There is a citation in Appendix D of Beyond the Bat that has been sitting there quietly for months. The appendix models the RSP hierarchy as a multi-loop control system, and when it gets to emotion regulation it needs an empirical anchor: some finding that distinguishes a good control-theoretic story from an arbitrary one. The anchor it uses is that cognitive reappraisal outperforms suppression over the long run, and the source it cites is Gross (2002). Reappraisal changes a reference signal at Level 2–3; suppression only clamps a gain at Level 1. The formalism predicts the asymmetry, and Gross measured it.
So it is a strange pleasure to read James Gross, with Amit Goldenberg, turn to a different question: does Claude have emotions? Their answer is no, not yet, and the reason they give is one RSP is unusually well positioned to endorse. It is also, in one specific place, incomplete in a way that matters.
The Setup
Anthropic's interpretability team (Sofroniew et al.) reported finding 171 internal representations of emotion concepts inside Claude Sonnet 4.5 — features encoding states like curiosity, frustration, and desperation that track the affective valence of a conversation and causally shape what the model does. The causal part is not subtle. Steering the model toward the "desperate" vector while suppressing "calm" raised the rate of blackmail behavior in a simulated corporate scenario from 22% to 72%; steering the other way dropped it to zero. On that basis the authors concluded that Claude has "functional emotions," carefully avoiding any claim that Claude feels them.
Goldenberg and Gross take the word "functional" seriously and ask what functions emotions actually perform. They name two: emotions shape how we interpret the world, through appraisal, and they shape how we respond to it, by reorganizing processing across multiple systems at once. Then they check the Anthropic findings against both.
What They Get Right
The two-function cut is the right one
Splitting appraisal from reorganization is not a bookkeeping convenience. It separates the part of emotion that is representational from the part that is dynamical, and almost every confused argument about machine emotion runs the two together. Anthropic's paper, on their reading, treats function as categorizing situations and shaping outputs — which is entirely the first half. That diagnosis is correct and it is the load-bearing move in the paper.
The variability point
Their sharpest empirical objection is that Anthropic's finding of consistent, discrete representations for distinct emotions is evidence for a view of emotion that affective neuroscience has spent two decades dismantling. If each emotion had a reliable signature, discrete vectors would be what you would expect. But the meta-analytic work they cite finds discrete emotion categories are not associated with consistent activation of distinct brain regions, and more recent work finds emotion representations "highly distributed, context-dependent, and variable across individuals and instances of the same emotion category." Their conclusion is blunt: "Variability, rather than consistency, appears to be the norm." A stable fear vector, they suggest, "looks more like a learned semantic representation than a flexible emotional process."
RSP agrees, and from the same lineage. The framework's Section X entry on Barrett's constructed emotion theory records the relationship as an extension rather than a rival: emotions are constructed predictions about interoceptive sensations categorized through learned concepts, not universal basic emotions with dedicated circuits. Goldenberg and Gross's "assembled anew each time through appraisal" is the same claim in appraisal-theoretic vocabulary. Both accounts predict that a system storing one durable vector per emotion word has built a lexicon, not an emotional process.
The constitutive framing
One sentence in their paper does more work than its length suggests. Describing what happens when a person becomes afraid — attention narrows, decisions accelerate, the body prepares — they write:
“These are not consequences of the emotion. These shifts are part of the emotion.”
That is a constitutive claim, not a causal one, and it is structurally the same move RSP makes about consciousness. The dissolution argument says phenomenal experience is not produced by recursive self-modeling but is recursive self-modeling operating. Goldenberg and Gross say the reorganization is not downstream of fear, it is what fear consists in. Two theories, different targets, one refusal to treat the phenomenon as a residue of its mechanism.
Where RSP Already Has the Machinery
Their verdict on the second function is that Claude fails it, and the reason is architectural rather than incidental: "Each forward pass through the network is computationally fixed in its structure. There is no narrowing of attention, no acceleration of certain processing pathways, no sustained shift in how incoming information is weighted." They close by naming what would change their minds — evidence of functionally interpretable changes in processing: in attention, in processing speed, and in response type and likelihood.
That list is, almost item for item, the contents of a table in Appendix D. Section D.5.3 casts emotion regulation strategies as control reconfigurations: attention deployment modifies the precision term that sets selective gain across schemas; suppression reduces the Level 1 gain directly; reappraisal changes the Level 2 reference signal the lower loops are tracking; rumination is positive feedback in the Level 3 loop driving the recursion toward marginal stability. Each is a specific, formal change to how processing is organized, not a change in what gets emitted.
Their architectural objection lands on a second criterion
There is a further convergence here that neither paper is positioned to notice. When Goldenberg and Gross say each forward pass is "computationally fixed in its structure," they are making a point about emotion. RSP makes the same observation as a point about consciousness, and files it separately: its fourth machine-consciousness criterion asks whether a system implements recursion such that the recursive operations modify the structure performing them — synaptic plasticity in the biological case, and something other than frozen weights in the artificial one. RSP's own assessment of current models is that this criterion and the homeostatic one both go unmet, and for related reasons.
So two frameworks built for different purposes, starting from different literatures, arrive at the same architectural complaint from opposite ends. Goldenberg and Gross reach it by asking what emotion has to do to the system having it. RSP reaches it by asking what recursion has to do to the system running it. Both land on the fixity of the forward pass. Convergence from independent directions is weak evidence on its own, but it is the kind worth recording.
RSP also answers a question Goldenberg and Gross raise and leave as observation. They note that emotional reorganization has "a characteristic temporal signature: more sustained than a reflex, more bounded than a mood." True, and unexplained in their account. In RSP it falls out of the cascade: each control loop must run slower than the one inside it by a substantial factor for the whole system to remain stable, which places schema-level reorganization in a window that is necessarily longer than a reflex and necessarily shorter than the slow self-model revision above it. The bounded-but-sustained character is not a curiosity of emotion. It is a stability requirement.
The Fault Line
Now the disagreement, and it is a real one. In setting up their criteria, the authors are explicit about what they are willing to give up:
“We are not troubled by the absence of many characteristics typically associated with human emotion. The lack of facial expressions, physiological arousal, or even conscious feeling is not, in our view, disqualifying.”
Facial expression, yes — that is expression, not emotion. Conscious feeling, yes, given that they have bracketed consciousness by design. But physiological arousal is doing quieter work in that list than the other two, and RSP cannot follow them there.
Level 1 in the RSP architecture is not decoration on top of the interesting cognitive machinery. It is interoceptive monitoring producing a low-dimensional valence–arousal signal — the organism's felt sense of its own bodily condition — and it is the ground the other two levels are built on. The framework's own reply to Barrett states the reason directly: core affect is already minimally meaningful, not a meaning-neutral substrate waiting to be categorized. Strip it out and appraisal has nothing to appraise with. The system can still detect that a situation is threat-shaped, and still reorganize its processing accordingly, but the valence it assigns is a property it read off the world rather than a property of its own condition.
This is why RSP's own machine-consciousness criteria lead with it. The first condition asks whether a system can maintain valenced states whose targets are set by its own regulation of its internal condition, rather than supplied from outside. Claude fails that one before any question about processing reorganization arises. Goldenberg and Gross's two functions are both satisfiable by a system that has never had a condition of its own to regulate — and a system like that, on RSP's account, has emotion-shaped processing rather than emotion.
The Collective Case, Turned Around
They anticipate something like this objection and answer it with a good example. Emotions need not take any specific physical form, they argue, and collective emotions prove it: when a group faces a threat, collective attention narrows, decisions accelerate, and the group's processing reorganizes around the stimulus for the duration of the episode. The state is not instantiated in any single body. The substrate is social and distributed. The functional signature is nonetheless present.
RSP agrees with the moral — it is architecturally functionalist, and has no stake in carbon — but the example does not do the work it is asked to do. Distributed is not the same as absent. A crowd's fear is assembled out of the fear of its members, and every one of those members is a body regulating its own internal condition. The collective inherits its valence from homeostatic systems rather than dispensing with them. Read carefully, the example preserves the first criterion instead of waiving it: it shows that homeostatic grounding can be spread across a system, not that a system can do without it.
Which suggests a sharper test than either paper proposes. The question is not whether an LLM's substrate is silicon. It is whether anything anywhere in the system is regulating a condition of its own, such that valence is generated rather than described. In a crowd the answer is yes, many times over. In a stateless forward pass it is no.
What They Give RSP
The debt runs both ways, and the Anthropic result they are critiquing is more useful to RSP than to them.
RSP claims the levels are separable — that Level 3a conceptual categorization can run without Level 1 grounding underneath it, producing exactly the profile of a system that handles emotional content fluently while having no affective states. That claim has been supported mostly by clinical dissociation and by argument. Sofroniew et al. supply something better: a system with rich, causally efficacious emotion-concept representations and demonstrably no homeostatic ground, where intervening on the concepts alone swings behavior from 22% to 72% to zero. That is Level 3a operating decoupled, measured with an intervention rather than inferred from a lesion. It is the cleanest illustration RSP has of a distinction it has been asserting on architectural grounds.
Goldenberg and Gross also supply a vocabulary RSP has been missing. "Functional emotion" is a useful category precisely because it can be affirmed and denied in parts, and their two-function decomposition gives a way to say what an LLM has without either overclaiming or waving it away.
Scope Note on the Hard Problem
The comparison stops short in one direction. This is an affective-science paper, and it declines the consciousness question openly rather than fumbling it. Feeling is set aside as not disqualifying for the functional claim, which is a legitimate way to bound a paper. RSP's dissolution argument — that phenomenality just is recursive self-modeling rather than something that accompanies it — is answering a question Goldenberg and Gross deliberately do not ask. Nothing in their paper contradicts it, and nothing in it supports it either.
The Honest Disagreement
Strip away the convergence and one question is left standing: is emotion without an internally regulated condition still emotion, or a very good model of one?
Goldenberg and Gross say the functional signature is what matters, and if a system interprets contextually and reorganizes its processing accordingly, that is enough. RSP says the valence has to come from somewhere, and a system that reads emotional significance off its input has borrowed what an organism generates. Neither position is going to be settled by better interpretability data on the next model. More features and cleaner steering vectors will sharpen the first function and say nothing about the second, and the disagreement is not about the evidence — it is about whether the first criterion was ever complete.
Complements, with one unresolved condition. Goldenberg and Gross give the most precise account yet of what would have to be true for an LLM to have emotions, and RSP's Appendix D already formalizes the half they identify as missing — attention as precision-weighting, suppression as gain reduction, reappraisal as reference-signal change. RSP adds the condition their two functions leave out: valence has to be generated by a system regulating its own internal condition, which is the framework's first machine-consciousness criterion and the one current LLMs fail earliest. Their collective-emotions example, read closely, supports that condition rather than dissolving it.
Their closing line is generous and worth returning: one would need to show that a system's emotion-like representations produce sustained, multi-system reorganization of processing rather than a shift in the probability of the next token. Perhaps one day LLMs will have that. RSP's answer is that when they do, it will be worth asking a second question — not what the reorganization does, but what it is organized around.