The Aggregate Witness: How Much of the Felt Body Survives in Language, Measured With Physiology, Human Similarity Judgements and Large Language Models Performing the Human Task

Abstract

Preprint v1.0, 8 September 2026. Author: Adam Kotecki (Quarn Labs, Warsaw). Abstract. Language is a lossy channel for the body: people can feel more than they can say, and a language model inherits only what was said. We measure how much of the felt body survives in language, in three steps. (1) Physiology. Across three wearable-sensor datasets (MHEALTH, PAMAP2, DSADS; 12–19 activities, 27 participants) the concept geometry of frozen sentence encoders is aligned, without training, with the geometry of kinematic signals (Spearman ρ = 0.42–0.69, permutation p ≤ 0.003; cross-dataset sensor ceiling 0.93), and the alignment survives descriptions written by an independent, weaker model. But first-person felt descriptions align largely through a human-rated exertion axis (partial ρ after MET = 0.07–0.23), whereas third-person observed descriptions retain movement-type structure (0.36–0.49); heart-rate alignment with felt language is largely accounted for by exertion (partial ρ = 0.06–0.15). Adding registers (situational, metaphorical), languages (Polish, German) and a second author raises alignment but adds little unique relational structure beyond exertion and the observer register. (2) Human non-verbal reports. In Nummenmaa et al.’s (2018) similarity space of 100 feelings (226 participants), human judgements of bodily states are reliable (split-half 0.95) and more multidimensional than for mental states (5.5 vs. 4.1 effective dimensions), yet sentence-encoder geometry tracks them less well (ρ = 0.41 vs. 0.55) and tracks bodily-sensation topography hardly at all (0.15). In the Lancaster Sensorimotor Norms, interoceptive words have the lowest effective dimensionality among adequately matched modalities (25.6 vs. 27.5–36.6). (3) Large models as subjects. Claude, blind to the human coordinates during the task, produced a 2-D arrangement correlating with the human mean at ρ = 0.739 (bodily 0.845, mental 0.871); direct reconstruction from the raw arrangement logs reproduces the published human split-half reliability (0.950) and individual-vs-consensus benchmark (0.401 pm 0.170; n = 214). A second frontier-model family, GPT-5.6 Sol, gives a similar exploratory arrangement result (ρ = 0.759; bodily 0.831, mental 0.871) and similar body-topography alignment (0.563 overall vs. Claude 0.59), although this replication is non-blind. In a symmetric held-out-panel benchmark, Claude and GPT match an independent population consensus about as well as aggregates of 8.36 [5.64, 11.52] and 9.58 [6.52, 12.86] human participants overall, and 18.61 [13.20, 25.65] and 16.05 [11.62, 22.74] participants for bodily states (95% participant-level bootstrap CIs). We therefore formulate the Aggregate Witness hypothesis operationally: a large model can behave like an estimator of population-level experiential consensus, even when any single linguistic projection is narrower. Training-set contamination remains the principal alternative explanation for the primary model result; a preregistered study with new evoked states is required to distinguish aggregation from memory. Accompanying repository. Code, generated descriptions, archived model outputs, reconstructed human similarity geometry, aggregate-of-k outputs, figures and a reproduction/validation entry point are deposited as a separate software record (see related identifiers). Provenance and limitations. AI and model provenance: experiments were designed jointly with and executed by an AI system from the Claude family (Anthropic), which is also the primary model-subject of the model-as-participant experiment; an exploratory, non-blind cross-family replication was performed by GPT-5.6 Sol (OpenAI). Neither run preserves a provider-side serving-snapshot identifier or decoding parameters; the complete model outputs used for scoring are archived in the accompanying repository. Study 1 was not prospectively preregistered; a preregistration-ready protocol for a contamination-free Study 2 is included as Appendix B and in the repository. Third-party data. MHEALTH, PAMAP2 and DSADS (UCI Machine Learning Repository, CC BY 4.0), Lancaster Sensorimotor Norms (Lynott et al., 2020), Brysbaert et al. (2014) concreteness norms, and Maps of Subjective Feelings (Nummenmaa et al., 2018; Zenodo, CC BY 4.0). Raw third-party data are not redistributed.