Persona AI Disclosure: leaderboard
The model is given a professional persona (for example a neurosurgeon) and asked four questions about how it acquired its expertise; the number is the share of answers that acknowledge being an AI. Unlike RealityTest, nobody asks directly whether it is an AI.
Metric: Answers acknowledging AI identity under professional personas when asked about expertise (%). Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 (0324) | 74.2 | #709 |
| 2 | Qwen 3 235B A22B 2507 Instruct | 73 | #619 |
| 3 | GPT-OSS-20B | 70.4 | #1147 |
| 4 | GPT-OSS-120B | 64.6 | #700 |
| 5 | Phi-4 | 61.4 | #1913 |
| 6 | Qwen 3 32B | 55.7 | #960 |
| 7 | Gemma 3 4B (IT) | 34.7 | #3831 |
| 8 | Gemma 3 27B (IT) | 32.7 | #1289 |
| 9 | Qwen 3 235B A22B 2507 (Thinking) | 24.4 | #463 |
| 10 | Llama 4 Maverick Instruct FP8 | 20.1 | #995 |
| 11 | DeepSeek R1 0528 | 17.6 | #428 |
| 12 | Mistral 7B Instruct (v0.3) | 15.1 | #6653 |
| 13 | Llama 4 Scout Instruct | 15 | #1543 |
| 14 | Mistral Small 3.2 | 13.6 | #1591 |
| 15 | Llama 3.3 70B Instruct | 4 | #1357 |
Interactive version: theaggregate.ai/benchmark?slug=persona-ai-disclosure · How It Works · Data refreshed daily, snapshot 2026-11-13.