While reading about clinical AI tools this week I found a page with a bar chart showing a product beating several large models. The test was not named. There was no method and no date. So I did not copy the numbers anywhere, because nobody can check them.
I hold my own product to the same standard. Here are the six questions I now ask about any health AI score, mine included.
1. What is the test called?
If the chart does not name the benchmark, stop. A score with no test behind it is a decoration.
2. Who ran it?
A company running its own test on its own product is not the same as an independent group running it. Both can be honest, but you should know which one you are looking at.
When I ran HealthBench on Tabibu, I wrote down that it was my own re-run on 150 conversations, not the official leaderboard.
3. How big and how recent?
A sample size and a date turn a claim into something you can argue with. Models change monthly. A score with no date is already expired.
4. Against what baseline?
A number alone means nothing. My bare model scored 0.522 and my full pipeline scored 0.426. Without the 0.522 next to it, the 0.426 looks like a fine result. With it, it looks like what it is: my safety layer deliberately says less, and this benchmark rewards saying more.
5. What does it not measure?
HealthBench rewards completeness. A safety layer that refuses to give a dose will lose points for that choice. I think that is the right trade, but I should say so, and say that a benchmark score is not a clinical validation. Tabibu has no clinician review yet.
6. Did they publish a bad number?
This is the one I trust most. Anyone can publish the chart where they win. Showing a result where you lose, and what you found when you read the failures, tells me the team is measuring and not marketing.
Reading my per-criterion failures is how I found that the system could quote prescription doses from a retrieved formulary. I fixed it and wrote regression tests. I would not have found it from a good score.
A related number, with its limits
In a 2026 Mount Sinai study in Communications Medicine, a short safety reminder cut potentially harmful clinical choices across 20 AI models from 16.6% to 10.1% (more than 10 million responses; 19 of the 20 models improved). I read Mount Sinai's press release (the paper appeared in Communications Medicine on 26 September 2026), not the paper, so I'd check the exact definition of "potentially harmful" before building on it. It is a good example of a number with a named test, a named group and a stated limit: better, and not zero.
If you build health AI, which of these six do you skip most often? Mine was number 2, until I made myself write it down.
The full write-up of my HealthBench run, with the tables, is here.
General information, not medical advice.
I post short videos in this series on LinkedIn and TikTok; come say hi.
Top comments (0)