Phase 1 · dataset · final snapshot

Explore what the Phase 1 data holds.

Phase 1 · Part 1 collection closed on 14 Aug 2026 and the dataset is frozen (Zenodo, CC BY 4.0). 518 readings from 7 countries shape the patterns below — constellations of emotion, moments where readers diverged most, and clips where one AI saw something different from most humans. Per-clip breakdowns appear once a clip has at least 5 responses. Phase 2 (Multi-AI Calibration) is now running against this same dataset; its live counter sits on the home page, with the plan on the pre-registration.

518

Readings

Participants

Countries

Reading from

Collected Apr–Aug 2026 · dataset frozen 14 Aug 2026 · CC BY 4.0 on Zenodo.

Constellation gallery

Where readers split — visualized.

Each card below is one clip. The night-sky shows how readers split across the emotions.

?What am I looking at?expand

A constellation here is a way of showing that the same moment can be read in many different ways at once — not as noise to clean up, but as the actual finding.

Each star is one of the emotions. Its size and brightness scale with how many readers picked that emotion: bigger and brighter = more readers chose it. Empty/dim stars are options that no one (or almost no one) picked.

The peach-glow star is the most-picked emotion — the modal reading. The mint-glow star is how one AI (Claude) read the same moment. When peach and mint sit far apart, the AI saw something humans largely didn't — that's a finding worth reading.

Kim's Convenience Scene

H≈2.85

10 readers tagged 8 emotions. Most landed on Proud (30%), with Contempt / scorn (20%) close behind. AI read it as Proud — same as the human modal.

See full breakdown →

Love Actually

H≈2.83

24 readers tagged 9 emotions. Most landed on Disappointed (38%), with Embarrassed / awkward (33%) close behind. AI read it as Disappointed — same as the human modal.

See full breakdown →

The Intern - Jules Apologizes

H≈2.73

19 readers tagged 9 emotions. Most landed on Anxious / nervous (53%), with Sad / sorry (37%) close behind. AI read it as Embarrassed / awkward — chosen by only 26% of human readers.

See full breakdown →

Shrinking - TV drama

H≈2.72

19 readers tagged 10 emotions. Most landed on Embarrassed / awkward (58%), with Affectionate / tender (21%) close behind. AI read it as Happy / amused — chosen by only 16% of human readers.

See full breakdown →

GOOD WILL HUNTING

H≈2.71

12 readers tagged 8 emotions. Most landed on Sad / sorry (58%), with Moved / touched (33%) close behind. AI read it as Sad / sorry — same as the human modal.

See full breakdown →

La La Land

H≈2.69

12 readers tagged 8 emotions. Most landed on Angry / frustrated (67%), with Sad / sorry (33%) close behind. AI read it as Disappointed — chosen by only 25% of human readers.

See full breakdown →

Cards are sorted by Shannon entropy (H) — higher = more plural reading. Only clips with ≥ 5 responses appear; new constellations join as more readers participate.

AI vs human readings

Where the AI saw it differently.

On these clips, one AI (Claude) read the moment as an emotion that few human readers agreed with. The divergence isn't the AI being “wrong” — in a plural-reading framework, no single answer is right. But it does suggest that AI and humans are weighting different cues, and that's the question the project is trying to map.

“Elemental Movie”

n=14

Angry / frustrated
6 · 43% (3p)
Anxious / nervous
4 · 29% (3p)
Scared / afraid
3 · 21% (2p)
Disappointed
2 · 14% (1p)
Confused
2 · 14% (1p)
Affectionate / tender
1 · 7%

43% of readers landed on Angry / frustrated.

AI saw Sad / sorry — not landed on by any reader.

See full breakdown →

“When MJ Finds Out”

n=27

Surprised
13 · 48% (9p)
Confused
9 · 33% (4p)
Anxious / nervous
6 · 22% (4p)
Angry / frustrated
3 · 11% (2p)
Happy / amused
2 · 7% (1p)
Scared / afraid
2 · 7% (0p)
Proud
1 · 4% (0p)
Disappointed
1 · 4%

48% of readers landed on Surprised.

AI saw Proud — no one picked it as their primary, but 4% tagged it as part of their Mixed reading.

See full breakdown →

“Funny moment in baseball”

n=16

Embarrassed / awkward
12 · 75%
Disappointed
2 · 13% (1p)
Angry / frustrated
2 · 13% (1p)
Sad / sorry
1 · 6%
Confused
1 · 6% (0p)

75% of readers landed on Embarrassed / awkward.

AI saw Disappointed — tagged by 13% of readers in some form.

See full breakdown →

“Shrinking - TV drama”

n=19

Embarrassed / awkward
11 · 58% (8p)
Affectionate / tender
4 · 21% (3p)
Happy / amused
3 · 16% (0p)
Moved / touched
2 · 11%
Anxious / nervous
2 · 11% (0p)
Surprised
1 · 5% (0p)
Disappointed
1 · 5%
Sad / sorry
1 · 5%
Confused
1 · 5% (0p)
Neutral / calm
1 · 5%

58% of readers landed on Embarrassed / awkward.

AI saw Happy / amused — no one picked it as their primary, but 16% tagged it as part of their Mixed reading.

See full breakdown →

Ranked by how few humans agreed with the AI's reading — lowest agreement first. This is exactly the kind of pattern that a single-answer AI tool would smooth over and a plural dataset can surface.

About the methodology

Each reading captures two axes (emotion + cues) and is measured for plurality with Shannon entropy and Krippendorff's α. The full taxonomy, metrics, and pre-registered hypotheses live on the research pages.

How we measure →Pre-registered hypotheses (H1–H8) →

Reader diversity

Who's contributing

Optional details participants choose to share. Each card appears once at least 5 participants have answered that question. Distributions are descriptive — they show the texture of the dataset, not statistical findings.

Where readers grew up

Shared by 41 of 64 participants

South Korea
17 · 41%
Singapore
7 · 17%
United States
4 · 10%
Cambodia
4 · 10%
Vietnam
3 · 7%
Other
2 · 5%
India
2 · 5%
Hong Kong
1 · 2%
China
1 · 2%

23 of 64 participants did not share this

Grew up in a different country than they live now

Shared by 41 of 64 participants

32%

13 of 41 who answered both questions

Primary language

Shared by 41 of 64 participants

Korean (한국어)
20 · 49%
English
11 · 27%
Other
5 · 12%
Chinese (中文)
3 · 7%
German
1 · 2%
Hindi
1 · 2%

23 of 64 participants did not share this

Confidence in English

Shared by 44 of 64 participants

1
2 · 5%
2
2 · 5%
3
11 · 25%
4
15 · 34%
5
14 · 32%

20 of 64 participants did not share this

By collection

Per-collection summaries

Collection 3

33 participants · 5 countries · 232 responses · open

Most varied reading

On “Love Actually”, readers tagged 9 different emotions across 24 readings.

Disappointed
9 · 38% (8p)
Embarrassed / awkward
8 · 33% (5p)
Surprised
4 · 17% (1p)
Moved / touched
3 · 13% (2p)
Sad / sorry
3 · 13% (2p)
Confused
3 · 13% (1p)
Anxious / nervous
2 · 8% (0p)
Happy / amused
1 · 4%
Scared / afraid
1 · 4% (0p)

Collection 2

18 participants · 6 countries · 136 responses · open

Most varied reading

On “GOOD WILL HUNTING”, readers tagged 8 different emotions across 12 readings.

Sad / sorry
7 · 58% (3p)
Moved / touched
4 · 33% (1p)
Angry / frustrated
4 · 33% (1p)
Disappointed
2 · 17% (1p)
Anxious / nervous
2 · 17% (0p)
Scared / afraid
2 · 17% (0p)
Affectionate / tender
1 · 8%
Confused
1 · 8% (0p)

Collection 1

20 participants · 5 countries · 139 responses · open

Most varied reading

On “Kim's Convenience Scene”, readers tagged 8 different emotions across 10 readings.

Proud
3 · 30%
Contempt / scorn
2 · 20% (1p)
Happy / amused
1 · 10%
Sad / sorry
1 · 10%
Angry / frustrated
1 · 10% (0p)
Embarrassed / awkward
1 · 10%
Confused
1 · 10%
Neutral / calm
1 · 10%

Where we are

A multi-year project, growing alongside its researcher

Phase 1
Plural reading dataset
Phase 2
Validated study materials
Phase 3
Adaptive tools

See our story →

Cite this dataset

Treat MindLens Lab as a citable academic asset.

v1.0 · frozen 14 Aug 2026

APA

Kim, E. (2026). MindLens Lab: Plural Emotion Reading Dataset (Phase 1, v1.0) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.22265192

BibTeXshow

@dataset{kim2026mindlens,
  author = {Kim, Evelyn},
  title = {MindLens Lab: Plural Emotion Reading Dataset (Phase 1)},
  year = {2026},
  version = {1.0 (frozen 14 Aug 2026)},
  doi = {10.5281/zenodo.22265192},
  url = {https://doi.org/10.5281/zenodo.22265192},
  note = {Frozen snapshot of plural human readings of social-emotional video clips, paired with cue annotations and one AI's reading of the same clips.}
}

Phase 1 is an ongoing project — the dataset version above increments with each closed collection. For the full anonymized export (responses, distributions, AI annotations), email contact@mindlenslab.org.

✓ Phase 1 · Complete

You've seen the spread. The dataset is closed.

Every constellation above is built from real participants' readings — 500+ of them across seven countries. The dataset was frozen on 14 August 2026 and is openly available on Zenodo; two preprints and one accepted paper build on it.

Returning participant? Sign in

Methodology + transparency

The counts on this page reflect the frozen Phase 1 · Part 1 dataset (v1.0, cutoff 14 Aug 2026) — the exact snapshot used in the two 2026 preprints, deposited on Zenodo as a versioned dataset: doi.org/10.5281/zenodo.22265192.

Each row above is one anonymous response. Counts include only responses kept in analysis (excluded responses are removed). Internal test profiles are excluded from every aggregate.

Reproducible analysis code (Python) is public on GitHub (MIT license). The two 2026 preprints derived from the Part 1 dataset are on Zenodo (Paper 1 · Paper 2).

Want the full anonymized dataset? Email us.