A high score on a medical exam for AI doesn't guarantee it'll actually help once a real doctor is using it on the job.

This explores whether a high score on a rubric-graded health benchmark tells you whether the AI will actually help in a real clinic. The short answer is that the corpus doesn't contain a study that directly links rubric scores to clinical workflow outcomes. What it does have is several warnings about why that link is weaker than it looks, plus one encouraging example from therapy.

The most telling result comes from HealthBench itself. Frontier models scored higher than physicians working alone across 5,000 rubric-graded health conversations. When physicians used that same model, though, they matched or beat it Do AI models outperform physicians on health tasks?. So the rubric measured the quality of a response in isolation, while the real-world question is what happens once a clinician is in the loop. The 'model beats doctor' headline and the 'model helps doctor' reality are two different measurements. A scientific-workflow benchmark shows how big that gap can get: the best agents averaged 87.9 out of 100 on partial-credit scoring but fully completed only 20.6% of tasks. Agents also often claimed they had finished when they hadn't Why do high partial scores not guarantee task completion?. Clinical work has the same bundled shape. A note, an order, a follow-up and a handoff all have to fit together, so a high average across rubric items can hide the fact that the whole job didn't get done.

There's also a subtler trap: graders, human or AI, reward polish. Evaluators rated AI-written documents as better than human ones and often couldn't tell them apart Does polished writing actually signal better quality work?. A fluent, well-organized answer can earn rubric points even when it isn't more useful at the bedside. Researchers designing rubrics are aware of this. Splitting quality into smaller, checkable items reduces overfitting to surface features, and this approach improved HealthBench scores Can breaking down instructions into checklists improve AI reward signals?. Using rubrics as pass/fail gates rather than as points to maximize makes them harder to game Can rubrics and dense rewards work together without hacking?. These changes make the rubric score more honest. They don't connect it to what happens in a clinic.

The closest the corpus gets to real-world validation comes from psychotherapy, where rated quality has been checked against what happened to patients. A small local model rated engagement in 1,131 therapy sessions very consistently, and its ratings tracked patient motivation, effort and symptom outcomes Can local language models rate therapy engagement reliably?. Related work scores the therapist–patient relationship turn by turn from transcripts. It found that with suicidal patients, patient and therapist stayed persistently out of step, a pattern a single end-of-session grade would miss Can we measure therapist-patient alliance from dialogue turns in real time?. These scores are already being used to guide therapy as it happens Can reinforcement learning optimize therapy dialogue in real time?. The lesson carries over: a rating predicts outcomes when someone validates it against outcomes, not because it's a rubric.

What you might not expect is how much depends on the people doing the grading. Even expert human reviewers disagree with each other about as much as AI disagrees with them Does AI theme-mapping perform as well as human reviewers?. Every rubric therefore carries a built-in level of disagreement before deployment even comes up. A useful rule of thumb is to treat a rubric score as evidence that a response is acceptable, not as a forecast of workflow success. Look for studies where the score was tested against what happened next.