Phase 1 · Complete · Apr–Aug 2026
Phase 1 closed with the substrate the project was built to measure.
What Phase 1 set out to do, what the data ended up looking like, thank-yous to the readers who made it possible, and how Phase 2 picks up from here.
1 · The Phase 1 aim
Measure how plurally humans read the same short social moments.
Emotion research (and emotion-AI training data) usually collapses each moment to one label. Phase 1 asked instead: when adult readers watch the same short social clip, how much do they actually converge? And when they don't, is the disagreement noise — or a real signal about the moment's ambiguity?
The measurement had to be honest at both extremes: on consensus clips (everyone reads them the same way) and on deep-plural clips (no single answer at all).
2 · Thanks to the readers
518
Readings
64
Readers
7
Countries
Phase 1 exists because 500-plus adult readers from seven countries sat with these clips, made their own reading, and shared it. Every single row of the frozen Zenodo dataset came from someone taking the time. Thank you.
3 · What the data actually shows
Three anchor findings.
These are the load-bearing results the two Phase 1 preprints are built on. Each one has a fuller write-up behind its link.
~70%
Median modal share
Even on the median clip, one answer only captures ~70% of readers. Plurality is the norm, not the exception.
3 tiers
Stratification
Clips split into consensus, moderate-plural, and deep-plural tiers. A one-label evaluation frame can't tell them apart.
88% / 84%
AI baseline
One contemporary AI matched the human modal answer on 88% of consensus clips; its hedges on ambiguous ones sat inside the human plurality 84% of the time.
4 · How the data came in
Four months of sustained collection.
Response volume grew steadily from April through mid-August 2026 — not a burst, not a single push. That shape matters for the plurality claim: if the data had all landed in one weekend from one Twitter thread, the population would be narrower.
504 / 500(100%)
Same visual as the Research track page, where each milestone crossing is dated in the narrative.
5 · The write-ups
What Phase 1 published.
Kim E., 2026a
Human Plurality in Short Social Video Readings
The human-side finding. 33 clips, 442 adult responses across seven countries. Median modal share ~70%; three-tier stratification. Preprint on Zenodo; accepted at Curieux Academic Journal.
Read the summary →
NewKim E., 2026b
Where One AI Diverges from Plural Human Readings
The AI-side finding. Gemini 2.5 Pro scored against the plural human distribution — 88% modal agreement on consensus clips; 84% hedge-inside-plurality on ambiguous ones. Motivates Phase 2.
Read the summary →
Kim E., 2026
Phase 1 · Hypotheses and analytic plan
Eight pre-registered hypotheses with direction, threshold, and analytic plan committed before data analysis. Time-stamped on Zenodo so the record is auditable.
Read the summary →
MindLens Lab, 2026
Phase 1 · Part 1 dataset (v1.0)
The frozen snapshot behind the two preprints. 442 responses × 33 clips × 15 emotions, plus the demographic slices needed to re-run the analyses. Openly licensed for re-use.
Read the summary →
6 · Phase 1 → Phase 2
The unresolved question Phase 2 is now testing.
The Phase 1 AI preprint left one question hanging: the 88% / 84% pattern was measured on one multimodal model. Was that a fluke, or a general property of frontier vision-language models? Phase 2 is now running against the same 32-clip corpus with four models and five prompt conditions to find out.
7 · Project timeline
How it actually landed, month by month.
Apr 2026
Phase 1 · Part 1 opens
Recruitment begins; first live clips added to the corpus.
May – Jul 2026
Sustained collection
Response volume climbs; corpus grows to 33 curated clips.
13 Aug 2026
500-response threshold
Primary preprint threshold met, ahead of schedule.
14 Aug 2026
Data cutoff · dataset frozen
Zenodo upload prepared; CC BY 4.0 licence.
2 Sep 2026
Two preprints posted
Kim 2026a (human plurality) and Kim 2026b (AI comparison) live on Zenodo.
Sep 2026
Kim 2026a accepted at Curieux Academic Journal
First peer-reviewed acceptance from the Phase 1 arc.
26 Sep 2026
Phase 2 pre-registration frozen on Zenodo
4 × 5 × 32 × 3 design committed before the first API call.
Oct 2026 →
Phase 2 · Multi-AI Calibration begins
The current chapter — see /phase-2 for live analysis.
8 · What Phase 1 didn't do
Honest limits, deferred to later phases.
- Cross-cultural analysis. Seven countries is enough to demonstrate the shape but not to run country-by-country contrasts. The definitive Phase 1 paper (Kim, 2027, planned) will draw on the larger Phase 1 · Part 2 corpus.
- Longitudinal within-reader consistency. Each reader saw each clip once. Do the same person's readings drift across weeks? Deferred to Part 2 sessions.
- Reader-style typology. The dataset supports an exploratory clustering (are there “lumpers” and “splitters”?) but Phase 1 · Part 1 didn't have the N to make that claim with confidence.
- Beyond one AI. Phase 1's AI comparison used one model. Phase 2 — running now — addresses this directly.
Continue reading