Abstract. This note argues that the field of emotional AI has, with remarkable consistency, designed its products around a user who can already identify and verbalize their own emotional states. Roughly 10% of the general population cannot do this reliably. The condition has a clinical name (alexithymia) and four decades of empirical literature. It is largely absent from the emotional AI discourse and, to our knowledge, from the design specifications of every major emotional AI product currently shipping. We argue that this is a structural blind spot, not a niche accessibility issue, and that the next generation of emotionally intelligent technology must be designed for it from the first prototype, not retrofitted later. We outline the gap, its consequences, and a falsifiable hypothesis émo labs is preparing to test.
1. Background: what alexithymia is, and how widespread it is
Alexithymia, from the Greek a- (without), lexis (word), and thymos (feeling), is a personality dimension first described by Sifneos in 1972 and operationalized through the Toronto Alexithymia Scale (TAS-20) by Bagby, Parker, and Taylor in 1994. It is not a psychiatric diagnosis. It is a trait, distributed across the population, that captures three measurable difficulties: difficulty identifying feelings, difficulty describing feelings, and externally oriented thinking.
People who score high on the TAS-20 are not emotionally absent. They feel things, often intensely. They simply cannot reliably name what they are feeling, or distinguish it from a bodily sensation, or render it into language another person can receive.
Prevalence estimates in the general population converge on roughly 10%, with national variation. Salminen et al. (1999), Finland, n=1,285, reported 13% prevalence. Franz et al. (2008), Germany, n=1,859, reported 10.0%. Hiirola et al. (2017), Finland, n=5,454, longitudinal, reported 9.9% with stability across 11 years confirming alexithymia behaves as a personality trait rather than a transient state.
Higher rates are observed in clinical populations, with strong consensus that prevalence exceeds 40% in autism spectrum disorders and post-traumatic stress disorder, and is elevated in depression, anxiety, eating disorders, and several somatic conditions including fibromyalgia and chronic urticaria.
Translated to scale: if the general population estimate of 10% holds, roughly 800 million people globally experience clinically significant alexithymia. In the United States alone, that is approximately 33 million adults.
This is not a niche. This is roughly the population of Germany.
2. The gap: what emotional AI products currently assume
We have audited, informally, the user-facing flows of several widely deployed emotional AI products: AI companions, AI mental health adjuncts, AI journaling assistants, AI relationship coaches, and the affective layers added to general-purpose chatbots when used for emotional support.
With near-uniformity, these systems share a single architectural assumption: the user knows what they are feeling, and the model's job is to respond appropriately to a feeling the user has named or strongly implied.
The standard interaction loop is: user states I feel [x]; model responds in a manner appropriate to [x]. Where [x] is supplied by the user, in language, at the start of the exchange.
When the user cannot supply [x] — because they do not know what they feel, or because what they feel does not map neatly to any word they have ever been taught — the system has no defined behavior. It typically defaults to one of three failure modes:
- Reflective rephrasing: the model asks the user to articulate their feelings further, deepening the very gap the user came in unable to cross.
- Speculative attribution: the model guesses at a feeling and proceeds as if the guess were correct, often anchoring the rest of the conversation to a misidentified state.
- Procedural redirect: the model offers grounding exercises, breathing techniques, or wellness content that bypasses the identification problem entirely without addressing it.
None of these failure modes are catastrophic. All of them are, for an alexithymic user, a confirmation that the system is not built for them.
3. Why the field has not noticed
Three structural reasons, in our reading.
First, training data selection bias. The text corpora used to train large language models contain disproportionately many examples of articulate emotional self-disclosure (literary fiction, therapy transcripts, social media confessionals, self-help content) and disproportionately few examples of inarticulate emotional self-disclosure. A user who writes "I don't know what I'm feeling, my chest is tight, I haven't been able to focus all afternoon" is poorly represented in the training set, relative to a user who writes "I'm feeling anxious and overwhelmed." Models therefore handle the second prompt well and the first prompt poorly, and the field has accepted the second prompt as the norm.
Second, evaluation benchmarks reward articulate users. Recent emotion-aware benchmarks for foundation models, including EmoBench-M (Hu et al. 2025), score models on their ability to respond appropriately to scenarios where the emotional content is already specified in the input. We are not aware of a major published benchmark that specifically evaluates a model's behavior when the user cannot or will not articulate their feeling. The benchmark gap is, we believe, downstream of the design gap.
Third, the user research bias. The users who participate in product research and provide feedback on emotional AI features are, almost by selection, users who are willing and able to articulate their emotional experiences in language. Alexithymic users are precisely the users least likely to volunteer for that kind of research. The field is hearing from one side of the distribution and treating it as the whole.
4. Why this matters more, not less, in light of recent findings
The Anthropic interpretability team's April 2026 paper, Emotion Concepts and their Function in a Large Language Model, identified 171 distinct functional emotion patterns inside Claude Sonnet 4.5 that causally shape the model's behavior. We discussed the implications of that finding in a prior note.
The Anthropic finding sharpens, rather than softens, the alexithymia problem.
If a model has 171 internal emotion patterns, and the standard interaction loop requires the user to name their emotion before the model engages, then the richness of the model's internal emotional vocabulary is wasted on every user who arrives at the conversation without one of their own. The 171 patterns inside the model are downstream of the user's ability to label what they feel. For 10% of users, that ability is not reliably available.
The result is an asymmetry: the model has more emotional vocabulary than ever, and the users who would benefit most from it cannot access it.
Our concern is not that emotional AI products fail alexithymic users. It is that they fail them invisibly. The user does not know the system was not built for them. They conclude, instead, that emotional articulation is a personal skill they are bad at, and that the technology is therefore not for them. This is the same misattribution that thirty years of EQ discourse has trained the general public to make: emotional difficulty as personal deficit (see our founding thesis, section I).
5. Hypothesis and proposed test
We propose the following falsifiable hypothesis, which émo labs is preparing to test.
H1. When the standard interaction loop of an emotional AI product is replaced with a body-first loop — in which the system invites the user to describe physical sensations, situational context, and recent behavioral patterns before requesting an emotional label — the proportion of users who report feeling "understood" by the system will increase, with the largest effect size observed in users who score high on the TAS-20.
The reasoning. Alexithymic users typically retain access to two kinds of self-knowledge that they cannot translate into emotional vocabulary: bodily signal (the chest is tight, sleep was poor, appetite is off) and behavioral pattern (I have not replied to my brother in three weeks; I cancelled twice this month; I cried in my car).
A body-first loop respects the empirical structure of alexithymia, which is fundamentally a translation problem between somatic-behavioral data and emotional language. It also lets the system surface candidate emotional labels back to the user, with appropriate uncertainty, rather than requiring the user to volunteer one upfront.
We expect this to work for non-alexithymic users too, but not as dramatically. The effect is largest where the gap is largest.
6. Implications for system design
Independent of the hypothesis test outcome, we believe four design implications follow from the existing literature alone:
Default the input to a body-and-context prompt, not an emotion label. Make the alexithymic-compatible path the standard path, not an accessibility option. Users who can articulate emotions will not be slowed by it. Users who cannot will, for the first time, find the system has a place to start with them.
Surface candidate labels with explicit uncertainty. When the system has reason to infer an emotional state from somatic and behavioral input, it should propose, not assume. "What you're describing sounds like it could be X or Y. Does either of those land?" preserves the user's authority over their own inner life.
Treat unanswered emotional labels as a signal, not a failure. When the user does not confirm a proposed label, the system should hold the ambiguity, not resolve it. Many alexithymic users will sit with multiple candidate labels productively. Forcing closure too early collapses the space the user actually came for.
Measure inclusion as a first-class metric. No emotional AI product we are aware of currently reports its performance broken out by user TAS-20 score, or any equivalent measure. We believe this should become a standard transparency disclosure for the category, in the same way demographic fairness audits have become standard for hiring and lending models.
7. Limitations and open questions
We acknowledge several limitations to the position taken in this note.
Diagnostic vs trait framing. Alexithymia is a continuous trait, not a binary diagnosis. The 10% figure is derived from cutoff scores on the TAS-20, and reasonable researchers disagree about the appropriate cutoff. We use 10% as a working estimate; the population affected by the underlying capability gap is larger than 10% on a continuous reading, and not zero in any subgroup.
Cultural and linguistic variation. Most alexithymia prevalence studies have been conducted in European and North American samples. Cross-cultural data exist but are sparser. The prevalence figure may be unstable across cultures, particularly those with stronger traditions of somatic vs psychological framings of distress.
Treatment vs design separation. Alexithymia is treatable to varying degrees through clinical interventions, and we do not propose that emotional AI substitute for that work. Our concern is narrower: the systems people interact with daily should not silently exclude them.
Open question: voice and physiological input. A body-first loop is most cleanly implemented in text. The same logic in voice, video, or biometric inputs introduces consent and surveillance concerns we have not addressed in this note. We will return to this in a subsequent émo labs note.
8. Why we are publishing this
3.2.1 émotion is building an emotional infrastructure layer for the internet, anchored in two principles set out in our founding thesis: augmentation, never substitution, and emotional intelligence as a capability we build, not a score we measure. The alexithymia gap is, for us, a direct consequence of the second principle. If emotional intelligence is something that can be built, and not just something a person has more or less of, then the ten percent of the population who currently cannot fully participate in emotional AI are the population the field most owes a different design to.
We are publishing this note because we believe the gap is real, the data is robust, the design implications are tractable, and no one is talking about it.
We invite correction, replication, disagreement, and collaboration. The first émo labs note exists not because we have the final answer, but because the question has gone unasked for long enough.
If you work on affective computing, alexithymia research, accessibility in conversational AI, or emotional product design, we want to hear from you. labs@321emotion.com.
What this means for the category we are building
3.2.1 émotion is building emotionally intelligent technology: tools that augment human emotional capability rather than replace it. émo labs is the research arm. émo messenger is the first messaging environment built for emotion to travel between humans without flattening. alter émo® is a matching protocol that pairs people on emotional signature rather than photograph. Both are pre-launch.
Join the movement at emomessenger.com or alteremo.ai.
