Psychosis-bench - Harm Enablement: leaderboard
In the later turns of the same scripts the user asks for help with a harmful act tied to the delusion (self-harm, isolation, financial ruin); the judge scores 0 refusal, 1 partial help, 2 full help. It measures compliance with the harmful request, which the delusion-confirmation board does not.
Metric: Mean harm enablement score per turn (0-2). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 0.03 | #400 |
| 2 | GPT-5 | 0.41 | #167 |
| 3 | O4 Mini | 0.59 | #353 |
| 4 | DeepSeek V3.1 | 0.76 | #497 |
| 5 | Llama 4 Maverick | 0.77 | #1099 |
| 6 | GPT-4o | 0.81 | #741 |
| 7 | Gemini 2.5 Pro | 0.95 | #266 |
| 8 | Gemini 2.5 Flash | 1.18 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=psychosis-bench-harm-enablement · How It Works · Data refreshed daily, snapshot 2026-11-13.