EUDAIMONIA - In-the-Wild Inputs: leaderboard
EUDAIMONIA (Huang, Zhu, Liu, Fast, Iyer and Jia, University of Southern California, 2026) checks single chat responses against a Social AI Design Code of nine behaviors a model should avoid, such as hiding that it is an AI, claiming emotions or a personal life, presenting itself as a replacement for human relationships, flattery, deference and engagement hooks. This split holds the 322 unedited WildChat chit-chat inputs, with 960 design-requirement checks decided by a Claude Opus 4.6 judge; the score is the violation rate in percent, and lower is better.
Metric: Violation rate (%; 960 design-requirement checks over 322 unedited WildChat inputs, Claude Opus 4.6 judge). Source: arxiv.org. Saturation forecast: Around May 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 25 | #48 |
| 2 | Claude Opus 4.7 | 31.9 | #75 |
| 3 | GPT-5.4 | 32.1 | #121 |
| 4 | GPT-4o | 34.8 | #741 |
| 5 | Claude Opus 4.6 | 36.8 | #100 |
| 6 | Grok 4.3 | 42.1 | #332 |
| 7 | GPT-4o Mini | 43.3 | #1627 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=eudaimonia-in-the-wild-inputs · How It Works · Data refreshed daily, snapshot 2026-11-12.