← Back to the site A0 portrait · 841 × 1189 mm
COLOR FISH TEST

The Grue Divide: When Cognitive Accuracy Doesn’t Predict Affective Alignment

Kai Park · Linguistics Department, Georg-August University of Göttingen · kai.hyeyeon.park@gmail.com

EMNLP 2026 Budapest: this talk on Underline
Seeing alike may not mean feeling alike!
In our data, Korean and Australian participants counted color gradients about equally well, yet described the colors in emotionally different ways.

The big idea Each number matches a section below.

It isn’t only whether a language has separate basic color words. How its color words are built, and how often they’re used, matter too. And because color words don’t map one-to-one across languages, feeling needs its own measure.

1The puzzleKorean uses one word, pureuda, for both blue and green. A naive reading of Whorf predicts worse discrimination there. Korean speakers were more accurate.
2Two cluesSpeed: faster only in the blue-to-green gradient, where Korean and English boundaries differ. This matches Winawer’s Russian finding.Accuracy: better across all four gradients. This fits how productive Korean color words are (many derived forms) and how often Grue is used.
3So we measure feeling, tooWords can’t be matched one-to-one across languages. SFF scores every keyword on the same three axes instead, so affect can be compared directly.

Why does this matter?FOR MULTICULTURAL AI

AI speaks many languages, but not every culture.Models trained mostly on English text miss cross-cultural differences in emotional expression and lean toward Western value systems. (Havaldar et al., 2023; Tao et al., 2024)
Translation isn’t enough.Output that reads as culturally generic, not locally situated, is judged lower-utility in that market. Color is a small, tightly scoped place to start.
Accuracy isn’t alignment.A model that gets color categories right can still miss what they feel like. Alignment may need to model perception and affect together.

1The puzzle: why Grue?

A pixel-art landscape: a blue sky above a green forest. Both are labeled with the same Korean word, pureuda.
One word for the blue sky and the green forest. Does that mean Korean speakers can’t tell green from blue? Our data say no.
  • Blue is the anchor. In Obangsaek, the five cardinal colors, blue stands for the East, wood and life; the everyday word for green, chorok, is a Sino-Korean loan.
  • Few forms, most used. In a free-recall study (42,000+ responses), Grue had the fewest derived forms yet the highest response frequency among Korean color categories (Lee and Kim, 1996).
  • Nuance lives in one word. Korean builds shades from a single root with prefixes, suffixes and vowel shifts; English mostly uses phrases.

MHow we tested it

RecruitmentProlific, X, HelloTalk, LinkedIn · 2,264 people
LLM #1: scene generationGemini 2.5 Flash · 3 scenes per color (9 total)
one web session per participant
Study 1: block countingBlue → White, Black, Yellow, Red gradients · 6 trials, 3 levels
Study 2: affect keywords1 scene per color (Blue, Green, Teal) → open-ended keywords
Cohort filterL1 matches country · ≥ 20 yrs residency904 KR / 255 EN · N = 9,534 trials
Cognitive measureserror rate RE · response time
LLM #2: SFF scoringGemini 2.5 Flash rates each keyword → (S, F, Fr)
Study 1 analysisage weighting · MWU, Kruskal–Wallis, Dunn · GLM · TOST + BF
Study 2 analysisper-image SFF means · Euclidean affinity · correlation · K-means
Cognitive vs. affectivecompared by country (e.g., the Australia Divergence)
Orange = where the LLM enters. Adapted from Figure 6 of the paper.

Study 1 stimuli: count the blocks (8–30) in bars running from blue to another cardinal color.

Error per trial = |actual − estimate| ÷ (actual × difficulty level 1–3).

RESULTS

2Seeing: more accurate on every gradientStudy 1 · 9,534 trials · 1,159 participants

0.056median error, KR
(vs. 0.147 EN)
1.16×EN error rate relative to KR, covariates held
d = 0.17small effect, clear with thousands of trials
Median error by gradient lower = more accurate
0.10.20.3
.023
.106
Yellow
0
.031
White
.120
.231
Red
.185
.286
Black
KoreanEnglish-speaking

Darker gradients are harder for everyone; the Korean bar is lower in all four. More than half of Korean White trials were counted exactly (median 0).

Accuracy gap from Korea, adjusted GLM coefficient, 95% CI
−0.4−0.20+0.2+0.4Australiap = .973UKp = .028USp = .039← more accurate than KRless accurate than KR →

With age, gradient and media habits held constant, the UK and US are less accurate than Korea. Australia’s interval sits squarely on zero.

3Feeling: the Australia DivergenceStudy 2 · SFF framework

The nine AI-generated scene images used in Study 2, three per color: blue, green and teal.
Blue #0000FFGreen #008000Teal #008080

Participants described each scene in their own keywords; Gemini 2.5 Flash scored every keyword from −1 to 1 on three axes:

SentimentvalenceFormalitysocial distanceFriendlinesswarmth

Korea ≈ Australia in seeing …

Error rate, KR − AU 90% interval of the difference
“the same” zone, ±0.068−0.1−0.0680+0.068+0.1

The whole interval sits inside the margin: equivalent (TOST p = .003; Bayes factor ≈ 4.6 for “no difference”). Response time was not equivalent (TOST p = .625).

… but not in feeling

How far is Australia’s feeling profile from…? distance in SFF space
Blue
KR0.223
US0.075 ← nearest
UK0.179
Green
KR0.279
US0.167
UK0.061 ← nearest
Teal
KR0.368
US0.477
UK0.315 ← nearest

For every color, Australia’s nearest neighbour is an English-speaking cohort, never Korea.