Phase 1 · Preprint 2
Where One AI Diverges from Plural Human Readings on Short Social Video Clips.
Kim, E. (2026b). Zenodo preprint · DOI 10.5281/zenodo.22253369. Companion to Preprint 1 (Kim, 2026a).
Publication
Open preprint on Zenodo (2 September 2026) under CC BY 4.0. AI runs use Google Gemini 2.5 Pro (snapshot 2026-08-13) with structured JSON output matching the Phase 1 human response protocol.
1 · What the paper asked
Where does one AI sit relative to the plural human distribution?
Preprint 1 documented that human readings on short social video clips are more plural than the single-label evaluation frame assumes. Preprint 2 asks the immediate follow-up: if we hand the same clips to a contemporary multimodal AI — evaluated against the actual response distribution rather than a canonical label — where does it sit? Where does it agree with humans, where does it diverge, and does its divergence sit inside or outside the human plurality?
The AI used is Google Gemini 2.5 Pro (2026-08-13 snapshot). The protocol matches Phase 1's human protocol at input time: same target-person description, no curator metadata beyond what a human reader saw. Three input conditions probe the metadata-leakage question separately.
2 · What it found
AI tracks the human plurality more than the field lets on.
Three descriptive findings on the 32 clips the AI successfully ingested (one clip failed ingestion; that failure is disclosed and discussed in the paper):
Finding i
Where humans agree, the AI agrees
On the eight clips where 80%+ of readers converged on a single emotion (the consensus tier from Preprint 1), the AI picked the same modal emotion on seven of eight clips. The AI is not systematically wrong on the clips where humans are unanimous — it agrees with them.
Finding ii
When the AI hedges, it hedges inside the plurality
On multi-emotion (“mixed”) AI responses, 41 of 49 hedges — 84% — listed constituent emotions that overlapped the human distribution on the same clip. The AI's hedging is not noise; it maps to the human plurality shape.
Finding iii
Where humans split, the AI splits too
Full-corpus divergence from the human modal is 41% — but that divergence is concentrated on clips where humans themselves were split. On the deep-plural tier, the AI is not diverging from a stable human answer; there is no stable human answer to diverge from. That is a different signal than raw error, and Phase 2 is the study built to interrogate it directly.
3 · What it means
A different reading of AI performance on this task.
A single-label scorecard would report the 41% divergence as a straightforward failure. The distribution-aware read is more useful: the AI is doing what the plural human distribution is doing, wherever we look. It agrees on consensus clips, hedges inside the plurality on ambiguous ones, and diverges most where the human answer is itself non-unique.
That reading reframes the evaluation problem. The right question is not “does the AI pick the correct label?” but “does the AI's response distribution mirror the human response distribution?” Phase 2 makes exactly that comparison across four contemporary multimodal models and five prompt conditions.
4 · What came next
From one AI to a calibrated fleet.
Phase 2 (Multi-AI Calibration) is now running. Four models — Gemini 2.5 Pro, Claude Sonnet 4.5, GPT-4o multimodal, Qwen 3.8 27B (open-weight) — run on the same 32-clip corpus across five prompt conditions, from zero-context to human-calibrated. The live counter on the home page shows run progress; the frozen pre-registration lives at /phase-2/preregistration.