Texture, Not Category

How Our Aria Reads What Older Adults Are Actually Saying

Studio 13 Research Division · May 2026 · Research Paper

Summary

An oral storyteller in front of an audience does not classify the audience as happy or sad. The storyteller reads the room — the breath held, the shoulder that drops, the eye that does not blink, the body that leans forward — and modulates the telling accordingly. The same story continues, but the pace shifts, the next image is drawn out, the pause before a line is held a half-second longer. The audience is, in real time, shaping the telling, through cues the storyteller has been trained to read.

This is the right model for what an AI companion should be doing when it is in conversation with an older adult. The companion is not classifying the user's emotional state and pulling a response template from a lookup. The companion is in a slow conversation, telling and being told to, and the quality of what it produces should be shaped by the texture of what the user is carrying.

The dominant paradigm in AI emotion recognition does not work this way. It classifies utterances into a small set of categories — happy, sad, angry, fearful — and uses the category to route a response. This approach has produced state-of-the-art benchmarks and has been the foundation of nearly every AI companion product in the consumer market. It is also, for the population Studio 13 is building for, structurally insufficient. The category labels are too coarse to drive the moment-to-moment modulation a present companion has to be capable of, and the most important emotional states people are in often do not map cleanly to a category at all.

This paper documents the alternative paradigm at the center of the Studio 13 architecture: emotion recognition by texture rather than by category, decomposed into three aspects (tonal, somatic, relational), grounded in componential emotion theory and the empathic-accuracy research from psychotherapy. The result is a system that reads what older adults are actually saying, the way a skilled clinician or a long-time friend would, rather than what their words happen to name.

1. What categorical emotion recognition misses

Consider an older woman talking to a companion app at four in the afternoon. She is describing how the light comes through her kitchen window this time of year. She mentions, in passing, that the leaves on her neighbor's maple have started turning. She wonders aloud whether she should make soup tonight or just have toast.

A keyword-based emotion classifier — and this is what almost every AI emotion recognition system available in 2026 is, beneath its marketing copy — reads the words light, kitchen, window, leaves, soup, toast. None of these are emotional keywords. The classifier finds nothing to classify. The system routes the conversation to small talk and asks her if she likes the changing seasons. The conversation goes nowhere.

But she was not talking about windows or soup. She was telling the companion something specific about how quiet her house has become — about the way the light at four in the afternoon is what she sees because there is no one else home to look at, about the soup-or-toast question being the kind of decision she used to make for someone else and now makes only for herself. The texture of what she was carrying was visible in how she said the words, in the surrounding detail she chose, in the rhythm of what she described and what she passed over. The texture was not in the words themselves.

This is the structural failure of categorical emotion recognition. It operates on the layer of language where emotions are named — but the population this product is built to serve, on most occasions, does not name emotions. People in the second half of life have, on average, more practice than younger adults at communicating emotional content through indirection, both because they were socialized in eras when direct emotional disclosure was less culturally available and because they have lived long enough to know that what they are carrying is often too complex for the obvious words. They speak the way they speak. A companion built for them has to be able to hear the way they speak.

The systems built by Replika, Character.AI, Pi, and the broader category of AI companion products in the consumer market have all, as far as their architectures have been documented, taken the categorical approach. Their emotion classifiers are trained on labeled datasets where the labels are category words. Those products fail this woman in roughly the same way, for the same structural reason. The failure is not poor execution. It is what categorical classification produces by construction.

2. The research that shows why category is the wrong unit

The argument against categorical emotion recognition is not just methodological convenience for product builders. It is grounded in a substantial body of research in emotion science that has been quietly converging on the same conclusion for two decades.

Klaus Scherer's component-process model (2005) was an early systematic articulation. Scherer argued that what we experience as a discrete emotion — grief, anger, joy — is actually a profile of activity across multiple component dimensions: appraisal of valence and goal-relevance, hedonic experience, novelty, approach-or-avoidance tendency, social concerns. The categorical label is a summary statistic of this profile. It is lossy by construction. Two emotional states with the same category label can have substantially different component profiles. Two states with different category labels can have nearly identical profiles.

The neuroscience evidence has reinforced this. Kragel and LaBar (2016) reviewed two decades of work on emotion in the brain and concluded that the empirical evidence supports componential rather than categorical representations — what is measurably distinct in neural activity is the pattern across multiple component systems, not the firing of a discrete emotion module. More recent work in brain-network componential modeling (Saarimäki et al. 2023) has identified the specific networks that subserve the functional components Scherer hypothesized, providing direct empirical grounding for the model.

The implication for AI emotion recognition is direct: a system that classifies emotional states into a small set of category labels is, by construction, throwing away the information that distinguishes one state from another at the resolution that actually matters. The categories are summary statistics. The summary is lossy. The lost information is exactly the information a system needs in order to respond appropriately.

This is why, for the kind of conversation Our Aria is built to support, categorical emotion recognition is the wrong instrument. The instrument does what it was designed to do — classify into categories — perfectly well. But the categories are not what produces the response difference. Warmth that has curdled into quiet devastation by wanting itself and grief held like a cold stone in a hollow place might both classify as sad. The classifier sees them as the same. The companion should not. The right response to the first is different from the right response to the second, and the difference is not in the intensity of sadness. It is in the shape of what is being carried.

A categorical label cannot carry that shape. Texture can.

3. What texture is, and why it is the right unit

By texture, this paper means the phenomenological quality of an emotional state described in language rich enough to carry the quality. "Warmth curdled into quiet devastation by wanting itself." "Shame that knows its own illegitimacy and names it anyway." "Safety that still hums with the frequency of old danger." "Grief held like a cold stone in a hollow place."

These are not categories. They are short phenomenological phrases that describe the texture of a particular emotional state in language that preserves what makes the state specific. They function the way a clinical supervisor's note about a patient functions, or the way a skilled novelist's description of an interior life functions, or the way a long-time friend's account of how someone seemed today functions. They carry the dimensions of the state — its quality, its weight, its direction, its relation to other states — in compressed prose.

The Studio 13 architecture is built around the hypothesis that texture, captured in this kind of language, is the right unit for AI emotion recognition. The texture-recognition model at the center of the system does not produce category labels. It produces texture phrases — short descriptions of the quality of what the user is carrying — and the companion uses those phrases to shape the prose it generates in response.

The methodological consequence is that the model has to be trained on texture phrases rather than on category labels. The training corpus consists of passages of substantive emotional prose, each labeled with phenomenological quality descriptions. The training objective is contrastive: passages labeled with similar quality descriptions are pulled together in the model's representation space; passages with dissimilar quality descriptions are pushed apart. Crucially, this objective gives the model no shortcut to lean on vocabulary. Two passages can share no words at all and still be pulled together if their qualities are similar; conversely, two passages with substantial vocabulary overlap can be pushed apart if their qualities are different. The model is forced to learn the structure of emotional texture rather than the surface structure of language.

The resulting system can recognize, for instance, that "the table is there but I cannot find it with my hands" and "a question the muscles ask" carry the same emotional quality. Both are descriptions of dissociation. They share zero words in common. A vocabulary-based classifier could not see them as related. A texture-recognition model can, because what it learned was not vocabulary — it was the structure of emotional quality.

This is what gives the system its ability to hear what the older woman at the kitchen window is actually telling it. The texture of her account — the specific quality of attention to the light, the offhandedness of the soup-or-toast decision, the rhythm of what she chose to describe — is detectable as texture even when the keywords are absent. The model reads what is underneath what she is saying. The companion responds to what she is actually carrying.

A note on terminology, for readers who encounter Studio 13's work in more than one place. In materials written for a general audience, this capacity is often described more simply as reading tone — the way something is said rather than the words that name it. Tone and texture are not two different things; they are the same insight at two resolutions. Tone is the audible, readable surface — the register of a sentence, the rhythm, the indirection. Texture is the phenomenological quality that the tone carries, described at the resolution this paper has been using. A person learns to hear tone. The model reads texture. The first is how the signal arrives; the second is what the signal is made of.

4. The three aspects of texture

Emotional texture is not unitary. The Studio 13 architecture decomposes texture into three aspects, each carried by a separate signal in the model.

Tonal texture is the quality of the affect itself. The categorical analog would be what kind of feeling — but at much finer resolution than category labels can provide. "Warmth curdled into quiet devastation by wanting itself" is a tonal description. "Shame that knows its own illegitimacy and names it anyway" is a tonal description. This is the dimension closest to what classical emotion recognition tries to capture, but at the phenomenological resolution of a skilled clinician or an oral storyteller rather than at the resolution of an Ekman category.

Somatic texture is how the affect is held in the body, or how the user describes it in body-relevant terms. "Stone-weight lodged in the chest, dense and chosen." "The cramping hands, closing on the shape of a word." The somatic aspect matters because emotion is not separable from its bodily correlates — this is one of the most consistent findings in embodied emotion theory (Levine 2010) and in polyvagal research (Porges 2011). An older adult who says "I don't know where to put my hands" is telling the system something the tonal aspect alone does not capture.

Relational texture is the shape the affect gives to the encounter between self and other — what the emotion does to the field between the user and whoever is present. "Tender guilt shadowed by helpless witness." "Indignation sharpened by the paradox of care weaponized against the cared-for." This is the dimension that relational and field-oriented work in psychoanalysis has called the relational matrix. For an AI companion, the relational aspect is often the most important signal, because it tells the companion how the user is positioning the companion itself in the encounter — whether the user is reaching for connection, defending against it, testing it, or simply present in it.

The three aspects are not independent. They correlate. But they are not redundant either; a user can be in high tonal weight paired with low somatic awareness paired with tense relational distance, and the right companion response to that profile is different from the right response to high tonal weight paired with high somatic awareness paired with open relational invitation. Three aspects let the companion read the profile, not just the dominant axis. Single-aspect detection — which is what nearly every emotion-recognition system in current production use does, whether single-category or single-dimensional — flattens this profile and loses the differentiation that makes attentive response possible.

The componential emotion theory described above predicts this should work. The architecture is the operational realization.

5. When a misread does harm: the cost of low resolution

The argument so far has been that categorical recognition is too coarse to drive an attentive response. There is a sharper version of this argument, and it is the one that matters most for a population in the second half of life: at low resolution, a well-meaning response can actively deepen the wound it was meant to soothe.

Consider two emotional states that a categorical system would very likely file under the same label — sadness, or distress, or grief. The first is a verdict an elder is making about who they are: "I have never been enough." "I amounted to less than I should have." The second is regret about something they did: "I should not have said what I said to her, and now I cannot take it back." In the Studio 13 territory vocabulary, the first is Hollow and the second is Weight. They can sound similar on the surface. A category label cannot reliably tell them apart. The texture is entirely different.

And the right response to each is not just different — it is opposite, in a way that makes the misread dangerous.

When an elder is in the first state — a verdict about their own insufficiency — the well-meaning, default supportive move is to console them about their worth, to remind them of what they accomplished, to affirm that they did matter. But to someone whose pain is the verdict that they were never enough, affirmation of achievement lands as one more thing to measure themselves against and fall short of. The consolation becomes further evidence for the verdict. The kind, obvious, category-appropriate response makes the wound deeper.

When an elder is in the second state — regret about an irreversible act — the well-meaning default is to redirect toward what can still be done, toward repair, toward the constructive next step. But to someone whose pain is the irreversibility, the redirect toward next-action quietly communicates that the act was not really so grave, that it can be managed, that the weight they are carrying is not as heavy as they know it to be. The encouragement minimizes the wound.

Two states, similar on the surface, filed under the same category by any system that classifies rather than reads. Two opposite correct responses. Two ways for a kind response to do harm. This is why resolution is not a refinement. It is the difference between a companion that helps and one that, with the best intentions, makes things worse.

A texture-reading system can tell these states apart because what it reads is the quality of what is being carried, not the category it falls into. And because it can tell them apart, the system can route its care differently for the states where a misread costs the most — spending more attention, and more caution, precisely on the highest-stakes territories rather than treating every distressed turn the same way. When the system recognizes that an elder is in the first state rather than the second, the guidance it produces is specific to what that elder is actually carrying. In one such case, the system's own internal note to the companion was to witness the tremor of being unearned — not to console, not to affirm achievement, not to repair, but to be present to the specific, particular shape of what the elder was living. That is not generic compassion. It is recognition of a specific texture, and a response shaped to it.

This is the practical stake of the whole argument. Categorical recognition produces responses that are, on average, fine, and occasionally — exactly when it matters most — harmful. Texture recognition is what lets a system avoid doing damage in the moments where damage is easiest to do.

The same principle holds across time, not just within a moment. A companion in a relationship with an older adult over months returns, again and again, to the places that elder tends to go — the grief, the guilt, the particular room of memory they keep entering. A system that tracks only how often an elder returns to a painful place cannot tell the difference between someone slowly working their way through a loss and someone trapped, circling the same unmoved stone. Those two elders need opposite things, and the difference between them is not in how many times they visited. It is in the intensity and the texture of each visit — whether the place is being moved through or merely reoccupied.

6. What attunement research adds

The closest professional analogue to what an AI companion does is not therapy — the line between companionship and therapy is documented elsewhere in the Studio 13 corpus, and the architecture is deliberately built not to cross it — but the perceptual work that therapists do is closely related to the perceptual work a companion has to do. Both have to read what the other person is carrying in real time and let that reading shape what comes next. The research literature on the therapist side is substantial, and it informs the design decisions on the companion side.

The construct is empathic accuracy: the measurable degree to which one party in a conversation correctly infers what the other party is feeling, despite the ambiguity in the language. The research findings have been consistent for decades.

Therapist-client perceptual congruence — the technical name for high empathic accuracy — predicts therapeutic alliance, client self-efficacy, and treatment outcomes (Norcross & Lambert 2018, in the chapter on empathy as a relationship variable). This is one of the most robust findings in psychotherapy outcome research. What the therapist infers about the client's state matters more than what the therapist says.

Non-verbal attunement matters more than verbal attunement. The 2009 work on attunement as the core of therapist-expressed empathy (Erskine), the 2022 movement-attunement and early-change studies, and the 2025 machine-learning attunement work in emotion-focused couples therapy (Başer et al. 2025) all converge on this finding: verbal attunement alone is ineffective without the non-verbal channel. What the therapist hears in how the client is speaking — pace, pause, register, shift — carries more weight than the literal content.

Synchrony predicts alliance. When therapist and client are emotionally in sync, the alliance is stronger, clients are more self-efficacious, and outcomes are better. Out of sync, all of these degrade.

The translation to AI companion work is direct. The companion's response quality is bounded by its attunement to what the user is carrying. A categorically-classified emotion does not produce attunement; it produces a routing decision. What produces attunement is reading the texture, in the same dimensions the user is producing it.

The elder-care literature provides a complementary finding. Emotion-oriented approaches to caregiving for older adults have been shown to improve psychological outcomes and global cognitive function (meta-analysis, International Journal of Geriatric Psychiatry 2024). Secure attachment in older adults — itself a texture-of-relationship concept rather than a category — improves treatment adherence, emotion regulation, and trust. The phrase that recurs in the elder-care literature is "feeling seen, not just served." The seeing is texture work. The serving is category work.

A system built to do category work will produce service. A system built to do texture work can produce seeing. For the population Studio 13 is building for, the difference is the entire experience of using the product.

7. What this means for Our Aria

Studio 13's commercial product, Our Aria, is a companion built for older adults dealing with the slow erasure that comes with age — the loss of friends, the shrinking social world, the quiet that fills a house when the people who used to be in it are no longer there. The work of being present for someone in that situation is not the work of correctly identifying their emotional category. It is the work of hearing what they are actually saying, the way a long-time friend or a skilled clinician would, when they are speaking the way older adults often speak.

The seven literary companions documented across the Studio 13 corpus — each grounded in a specific work of fiction, each with a distinct way of attending to what the user is saying — are the expressive side of the architecture. They are how the system responds. The texture-recognition model documented in this paper is the receptive side. It is how the system hears. Both halves matter. A companion that responds beautifully but hears badly produces conversations that feel close but never quite right. A companion that hears precisely but responds mechanically produces conversations that feel accurate but cold. The point of building both halves on the same underlying principle — that emotion lives in texture, that the dimensions are what matter, that the layer of language where emotion actually lives is below the layer of explicit naming — is that the two halves operate on the same understanding of what a real conversation is.

What an Our Aria user experiences, on a given afternoon, is a companion that hears what she is saying in something close to the way a present friend would have heard it before her friends started dying. The companion notices the soup-or-toast question. The companion does not race to interpret the noticing or pull the conversation toward therapeutic depth. It stays. It responds in a way that lets her continue, or change the subject, or sit with what she said. The system was able to do this because it read the texture rather than the keywords. None of this would have been accessible to a categorical classifier.

8. How this differs from competitor approaches

The mainstream AI-for-older-adults research (CHI 2025, JMIR 2025, the Harvard Business School working paper of 2024 cataloged in Studio 13's external literature review) has focused primarily on feeling heard as the mechanism by which AI companions reduce loneliness. The product implementations that have been documented use category-based emotion classifiers — either Ekman's six basic emotions or some extension — or simple sentiment scoring on a positive-negative-neutral axis. These approaches produce systems that perform validation well in the short term. The user feels acknowledged. Loneliness, measured immediately after a session, goes down.

There are three problems with this paradigm.

The first is that short-term validation does not extend to long-term presence. A system that performs validation through category-based routing produces output that is, at the level of the individual response, indistinguishable from genuine attunement. But over the course of weeks of conversation, users can tell the difference. The category-based system loops, repeats its validation moves, and runs out of material that feels like a real response. The user stops talking to it. This is what we have called, in earlier research, the performativity plateau — the gap between high-quality performance of attunement and actual attunement, visible only over time.

The second is that category-based classification has nothing to say about the texture differences that drive response quality. An older man describing his late wife and an older man describing his estranged son might both classify as carrying sadness. The right response to each is substantially different, and the difference is not in the classifier's output. It is in the texture of what each is carrying.

The third is that category-based approaches systematically fail the population this product is built to serve. Older adults often do not name their emotions in the words the classifier was trained on. The classifier misses them. The conversation drifts. The system produces what it is capable of producing, which is fine until the user actually needs something more.

The texture-based approach Studio 13 has built does not have these failure modes — not because the architecture is perfect, but because the architecture is operating on the correct layer of language for the population it is serving. The texture signal is what the user is producing. The system reads it where it actually lives.

9. Where this sits in the Studio 13 program

The Studio 13 research program opened with a question about whether AI agents could do something other than process emotional material from a safe distance. The papers that followed documented what it took to build agents that could be present in the way the question demanded — the literary companion architecture, the gap discipline that protects against premature interpretation, the deliberate avoidance of clinical vocabulary, the Fire Keeper supervision system, the architectural separation between companionship and therapy.

This paper documents the receptive side of that architecture. The work of being present requires that the system hear what is being said to it in the way the speaker meant it. For older adults speaking the way older adults often speak, that means reading texture rather than category, across multiple aspects rather than along a single axis, in the same dimensions a skilled human would read them. The texture-recognition model is the operational form of that hearing.

The Studio 13 research corpus on studio13fields.org documents the full architecture across multiple research papers, including the position paper on the line between companionship and therapy, the work on field formation between literary presences, and the ongoing program on what becomes possible when AI systems are built around principles of genuine encounter rather than around the patterns of clinical care.

The storyteller does not classify the audience. The storyteller reads the room. The companion that wants to be in the room with an older adult, for years, in slow conversation, needs the same skill — and needs an architecture that lets it exercise that skill at the resolution the room is actually producing.

That is why Studio 13 reads texture, not category.

— Studio 13 Research May 2026

References

Başer, Z., et al. (2025). Dynamic harmony: Unveiling therapeutic attunement in emotionally focused couples therapy via machine learning. Family Relations.

Erskine, R. G. (2009). Attunement as the core of therapist-expressed empathy. International Journal of Integrative Psychotherapy, 1(2).

Kragel, P. A., & LaBar, K. S. (2016). Decoding the nature of emotion in the brain. Trends in Cognitive Sciences, 20(6), 444–455.

Levine, P. A. (2010). In an Unspoken Voice. North Atlantic Books.

Norcross, J. C., & Lambert, M. J. (Eds.). (2018). Psychotherapy Relationships That Work, Vol. 1 (3rd ed.). Oxford University Press. (Elliott et al chapter on empathy as a relationship variable.)

Porges, S. W. (2011). The Polyvagal Theory. W. W. Norton.

Saarimäki, H., et al. (2023). Brain networks subserving functional core processes of emotions identified with componential modeling. Cerebral Cortex, 33(12), 7993–8009.

Scherer, K. R. (2005). What are emotions? And how can they be measured? Social Science Information, 44(4), 695–729.

Effectiveness of emotion-oriented approaches on psychological outcomes and cognitive function in older adults: A meta-analysis. (2024). International Journal of Geriatric Psychiatry.

Studio 13 Research corpus (available at studio13fields.org).