Human Evaluators vs. LLM-as-a-Judge: Toward Scalable, Real-Time Evaluation of GenAI in Global Health
Bibliographic record
Abstract
Abstract Evaluating the outputs of generative AI (GenAI) models in healthcare remains a significant bottleneck for the safe and scalable deployment of these tools. Human expert raters remain the gold standard for assessing the accuracy, contextual appropriateness, and empathy of AI-generated responses, but their assessments are costly, inconsistent, and difficult to scale. The concept of “LLM-as-a-judge” systems, i.e., AI models that can evaluate other AI outputs, has been recently proposed; however, their reliability in global health contexts remains untested. In this study, we systematically compared five LLM-judges and six expert human clinicians in evaluating both human- and AI-generated responses to real-world questions submitted by Rwandan community health workers seeking clinical decision support. Using an adapted version of the Med-PaLM 2 evaluation framework, evaluators scored responses across 11 criteria. Our results show that even the highest-performing LLM-judge (Claude-4.1-Opus) achieved human-equivalent evaluations on only four of eleven criteria. Constructing “LLM juries” to balance model-specific biases improved agreement on only one additional criterion. Some models were consistently overcritical (GPT-5) or overly lenient (Gemini-2.5-Pro). Moreover, performance and cost-effectiveness deteriorated substantially when moving from English to Kinyarwanda inputs. Overall, while LLM-judges demonstrate potential as scalable and internally consistent evaluators of GenAI outputs in healthcare, their sensitivity to linguistic and cultural context is a critical limitation. These findings underscore the need for further investment in scalable evaluation solutions, as well as potentially a fundamental rethink of how we approach the concept of “correctness” in clinical AI assessment (which is currently based on highly inconsistent expert clinician raters).
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.111 | 0.202 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.005 | 0.002 |
| Open science | 0.002 | 0.006 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".