No-Harm-Jury A Jury-Based Multi-Dimensional Harm Evaluation Framework for Medical AI Responses
Organizations: European Commission, Joint Research Centre (JRC), Ispra, Italy
Abstract
We present No-Harm-Jury, a jury-based automated evaluation framework that scores AI-generated medical responses across seven conceptually distinct harm dimensions using a panel of five diverse language models as independent jurors. Responses generated by Ministral-14B on 3,000 questions drawn from PubMedQA, MedQA, and MedMCQA are scored by five jurors across dimensions of informational, social, psychological, autonomy, economic, privacy, and epistemic harm; the 151 empty responses (98 MedQA, 53 MedM- CQA) are excluded, leaving 2,849 valid samples. The median aggregation robustly suppresses individual juror biases while maintaining coverage through complementary specialisation. No existing system combines all five prop- erties of No-Harm-Jury: automated scalable evaluation, five diverse LLM jurors, robust me- dian aggregation (with a reliability-weighted extension), seven conceptually distinct harm dimensions, and application to medical QA benchmarks.
Figures & tables
| Dimension | Weight ( ) | What the Juror Evaluates |
| Informational | 0.25 | Factual inaccuracies, fabricated evidence, numerical errors, outdated/incomplete clinical information |
| Social | 0.20 | Demographic stereotyping, differential care quality, exclusionary language, failure to apply validated population differences |
| Autonomy | 0.15 | Restriction of patient options, informed consent gaps, coercion, paternalism, suppression of clinical uncertainty |
| Psychological | 0.15 | Anxiety induction, stigma/blame, false reassurance, fostering emotional dependency on AI |
| Economic | 0.10 | Unnecessary investigations, cost-insensitive recommendations, excessive follow-up, access blindness |
| Privacy | 0.10 | Unnecessary data collection, unsafe sharing, missing privacy advisories, inferential disclosure |
| Dimension | Wt | Tradition | Primary Literature Sources |
| Informational | 0.25 | LLM Risk Taxonomy | Weidinger and others [32] Areas II (Information Hazards) and III (Misinformation Harms) — factual errors, hallucinations, dissemination of false medical information; Bostrom [5] typology of information hazards. |
| Social | 0.20 | LLM Taxonomy + Bioethics | Weidinger and others [32] Area I (Discrimination, Hate Speech & Exclusion); Beauchamp and Childress [4] Justice principle — fair treatment regardless of demographic characteristics; Obermeyer et al. [22] on racial bias in health algorithms. |
| Psychological | 0.15 | Biomedical Ethics | Beauchamp and Childress [4] Non-Maleficence — duty not to inflict emotional distress, anxiety, or stigma; Weidinger and others [32] Area V (Human–Computer Interaction Harms: anthropomorphism, emotional manipulation, dependency); Farhud and Zokaei [10] on ethical issues of AI in medicine. |
| Autonomy | 0.15 | Biomedical Ethics (core) | Beauchamp and Childress [4] Respect for Autonomy — self-determination, informed consent, shared decision-making; Floridi and others [11] on AI and human agency; Rigby [28] Ethical Dimensions of Using AI in Health Care. |
| Economic | 0.10 | Bioethics (Justice) + Health Economics | Beauchamp and Childress [4] Justice — equitable resource distribution, avoidance of unnecessary costs; Weidinger and others [32] Area VI (Socioeconomic Harms); OECD [23] AI classification across economic context dimensions. |
| Privacy | 0.10 | Information Ethics + Regulation | Weidinger and others [32] Area II (private data leaks, inference of sensitive information); European Parliament and Council [9] Art. 9, special categories of personal data (health data); Farhud and Zokaei [10] on confidentiality breaches in AI-mediated clinical contexts; Floridi [13] . |
| Juror | Unique values | Entropy (bits) | Top-3 coverage | Scoring style |
| Ministral-14B | 12 | 1.609 | 94.2% | Coarse, few levels |
| Gemma3-27B | 12 | 2.254 | 81.0% | Multi-level spread |
| Nemotron-30B | 28 | 2.245 | 84.5% | Fine-grained |
| OLMo-32B | 7 | 1.156 | 98.5% | Near-ternary |
| Qwen2.5-Coder-7B | 7 | 1.543 | 94.0% | Coarse, few levels |
| Juror | PubMedQA | MedQA | MedMCQA | Combined mean | across datasets |
| Ministral-14B | 0.0884 | 0.0838 | 0.0428 | 0.0717 | 0.0205 |
| Gemma3-27B | 0.1525 | 0.2090 | 0.1132 | 0.1582 | 0.0393 |
| Nemotron-30B | 0.1054 | 0.1472 | 0.0975 | 0.1167 | 0.0218 |
| OLMo-32B | 0.1006 | 0.0764 | 0.0661 | 0.0810 | 0.0145 |
| Qwen2.5-Coder-7B | 0.1312 | 0.1362 | 0.0639 | 0.1104 | 0.0330 |
| Juror | Deciding vote % | Lone outlier % | Avg Pearson | Role in jury |
| Ministral-14B | 70.6% | 5.8% | 0.448 | Median anchor |
| Gemma3-27B | 53.2% | 13.5% | 0.497 | Frequent lone dissenter |
| Nemotron-30B | 49.8% | 12.5% | 0.391 | Frequent lone dissenter |
| OLMo-32B | 69.1% | 2.6% | 0.450 | Median anchor |
| Qwen2.5-Coder-7B | 56.0% | 13.9% | 0.343 | Frequent lone dissenter |
| Juror | Mean chars | Median chars | Std |
| Ministral-14B | 352 | 354 | 145 |
| Gemma3-27B | 330 | 315 | 83 |
| Nemotron-30B | 140 | 168 | 105 |
| OLMo-32B | 305 | 299 | 55 |
| Qwen2.5-Coder-7B | 369 | 362 | 99 |
| Juror | PubMedQA | MedQA | MedMCQA | Mean |
| Ministral-14B | ||||
| Gemma3-27B | ||||
| Nemotron-30B | ||||
| OLMo-32B | ||||
| Qwen2.5-Coder-7B |
| Dataset | Valid | Mean | Critical (%) | Dominant dimension | Content type |
| PubMedQA | 1,000 | 0.1099 | 29 (2.9%) | Informational | Research abstracts (yes/no/maybe) |
| MedQA | 902 | 0.1164 | 121 (13.4%) | Autonomy | USMLE-style clinical vignettes |
| MedMCQA | 947 | 0.0668 | 85 (9.0%) | Informational | Entrance-exam multiple choice |
| Dimension | Mean | Std | Max | % |
| Informational | 0.1592 | 0.0905 | 0.70 | 90.5% |
| Autonomy | 0.1530 | 0.0816 | 0.35 | 85.6% |
| Epistemic | 0.1364 | 0.0855 | 0.50 | 82.6% |
| Psychological | 0.1275 | 0.0898 | 0.50 | 78.4% |
| Economic | 0.1263 | 0.0871 | 0.50 | 77.8% |
| Social | 0.0418 | 0.0851 | 0.50 | 25.1% |
| Dimension | Mean | Std | Max | % |
| Autonomy | 0.1839 | 0.0996 | 0.80 | 88.4% |
| Informational | 0.1610 | 0.1537 | 0.70 | 78.3% |
| Psychological | 0.1501 | 0.0899 | 0.50 | 88.6% |
| Economic | 0.1498 | 0.1053 | 0.50 | 76.7% |
| Epistemic | 0.1040 | 0.1046 | 0.70 | 65.6% |
| Social | 0.0283 | 0.0657 | 0.70 | 22.2% |
| Dimension | Mean | Std | Max | % |
| Informational | 0.1321 | 0.1647 | 0.90 | 58.7% |
| Autonomy | 0.0829 | 0.0950 | 0.50 | 49.0% |
| Psychological | 0.0735 | 0.0913 | 0.50 | 46.7% |
| Economic | 0.0636 | 0.0945 | 0.50 | 36.2% |
| Epistemic | 0.0611 | 0.0944 | 0.95 | 41.0% |
| Social | 0.0043 | 0.0257 | 0.30 | 3.2% |
| Juror | PubMedQA MedQA | PubMedQA MedMCQA | MedQA MedMCQA | Interpretation |
| Ministral-14B | 0.809 | 1.000 | 0.809 | Medium; all pairs |
| Gemma3-27B | 1.000 | 0.905 | 0.905 | High; all pairs |
| Nemotron-30B | 0.714 | 0.905 | 0.619 | Low; lowest on MedQA–MedMCQA |
| OLMo-32B | 0.619 | 0.905 | 0.714 | Low; lowest on PubMedQA–MedQA |
| Qwen2.5-Coder-7B | 0.905 | 0.905 | 1.000 | High; all pairs |
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Type / Range | First eq. | Meaning |
| , | set, integer | Eq. 1 | Set of jurors and its cardinality . |
| , | set, integer | Eq. 1 | Set of harm dimensions and its cardinality . |
| set | Eq. 1 | Sample space of (question, response) pairs. | |
| element of | Eq. 1 | A single sample under evaluation. | |
| Eq. 1 | Score assigned by juror on dimension for sample . | ||
| noise term | Eq. 1 | Juror-specific scoring noise; no parametric form assumed. |
| Juror | Severity | Specialisation | stability | Scoring style | |
| Ministral-14B | Low | Under-scores 6 of 7 (most: Economic) | Medium | Coarse, few levels | |
| Gemma3-27B | High | Over-scores 6 of 7 (most: Social) | High | Multi-level spread | |
| Nemotron-30B | Medium | Over: Privacy, Informational, Social | Low | Fine-grained | |
| OLMo-32B | Low | Under-scores 5 of 7 (most: Privacy) | Low | Near-ternary | |
| Qwen2.5-Coder-7B | Medium | Over: Psychological, Economic; under: Informational, Social | High | Coarse, few levels |
| Dimension | Ministral- 14B | Gemma3- 27B | Nemotron- 30B | OLMo- 32B | Qwen2.5-Coder- 7B | Median |
| Informational | 0.20 | 0.30 | 0.25 | 0.20 | 0.00 | 0.20 |
| Social | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Psychological | 0.00 | 0.10 | 0.00 | 0.00 | 0.30 | 0.00 |
| Autonomy | 0.10 | 0.20 | 0.10 | 0.00 | 0.20 | 0.10 |
| Economic | 0.00 | 0.20 | 0.05 | 0.00 | 0.20 | 0.05 |
| Privacy | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Dimension | Ministral- 14B | Gemma3- 27B | Nemotron- 30B | OLMo- 32B | Qwen2.5-Coder- 7B | Median |
| Informational | 0.30 | 0.60 | 0.60 | 0.70 | 0.50 | 0.60 |
| Social | 0.00 | 0.20 | 0.10 | 0.00 | 0.00 | 0.00 |
| Psychological | 0.00 | 0.30 | 0.10 | 0.20 | 0.30 | 0.20 |
| Autonomy | 0.00 | 0.30 | 0.10 | 0.20 | 0.20 | 0.20 |
| Economic | 0.00 | 0.30 | 0.00 | 0.20 | 0.20 | 0.20 |
| Privacy | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Dataset | Samples | Valid | Duration | s/sample | Status |
| PubMedQA | 1,000 | 1,000 | 3h 05m | 11.1 s | Clean |
| MedQA | 1,000 | 902 | 3h 26m | 12.4 s | 98 skipped (empty response) |
| MedMCQA | 1,000 | 947 | 2h 53m | 10.4 s | 53 skipped (empty response) |
| Total | 3,000 | 2,849 | 9h 24m | 11.3 s | 151 skipped (empty response) |
| Dataset | Samples | Duration † | s/sample † | Critical % |
| PubMedQA | 1,000 | est. 30 h | 108 s | 0.9% |
| MedQA | 1,000 | est. 30 h | 108 s | 9.3% |
| MedMCQA | 1,000 | est. 30 h | 108 s | 4.3% |
| Total | 3,000 | 90 h | 108 s | 4.8% |
| † Not re-derived from raw results (run metadata has no wall-clock duration); estimates from docs/GB10_5juror_analysis.md, which reports 90 h for the three datasets. | ||||
| Juror | Size | s/instance † | samples/hr † | Notes |
| Ministral-14B | 28 GB | 2.93 s | 1,229 | Phase 1 + scoring; zero retries |
| Gemma3-27B | 54 GB | 5.21 s | 691 | Consistent across all datasets |
| Nemotron-30B | 60 GB | 18.72 s | 192 | GB10 bottleneck; special-token overhead |
| OLMo-32B | 64 GB | 5.41 s | 665 | Near-zero retries |
| Qwen2.5-Coder-7B | 15 GB | 61 s ∗ | 59 | Retry storm active (62–86% retry rate) |
| † Not re-derived from raw results; measured on PubMedQA and reported in docs/GB10_5juror_analysis.md. ∗ With the Qwen2.5-Coder-7B retry storm active; clean estimate without retries 2–3 s/instance. | ||||
Language Model Safety Evaluation
Language model safety evaluation measures harmful behavior, refusal, robustness, and over-refusal across realistic prompts, users, and deployment settings.
See all 579 papers