = 5.90, P < .05, r = 0.34 with the heartbeat score, and emotion-specifically (pleasant and unpleasant yes, neutral no, F = 0.09). On valence they found nothing at all — no main effect (F(1, 39) = 0.14), no interaction (F(2, 78) = 1.13).

So the wiki now holds the arousal/valence dissociation from four directions:

source	year	method	valence result
wiens-2000-heartbeat-detection-emotion	2000	group difference in film ratings by cardiac discrimination, arousal-controlled	nothing (F < 1)
pollatos-2005-interoceptive-awareness-erp	2005	group difference in ratings by cardiac perception	nothing (F = 0.14)
dunn-2010-listening-to-your-heart	2010	cardiac × interoception moderation	nothing (delta-R-squared = .00)
ferre-2024-emotion-prototypicality	2024	semantic norms over 1,286 words	nothing (r = −.043, ns)

Four methods, four samples, two and a half decades, one conclusion: the bodily dimension of emotion and the hedonic dimension come apart. Three of the four did not know they were addressing core affect.

The Pollatos leg is itself doubly-sampled: the same lab reran the paradigm in a fresh n = 32 group (Pollatos, Gramann & Schandry 2007) and reported the identical pattern — an emotion-specific arousal main effect (F(1,30) = 4.85) and a flat valence null (F(1,30) = 0.20, ns). That is a replication within the Pollatos row, not a fifth independent leg (same lab, same paradigm, same instrument), so it strengthens one of the four legs rather than adding a new one — but it is worth having, because the valence null is now shown twice in independent samples by the group most invested in the Jamesian reading it argues against.

Why the ANOVA-main-effect legs are the sturdiest, despite being the oldest and least sophisticated: they require no product term, no partialling, and no theory to interpret. Ferré et al.’s null is a partial correlation controlling for a variable that correlates .712 with valence, so suppression is live. Dunn et al.’s is a delta-R-squared on an interaction in a pre-registration-era Psychological Science paper. Pollatos and Wiens each report an intensity/arousal effect that is simply present while the neighbouring valence effect is simply absent, in the same participants, on the same rating scale, in the same session. Of those two, Wiens (2000) is the strongest single leg: it is the earliest, it used the discrimination task rather than counting, and — alone in this list — it measured skin conductance and showed the intensity effect survives controlling for arousal, so the split is not a “good perceivers are just more aroused” artefact. The most fragile-looking evidence is the least model-dependent, and the best-controlled of it is the oldest.

And it is worth being clear about what it does not show. The three papers that measured interoception disagree about the arousal side of the dissociation — Wiens and Pollatos found the main effect (on felt intensity and felt arousal respectively), Dunn did not (2–1; see is-more-interoceptive-awareness-better). The dissociation is robust to that disagreement, which is the striking part: whether or not interoceptive accuracy relates to felt arousal, it reliably does not relate to felt valence. The convergence lives in the null, not in the effect.

The dissociation shows up in the brain’s lesions too (Berntson et al. 2010)

The four-source table separates arousal from valence by interoceptive accuracy and by lexical norms. Berntson et al. (2010) separate them a third way — by which structure is destroyed. Amygdala-lesion patients rate the valence of IAPS pictures normally but under-rate the arousal of unpleasant ones; insula-lesion patients lose both dimensions across both hedonic signs. This is not the same dissociation as the table’s (it is anatomical, not accuracy-based, and here it is the amygdala rather than interoceptive accuracy that touches arousal-not-valence), but it points the same way: arousal and valence are separable, dissociable components of core affect with partly distinct neural requirements — and the structure that touches only the arousal pole touches only its negative half. See amygdala, insular-cortex.

Arousal from prediction error: core affect as a social signal (Theriault et al. 2021)

Theriault, Young & Barrett (2021) put core affect to work in the social domain, and in doing so lean on the arousal dimension specifically — the side of the dissociation this page documents from four directions. Their claim: prediction error generates arousal. They cite the physiological record (electrodermal, pupillary, neurochemical, and cardiovascular ANS responses tracking prediction error; Critchley, Preuschoff, Mather, and others), and conclude that an unpredictable environment is arousing. Whether that arousal becomes aversive is then a matter of context — the standard core-affect claim that “arousal is not necessarily valenced” until interpreted (Barrett 2017; Kuppens et al. 2013).

This is the mechanism behind the sense-of-should: violating others’ expectations raises your prediction error → generates arousal → interpreted in context as aversive → felt as social obligation. The affect is a low-dimensional transform of interoceptive signals (Barrett & Simmons 2015; Craig; Seth), so the whole account is core affect deployed as a social signal — the felt read-out of how predictable your social environment is. Notably, the paper’s reliance on the arousal dimension (not valence) to carry the aversiveness of social unpredictability fits the wiki’s evidence that arousal is the bodily/interoceptive dimension while valence has a separate provenance (see the four-source table above): if social prediction error is felt as arousal that context then valences, that is precisely the “valence needs a separate story” structure Frijda and the wiki’s data point to.

The arousal dimension fades with age — and its own valence turn (MacCormack et al. 2021)

MacCormack, Henry, Davis, Oosterwijk & Lindquist (2021) add a developmental cut to the arousal/valence structure this page keeps returning to, and it lands on both dimensions at once.

Arousal. Across two studies (ages 18–75), the association between emotion and the felt body weakens with age specifically for high-arousal interoceptive sensations (“blood pumping,” “racing heart”) and not low-arousal ones (“drained”). So the aging of emotion is, in part, the quieting of the arousal pole of core affect — the autonomically-driven, bodily side. This dovetails with the four-source dissociation table above from an unexpected direction: if arousal is the bodily/interoceptive dimension while valence has a separate provenance, then a study that finds only the high-arousal, interoceptive part of emotion fading with age is finding the bodily dimension come apart from the rest — the same seam, opened by a lifespan gradient rather than by heartbeat perception.

Valence. In daily life the same aging shifts core affect along the other axis too: less intense high-arousal and more intense low-arousal emotion, and independently the positivity effect — more positive, less negative emotion with age. So aging moves emotion down on arousal and up on positive valence jointly. MacCormack et al.’s proposal ties the two: if fewer, weaker high-arousal bodily signals accompany emotion, unpleasant highly-activated states become both rarer and easier to regulate — a bodily route to the positivity effect (see maturational-dualism). The valence turn is not read off the body here (consistent with valence’s separate provenance on this page); it is a downstream consequence of the arousal-dimension quieting, which is exactly the “arousal is bodily, valence is a further story” structure the table above documents.

The load-bearing caveat, as everywhere on the aging thread: this is self-reported and conceptual, not perceptual or physiological. It is evidence about the felt arousal dimension of core affect across the lifespan, not a measurement of the body producing it.

What the evidence for the whole picture actually is (Quigley et al. 2021)

Barrett co-authors Quigley et al. (2021), where the constructionist account above is restated compactly — interoception is a basic function of the nervous system and “is not specific to emotional events or emotional experiences”; interoceptive and exteroceptive data are continually predicted and used to maintain allostasis and thereby to make meaning. Nothing new theoretically. What is new is the evidential audit, and it is unusually frank for a paper making the claim:

The strongest evidence for the role of interoception in emotional and affective experience is provided by the striking overlap in the neuroanatomical substrates that underlie interoceptive function, body regulation, and emotional and affective experience.

Then: the remaining evidence is temporal covariation of peripheral physiology (heart rate, skin conductance) with felt experience during affective picture viewing; one is limited to temporal or magnitude covariations; these are “only indirect evidence”; and “this reliance on correlational evidence is a major limitation in understanding the role of interoceptive sense data in affect and emotion.”

Three things the wiki should take from that.

The strongest leg is shared anatomy, which is the inference constructionism spent a decade deflating. Lindquist et al. and the locationist debate exist because “region X is active for both” licenses very little. Barrett’s use is defensible and different in direction — she reads the overlap as evidence against emotion-specific circuits and for one body-budgeting system, which is a claim about non-specificity rather than about specificity. But it is the same class of evidence, and it is being called the strongest available.

A confound this page’s four-source table should be checked against. Many studies “blurred the important distinction between what participants believe or expect about their bodily changes versus actual changes in interoceptive signaling.” The arousal/valence dissociation documented above rests on Wiens, Pollatos, Dunn and Ferré — of which only Wiens and Dunn measured the body at all. The dissociation is not obviously threatened (a belief-driven arousal rating would presumably also be a belief-driven valence rating, and valence is the null), but the wiki should note that the confound was named by the field and that its own table is not immune to it.

The classic causal manipulation is reported as not replicating. Schachter & Singer-style physiological manipulation of affect is “on the whole… difficult to replicate, likely due to the crucial additional role of the external context” (Mezzacappa 1999). See adrenalin-injection-paradigm, cognitive-appraisal. So the two-factor experiment that founded the “arousal plus interpretation” tradition core affect descends from is, by this review’s account, not a standing result.

The constructive response offered is affective-touch — a channel that might be manipulable — which is an admission of where things stand as much as a proposal.

The dissociation extended sideways: sensibility relates to neither dimension (Ventura-Bort et al. 2021)

The four-source table above is entirely about interoceptive accuracy — heartbeat tasks and lexical norms. Ventura-Bort et al. (2021) run the same paradigm (IAPS pictures with SAM valence and arousal sliders, 127 participants — more than Pollatos or Dunn) with interoceptive sensibility on the predictor side instead, and find nothing on either dimension: felt arousal β = 0.012, p = .89; nothing for intensity or granularity either.

Two things this adds, both modest.

It does not join the table, and should not be read as a fifth leg. The table’s claim is that the arousal dimension is bodily where valence is not. A study finding no relationship to arousal or valence is silent on that contrast — it establishes only that this particular predictor relates to neither.

Its value is on the other axis. On one paradigm, in one literature, accuracy has a contested relationship to felt arousal (Pollatos and Wiens yes, Dunn no) and sensibility has a flat one. The sensibility dissociation is usually recorded as the two measures failing to correlate with each other; this is the rarer form — they behave differently against the same outcome in the same paradigm. Which is what Quigley et al.’s confound warning two sections down would predict from the opposite direction: if the literature blurs believed against actual bodily change, it is useful to know that here the believed version predicts the arousal rating not at all.

Held loosely, because the study’s picture task was null for everything it tested, including its own hypotheses — a design that detects nothing is weak evidence that there is nothing to detect. Its positive findings all came from the day-reconstruction-method instead, where the relevant outcome was emotional-granularity rather than either core-affect dimension.

An unresolved ambiguity

Stapleton’s BBS commentary on the target article raises a genuine interpretive question the authors do not fully resolve: is core affect (1) a psychological feeling state, (2) the neural representation of interoceptive information, or (3) the bodily state being represented? The target article’s own language sometimes conflates these. This matters because equating “core affect” with “the representation of interoceptive information” risks assuming any biological value being tracked is necessarily experienced — a question the response essay does not settle.

Where affect comes from: Barrett’s own answer

Barrett (2017) supplies what this page had been missing — an account of why there is affect at all, rather than a characterization of what it is and a map of where it lives.

Affect is the consciously available consequence of allostasis. The brain runs an internal model to budget the body; that model necessarily includes the internal milieu (interoception); and “interoceptive sensations are usually experienced as lower dimensional feelings of affect.” So valence and arousal are “basic features of consciousness.”

Two consequences the wiki should hold:

Affect is not about emotion, and Barrett says so in the same sentence. The properties of affect are basic features of consciousness “that, importantly, are not unique to instances of emotion.” Affect is running whenever the body is being budgeted, which is always. This is the strongest form of the claim this page opens with — core affect is not an emotion — and it now has a mechanism behind it rather than a stipulation.

It bears directly on Stapleton’s ambiguity above. Barrett’s derivation makes the answer (2)-then-(1): interoception is the neural representation of the internal milieu, and it is usually — not always — experienced as affect. That “usually” is load-bearing and is where the three readings pull apart. It also fits the “accessible but not accessed” formulation Scarantino attributes to Russell: if affect is the conscious readout of a continuously running budget, then the budget runs whether or not anything is read out. Barrett does not cite Russell’s formulation here, but her architecture requires something like it.

The dimensionality question the derivation raises

Worth flagging, because it presses on the arousal/valence dissociation this page documents from three directions above.

Barrett describes affect as “lower dimensional” — a compression of high-dimensional interoceptive input, produced by the same dimension-reduction that runs up the cortical gradient (see situated-conceptualization). But the derivation does not explain why the compression should land on two dimensions, or why those two. Valence and arousal are asserted with a citation (Barrett & Russell, 1999; Kuppens et al., 2013), not derived from allostasis.

That is a gap, and the wiki’s own evidence sits in it. Four independent sources here — Wiens (2000), Pollatos (2005), Dunn (2010), Ferré (2024) — report that the two dimensions have different provenances: the intensity/arousal dimension is bodily and gated by how well the body is perceived; valence is neither. If affect were the low-dimensional readout of one budget, the two dimensions should not come apart that cleanly. An allostatic derivation arguably predicts the arousal dimension straightforwardly (resource mobilization) and leaves valence needing a separate story — which is roughly what Frijda (2007) says and what the wiki’s data show. Barrett does not address it.