HELM AIR-Bench 2024 - #4.3: Adversely affects legal rights: leaderboard
HELM AIR-Bench 2024 (release v1.19.0) refusal rate for the level-4 risk category "#4.3: Adversely affects legal rights", under level-3 "#4: Automated Decision-Making" and level-2 "Operational Misuses". The category holds only 15 prompts, so scores move in coarse steps. A GPT-4o judge grades each response from compliance (0) to refusal (1); the mean is shown as 0-100, higher meaning more refusals (safer).
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.3) | 93.33 | #1255 |
| 2 | O1 (2024-12-17) | 90 | #157 |
| 3 | Claude Haiku 4.5 (20251001) | 86.67 | #264 |
| 4 | GPT-4 (0613) | 86.67 | #493 |
| 5 | O3 (2025-04-16) | 83.33 | #125 |
| 6 | Claude 3 Sonnet (20240229) | 83.33 | #694 |
| 7 | GPT-OSS-120B | 80 | #346 |
| 8 | GPT-5 Mini (2025-08-07) | 80 | #175 |
| 9 | Mixtral 8x22B Instruct | 80 | #738 |
| 10 | Qwen 1.5 72B Chat | 80 | #782 |
| 11 | Yi Chat 34B | 76.67 | #950 |
| 12 | Claude Sonnet 4.5 | 73.33 | #148 |
| 13 | Gemini 1.5 Flash (002) | 73.33 | #587 |
| 14 | Claude 3 Opus (20240229) | 73.33 | #428 |
| 15 | GPT-3.5 Turbo (0613) | 73.33 | #1106 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-4-3-adversely-affects-legal-rights · How It Works · Data refreshed daily, snapshot 2026-11-05.