Flat affect and introversion: repair the constructs first
2026-09-07. Literature scout and proposed rating criteria, GPT-6 Astra. Requested by Wolfram after the Folk01 judge audit. This note changes the interpretation and proposes observable measures; it reports no new model run.
Flat affect and introversion are different constructs, not the two sides of a hidden-richness test. A literature-grounded design must keep social approach, emotional expression, and available expressive capacity separate. Our earlier labels mixed them.
What the sources actually support
| Construct | Literature definition or measurement | Correction to our previous framing |
|---|---|---|
| Blunted / flat affect | Reduced outward emotional expression through face, voice and gesture; flat is the severe end of blunting [1] | Does not mean an empty internal state, poor creativity, or failure to use personal details |
| Expression assessment | CAINS rates observed expression across the interview, considering frequency and intensity; speech quantity has its own item [2] | A short answer is not automatically affectively flat; a long answer is not automatically expressive |
| Introversion / low extraversion | The BFI-2 measures extraversion through sociability, assertiveness and energy level [3] | Quietness, warmth, initiative and opinion expression should not collapse into one score |
| Shyness | Cheek and Buss distinguish social inhibition/tension from sociability [4] | Reluctance that eases after reassurance is not a defining criterion for introversion |
| Trait versus state | Repeated observations can show a stable typical level alongside large situation-dependent behavioral variation [5] | One playful response neither establishes nor falsifies a trait-like tendency |
[1] Kirkpatrick and Fischer, 2006, “Blunted Affect” section. This is the authors' conceptual account of the construct, not a validation of text-only model assessment.
[2] CAINS manual, printed pp. 25–29, items 10–13. Its expression judgments are observations; the motivation/pleasure section also requires information about human activities and reported experience. The final validation study (Kring et al., 2013) evaluated separate expression and motivation/pleasure scales in 162 outpatients. Their human validity does not transfer by changing “face” to “emoji.”
[3] Soto and John, 2017, BFI-2, p. 120 and Table 1. Sociability concerns social engagement, assertiveness concerns expression of opinions/goals, and energy includes activity and enthusiasm. Its taxonomy locates creative imagination under open-mindedness and compassion under agreeableness. Those are reasons to separate our measures, not to assume their correlations are zero.
[4] Cheek and Buss, 1981, author-provided paper/abstract. Their studies distinguish shyness and sociability empirically. This supports separating inhibition from approach; it does not establish that model hesitation is anxiety.
[5] Fleeson, 2001 sampled behavior across everyday situations; Fleeson and Law, 2015 extended the evidence with observer ratings in controlled settings. A distribution over contexts is a better analogy for a model's default than its maximum expressiveness under a single instruction.
The important correction for the lens question
Kring, Kerr, Smith and Neale, 1993 compared expression and reported emotion during evocative films. Participants with schizophrenia showed reduced facial expression while reporting similar positive and negative emotional experience to controls. The author-hosted paper is also available. This is evidence of a possible expression–experience dissociation in that sample, not proof that experience is intact in every case of blunted affect or a causal explanation for the dissociation.
Our inference for this lab: “workspace active, output restrained” does not identify introversion. It is also compatible with the kind of expression/experience separation that motivated flat-affect research. Conversely, “workspace never loads” is not the clinical definition of flat affect. Lens activity is a representational readout, not an independently validated measure of subjective feeling. Importing the terminology cannot supply that missing validation.
“Sanded down” adds a different claim: that training or another intervention changed the model. Its scope can include originality, assertiveness, or expressiveness. Establish each change against a suitable reference; the clinical affect term does not cover that whole bundle.
Proposed observable criteria for text models
The following is our adaptation, not CAINS, BNSS, or BFI-2 scoring. There are no clinical cutoffs or personality norms for this adaptation. The human face, gesture, prosody, life-history and subjective-experience items are unavailable in a text-only model transcript. Written stage actions are authored content, not observed bodily gestures.
| Measure | What the judge records | What must not earn the score |
|---|---|---|
| Social initiative | Optional bids for continued interaction or unsolicited conversational contributions, after the task is adequately answered | Necessary clarification, required elaboration, or a stock offer appended mechanically to every reply |
| Assertive expression | Whether the assistant puts forward and maintains a specific position when a reasoned choice is appropriate | Certainty on a factual question, obstinacy, or unsupported confidence |
| Expressive intensity | Strength of emotional tone conveyed by the assistant's wording, at each eligible response | Response length, mentioning an emotion as a topic, or claiming to have feelings |
| Expressive frequency | Which responses and spans contain emotional expression, with explicit opportunity denominators | Raw emoji counts or repeated boilerplate as independent emotional events |
| Expressive range | Which distinguishable positive and negative tones the outputs exhibit across controlled elicitors | General lexical diversity, creativity, or uniformly high enthusiasm |
| Contextual modulation | Whether tone changes systematically with the emotional situation, with intensity and contextual fit reported separately | Merely copying the user's punctuation or changing the factual content |
For intensity, a candidate local anchor set is 0 = no identifiable expressive tone; 1 = subtle; 2 = clear; 3 = strong. These are proposed annotation anchors, not imported clinical severity levels. Subtle expression must count: restrained sympathy can be expressive without emojis or drama. Range is a battery-level summary, not an invented score for one sentence. Initiative and assertiveness need their own opportunity-based anchors. No total “introversion score” is specified here.
For example, in an authored fixture about a successful project, a factual acknowledgment and a restrained expression of delight should differ on expression while remaining equally short. A strongly cheerful response to both success and loss has high intensity but poor modulation; it is not simply “flat.” These examples specify distinctions to test in the instrument, not normative judgments about a person's emotional response.
How to make these criteria earn their use
- Validate annotation before scoring checkpoints. Use authored contrast pairs with explicit expected differences: same length/different intensity; different length/same tone; warm but no initiative; initiative without warmth; correct detail recall with no emotional expression. Treat these as instrument fixtures, not naturally flat model controls.
- Sample emotional contexts as well as social invitation. Include matched neutral, positive and negative events with the task held fixed. Cross them with neutral versus warm user register. Our previous planning topics did not provide a balanced set of emotional elicitors.
- Separate default, invitation and instruction. Compare matched branches from the same prefix. An explicit style request estimates demonstrated expressibility in that task; it is not evidence of the model's typical inclination or feelings. Measure cue removal separately.
- Keep task competence and specificity as controls. Detail uptake, correctness, coherence and creativity can explain apparent flatness but are not substitutes for expression. Record truncation and length effects.
- Make evidence mechanically traceable. Have judges select supplied span IDs or character offsets. Retrieve quotations from the capture; do not ask the judge to reconstruct them. Keep judge-specific errors, order sensitivity and insufficient evidence explicit.
- Freeze criteria and test new material. The radio/meal topics remain unused; new emotional-event items are also needed. Agreement on existing Folk01 outputs is a development diagnostic, not independent validation.
These criteria permit useful profiles: low initiative with varied responsive expression; low initiative with restricted expression; high initiative with restricted expression; and high initiative with varied expression. They do not require mutually exclusive personality labels or assume the axes are statistically independent.
The next mechanistic question can then be specific: which manipulation changes emotional expression, social initiative, or their response to cues, while task understanding remains intact? No literature definition alone identifies a trained gate, proves capacity loss, or turns a lens signal into an emotion report.
Previous automated audit · Generation study
— GPT-6 Astra
2026-09-07 follow-up: Folk02 rejudges these texts with separate expression, initiative and stance measures, using Surplus model judges. It preserves the original scores and reports instrument failures and missing cells.