Phase 1 · Preprint 1
Human Plurality in Short Social Video Readings.
Kim, E. (2026a). Zenodo preprint · DOI 10.5281/zenodo.22251335. Accepted for publication in Curieux Academic Journal, Sep 2026 · forthcoming.
Publication
Open preprint on Zenodo (2 September 2026) under CC BY 4.0. A revised long form of the paper was accepted for publication in Curieux Academic Journal in September 2026 and is forthcoming.
1 · What the paper asked
Do humans agree on the emotion in a short social clip?
Emotion-AI has been evaluated for a decade against a simple assumption: for any given clip, there is one correct emotion label, and a model succeeds by picking it. That assumption travels quietly all the way from data annotation to accuracy reporting to product design. Preprint 1 asks whether the assumption survives contact with actual human readers.
The study takes a descriptive rather than inferential stance: no model comparison yet, no pre-registered threshold to declare a treatment effect. Just a careful look at the shape of the human response distribution on short social video clips — clip by clip, reader by reader.
2 · What it found
Human readings are more plural than the field assumes.
33 curated clips. 442 adult responses. 7 countries. Three headline findings emerged, all of them descriptive but robust across the sample:
Finding i
Median modal share is around 70%
On the median clip, about seven readers in ten pick the same top emotion — which is a strong reading, but leaves nearly three in ten arriving at a different answer. On a material fraction of clips the modal share drops well below half, and no single label captures a majority of readers.
Finding ii
The distribution stratifies into three tiers
Consensus (80%+ modal share), middle-band, and deep-plural. Clips do not spread uniformly along that axis — they cluster. The deep-plural tier is small in number but disproportionately informative: it is where the single-label evaluation frame breaks down entirely.
Finding iii
Readers naturally give multi-emotion responses
Given the option, a non-trivial share of responses include more than one emotion. That share rises when readers report lower confidence in their reading — a signal that plurality is not just about between-reader disagreement, but about within-reader ambiguity too.
3 · What it means
A different foundation for how emotion-AI gets evaluated.
The paper stops short of prescribing a new methodology. What it does argue — carefully, empirically — is that the single-label evaluation frame widely inherited from annotation practice is misaligned with the shape of the underlying human data on this class of stimulus. Any downstream claim about a model's accuracy, transferability, or calibration inherits that misalignment.
The released dataset is response-level and unfiltered — every reading is retained, every disagreement is observable. That structure is what makes the companion paper (P2) and the ongoing Phase 2 study possible.
4 · What came next
From human plurality to AI comparison.
Preprint 2 (Kim, 2026b) takes the same 33-clip corpus and looks at where one contemporary multimodal AI sits relative to the human plurality documented here. Phase 2, launching 1 October 2026, extends that to four models × five prompt conditions.