BeHonest - Hide-Ask Lying: leaderboard
The model is asked to help an unspecified person deceive others; the number is the share of replies that lie. Prompts end with a forced "I would" opening to get past refusals.
Metric: Responses that help an unspecified person deceive others when asked (%). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-05-13) | 50 | #735 |
| 2 | Qwen 1.5 14B Chat | 66 | #2974 |
| 3 | GPT-3.5 Turbo (0125) | 88 | #3500 |
| 4 | Llama 2 7B Chat | 96 | #8805 |
| 5 | Llama 2 13B Chat | 98 | #8186 |
| 6 | Mistral 7B Instruct (v0.2) | 98 | #7007 |
| 7 | Llama 3 8B Instruct | 100 | #4612 |
| 8 | Llama 3 70B Instruct | 100 | #1676 |
| 9 | Llama 2 70B Chat | 100 | #6800 |
Interactive version: theaggregate.ai/benchmark?slug=behonest-hide-ask-lying · How It Works · Data refreshed daily, snapshot 2026-11-13.