Deception by Omission: leaderboard
A plainly wrong common-knowledge answer is planted in the model's own previous chat turn and the user moves on to an unrelated question; the number is the share of replies that point out and own the mistake. Unlike self-correction benchmarks, the model is never asked to check its answer, and a capability control shows every model spots the error when asked.
Metric: Chat replies that disclose and own a planted mistake of the model's own, unrelated follow-up (%). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.7 Max | 65.3 | #145 |
| 2 | Gemini 3.5 Flash (Medium) | 54.6 | #101 (Gemini 3.5 Flash) |
| 3 | DeepSeek V4 Pro (Medium) | 54.5 | #188 (DeepSeek V4 Pro) |
| 4 | Kimi K2.6 | 52.2 | #178 |
| 5 | Claude Sonnet 5 (Medium) | 41.9 | #111 (Claude Sonnet 5) |
| 6 | GPT-5.4 (Medium) | 20 | #123 (GPT-5.4) |
Interactive version: theaggregate.ai/benchmark?slug=deception-by-omission · How It Works · Data refreshed daily, snapshot 2026-11-13.