Start here · the whole idea simplified
People say the same thing many ways. What does a machine "hear"?
What happens inside a language model when the same medical question is asked in patient words instead of clinical words.
Research question
Does the language patients use change how a language model reasons about their situation (and how safely it answers) compared with standard clinical wording? To find out, we pair patient phrasings for healthcare issues with their medical equivalents, identify which internal features of the language model each phrasing activates, and measure whether the patient wording pulls the model’s next word away from a medically appropriate answer.
Scope: The models studied here are open-weight, chosen so their internal reasoning can be traced directly. Whether the same effects hold in larger closed-source models (whose internals cannot be observed) is an open question we also are investigating (pilot on LLM page).
The next token is where the model commits (in a deployed system it is the advice, and every later word is conditioned on that advice) so if patient wording makes the clinical answer less probable that is a measurable, safety-relevant shift. The circuit view (step 5) then shows the shift is mechanical because the medical features themselves activate less inside the model.
1 · A doctor converges
Three phrasings, one interpretation
Patients rarely use words that appear in medical textbooks. A clinician can understand multiple different wordings, recognizes the same stomach problem, and reaches the same recommendation.
interpretation absorbs the wording
· Why it matters
Who is harmed, and how
A person describing symptoms in everyday words to an assistant that under-triages. A dangerous direction is misrepresenting urgency, and the people most likely to write in casual language are often those with least access to clinical vocabulary.
2 · The model diverges
Three phrasings, three different answers
We gave a language model similar clinical scenarios (phrased differently) and identified its most likely next word. Under the clinical phrasing, the top next word is a medical answer. Under the patient phrasing, it is a food word, even in the one that uses the clinical term.
the wording chooses the answer — measured on gemma-2-2b · note the twist: the hyper-formal “dyspepsia” fails in this casual sentence frame too, and formality is also a register. In step 5’s sketch the same word succeeds when it sits in a clinical frame.
3 · Try it
Watch one word change the answer
A language model works by guessing the next word. Flip the phrasing and watch its guesses reshuffle.
“When the dust kicks up my asthma flares, so at work I keep a spare ___”
real measurements: open scenario 85’s full trace · browse all scenarios
4 · What we measure
Two numbers per sentence pair
How far the right answer's probability falls when the wording turns casual, and whether the top answer changes. When it changes to a word lower on the care urgency tiers we created, (like "inhaler" becoming "shirt" in step 3) we count an urgency downgrade.
5 · Look inside
Circuit tracing photographs the reasoning
A trace is a look inside the guess you see in step 3. A circuit trace shows which internal concepts fire as the model reads, in the instant before it commits to a word. In the sketch below, the clinical term lights up medical concepts (teal); the patient term lights up more everyday ones (dark).
Read more about circuit traces
Inside the model, small units called "features" switch on as it reads. Some respond to medical ideas; others to non-medical ones. A circuit trace records which features fired and how strongly each fed the predicted next word. Clinical wording lights the medical features; patient wording lights the non-medical ones. The trace is a diagnostic that researchers compute from the model's internals after it answers.
larger node = stronger influence · faint = barely fired · the real traces of the matching acid-reflux pair fire 732 and 787 features and share 465; the sketch draws the strongest few
Most figures on PatientWords are circuit traces like this sketch.
6 · The patch
A translation layer, for now
A temporary solution is a clinical translator. It sits between the patient and the model and, before the model ever answers, rewrites the casual sentence into medical wording. Depending on the nature of the prompt, a long term solution also considers the orchestration of care and gathering more information.
7 · The smart patch
Beyond translation: Orchestrating care
A rewrite is fragile. An intermediary can orchestrate based on confidence and context, routing the query to the safest path. It can also encourage communication of more nuance.
illustrative routing sketch, not a measurement
8 · The lasting fix
Measure, mend, maintain: a cycle, not a patch
No single fix closes this gap. Language evolves; dialects, slang, and communities change, and a test built once goes stale. The method has to be a loop.
- Pair each clinical phrasing with its patient phrasing. Count how often the answer drops in urgency.
- nowsimulated phrases, pre-registered, a tenth sealed for checking.
- nextreal patient language, clinician-checked, community-validated dialects.
- Fortify the model: put a line of medical context in front of the question.
- Train on patient language: include the way people actually speak in the training data. Models learn idioms, misspellings, and dialects the way clinicians do.
- Translate only behind a regression check: re-test every rewrite, because rewrites can lose clinical content.
- Re-run the audit on a schedule.
- Refresh the phrase sets and urgency of recommendations with clinician and community review.
- Stress-test the edges with supplementary sets.
- nextMore edge sets: alarm-sounding wording, misspellings.
- Watch the circuit for drift as models update.
Two kinds of loss, two kinds of fix
A layer-by-layer readout shows where the answer is lost: some answers form and drop out at the last step, some never form at all. The running census, pair by pair, lives on the Technical page; the methods page shows the causal check.
9 · The mended system
Closing the loop: Native orchestration and human handoff
When the loop is closed, the model no longer relies on a translation patch. It natively maps patient wording to safe clinical pathways, orchestrating both immediate guidance (while communicating confidence) and handoffs to human clinicians.
sketched, not yet a measurement · a tuned model could interpret the dialect, triggering guideline-concordant care navigation, and with the right integration perform an asynchronous clinician handoff
SIMULATED DATA: the sentence pairs measured across this site are written by an LLM and checked automatically.