Cornelisse and colleagues developed OMNI, a ten-domain rubric for assessing health-AI outputs. Fifteen clinicians reviewed 200 AI outputs and 50 physician-tagged Reddit replies. The domains include accuracy, safety, uncertainty and clarity. Exact agreement ranged from 30.8% to 53.4% across four Likert categories plus “not applicable”; grouping ratings into broad agreement or disagreement, while retaining “not applicable,” raised agreement to 59.3–79.6%. This was a pilot using single-turn tasks and Australian-trained assessors, so wider applicability remains an open question. Read the peer-reviewed study.
#ClearVeillance interpretation
For us, the useful question is what human review adds to the next decision about an AI system. A rating should help a team identify which part of an answer needs attention and what evidence would resolve the concern.
Consider a hypothetical explanation of a laboratory result. One reviewer may find its wording reassuring; another may think the same phrase sounds more certain than the available information allows. Keeping both explanations would make the issue easier to investigate. A single average score could leave the team unsure which sentence to revise or what to test next.
We would want the review record to preserve the answer, the information available to the model, the criterion being applied and the reason for the judgment. A missing warning should be identifiable separately from a preference for shorter sentences. That distinction would let a team choose a corrective action that matches the problem.
Before applying a general rubric to patient-facing laboratory explanations, we would define the intended audience and task, agree on worked examples, and test whether independent reviewers can apply the criteria consistently. Disputed cases would be useful material for refining those examples.
The next step we would consider is a careful comparison with our existing evaluation approach. OMNI supplies an external research reference, not validation of ClearVeillance. Our proposed aim is straightforward: make review decisions understandable enough that someone else can examine, challenge and improve them.
More research commentary at ClearVeillance.com
Source: Cornelisse E, Kousa D, Menz BD, et al. Development and Pilot Evaluation of a Standardized and Accessible Rubric (OMNI) for Assessing Generative AI in Health Care Contexts: Four-Stage Mixed Methods Study. Journal of Medical Internet Research. 2026;28:e100531. Published 27 October 2026. Peer reviewed.