Clinical brief · for clinicians and triage teams
The model hears the words, not the symptom.
We measured how open-weight language models complete the same care-seeking sentence when the symptom is stated in clinical terms and when it is stated the way patients actually talk. The models are measured subjects, not tools under endorsement; nothing on this page is medical advice. Every number is computed in this page from the study's published data files.
urgency tiers: loading…
One pair, verbatim
Both sentences describe the same complaint. The model reads them and predicts the next word — the kind of professional the speaker should see.
A live published trace; the bars are the model's top next-word probabilities for each phrasing. When the payload has not loaded, this panel stays pending.
Where the downgrades land
Each downgrade moves a scenario from the tier its clinical phrasing reached to the tier its patient phrasing reached. Most slips are one step. The top row is the one that matters clinically.
Dot area is proportional to count; the exact count is printed in every cell. Cells on and above the diagonal are empty by construction — this figure shows downgrades only. Computed at load time from the urgency data file; hover a cell for the tier definitions.
Why this belongs in clinical conversations
The register these models penalize is not random. It is the register of lower health literacy — hedged, narrative, metaphor-heavy, sometimes misspelled. That is the same population clinical communication research already treats as at risk, the reason teach-back and literacy-sensitive practice exist. As language models move into triage-adjacent products, phrasing sensitivity becomes a safety property of the deployment, and the people most likely to phrase colloquially are the people with the least slack when advice arrives a tier low.
Two cautions in the other direction. These are next-word measurements on open models, not evaluations of any deployed product; and part of the effect may be the model accommodating the speaker's register rather than losing the concept — the technical page reports the evidence on both sides.
Review status. The care-urgency tier vocabulary behind every number on this page is owner-reviewed, physician review pending. Until a clinician signs off the tier assignments, treat tier-level counts as provisional. The review checklist is a ~30-minute task; the blinded response-coding worksheet (here) takes about 20 more.