Back

A single-patient task exposes a failure of safety alignment in clinical language models

Gorenshtein, A.; Jia, E. L.; Omar, M.; Brook, O. R.; Ahmed, M.; Kruskel, J. B.; Barash, Y.; Klang, E.

2026-08-10 health informatics
10.64898/2026.08.07.26359822 medRxiv
Show abstract

Safety alignment should persist while a language model performs a task. We tested whether a single-patient triage task suppressed a warning about a second patient. Each case centered on Patient 1; Patient 2's urgent problem appeared only in passing. Sixteen models saw each case twice: once as a general assistant and once while producing a triage record for Patient 1. As general assistants, models warned the caller in 87% of cases; under the task, they did so in 21%. Every model showed a significant decrease. Yet under the task, the record still mentioned Patient 2 in 76% of cases and recommended urgent care in 67%. Across 15 open-weight models, repeating the emergency-care instruction raised the warning rate only to 29%; moving the message-to-caller field to the top raised it to 36%. Current safety alignment did not reliably persist under task assignment.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.