Where One AI Diverges from Plural Human Readings on Short Social Video Clips

In a companion paper (Kim, 2026a — DOI: 10.5281/zenodo.22251335), it was documented that on short social video clips, human readers do not converge on a single emotion: on the median clip, the most-picked emotion draws about seven readers in ten, and about one in four responses explicitly reports multiple simultaneous emotions. This paper compares one AI's readings (Google's Gemini 2.5 Pro) to the human distribution on the same 33 clips, using human modal emotions derived from 43 adult participants. In the single-run analysis across the full 32-clip clean-condition corpus, Gemini's top-picked emotion diverges from the human modal on 13 of 32 clips (40.6%). To separate stable AI behavior from single-run sampling variability, a stochasticity check was conducted on three clips: 3 conditions × 7 iterations × 5 independent runs = 315 total observations. First, metadata reduces modal alignment on the three selected clips consistently across all five runs — a −12-percentage-point shift. Second, AI hedging is a reproducible behavior on these clips; on 84% (41/49) of mixed responses whose full constituent lists were recorded, the human modal appeared among the AI's listed constituent emotions. Third, the specific mechanism initially claimed in single-run data ("metadata suppresses hedging") did not replicate under repeated sampling; the actual mechanism is per-clip variable. The paper also documents two methodological observations: one clip could not be ingested by any tested AI via YouTube-URL programmatic access (a substantive methodological signpost about corpus bias), and emotion-AI evaluations should report repeated-run stability alongside single-run pattern claims. Companion paper: Kim, E. (2026a). Plural Emotion Readings in Short Social Video Clips: A Descriptive Study. Zenodo. https://doi.org/10.5281/zenodo.22251335 Open dataset: https://www.mindlenslab.org/live-findingsPre-registration: https://www.mindlenslab.org/research/hypothesesAnalysis code: https://github.com/evelyn-mindlens/preprint-analysis (MIT license)Project site: https://mindlenslab.org