Do frontier AIs read emotion the plural way humans do?

The current research asks

Phase 1’s finding that humans don’t fully agree led to the Phase 2 question — what about AI?

1Do the AIs read plurally too?
2Does the prompt matter?
3Which model comes closest to humans?

What emotion do you read?

HappyMovedSurprisedConfusedAnxious+ more than one

Confidence

middle

Timeline · Phase 2

Sep 2026· done
Phase 1 preprints posted · Phase 2 design frozen. Pre-registration →
Oct 2026· now
Runner kickoff · all four models in parallel. Live analysis →
Nov 2026
Sustained collection · within-cell reliability checks.
Dec 2026
Analysis · preprint drafting · target close 12/31.

Live JSD grid · 4 models × 32 clips

Progress

617/ 1,920· 32%

· last run 9h ago

GeminiGemini 2.5 Pro

ClaudeClaude Sonnet 4.5

GPT-4oGPT-4o

QwenQwen 3.8 27B

<0.120.12–0.240.24–0.300.30–0.400.40–0.500.50–0.600.60+waitingLive · updating as runs complete

ModelsGemini 2.5 Pro · Claude Sonnet 4.5 · GPT-4o · Qwen 3.8 27B

Each row is a model; each column is one clip. Cell color = that model's Jensen–Shannon divergence from the plural human reading on that clip. Greener = closer to humans.

Overview Pre-registration

Completed Research·2024 – Aug 2026

Four years ago, a small question — how hard it is to read someone else’s emotion — turned out to be much larger.

Phase 0 · 2024 – 2025

What AI tools for emotional-intelligence learning already exist — and where do they fall short?

KIM Evelyn, 2025

Artificial Intelligence in Emotional Intelligence Training for Autism

Origin paper

Phase 1 · 2026

How do people read emotion plurally — across cultures, on the same clip?

KIM Evelyn, 2026a

Human Plurality in Short Social Video Readings