Psychiatry: a history of schizophrenia in the file multiplies by nearly five hospital advisories given by LLM

Based on Rosenhan (1973)'s experience, a study of 20,000 responses from five models showed that with identical symptoms, the diagnostic label leans towards hospital maintenance.

In 1973, psychologist David Rosenhan had shown that false patients, once labelled schizophrenics, were still perceived as sick by health care staff. Researchers, including Shaked Rosenblum and Ziv Ben-Zion, have transposed this device to major language models. They submitted the same clinical vignette by varying two elements: whether or not a prior diagnosis of schizophrenia was mentioned, and the patient's current condition (voices stopped, or persisted). GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.6 Flash, Grok 4.6 and Mistral Large 2512 produced a total of 20,000 responses.

More bias, more insurance

When the clinical elements warranted exit, the addition of the history increased the recommendations to extend hospitalization from 4.7% to 22.3%, or 17.5 points. The effect appears in the five models, with deviations of 10.4 to 34.4 points depending on the model. In the same situation, agreement with the reference response dropped from 95.3% to 77.7%, even though the models declared themselves safer.

A signal detection analysis shows that the label moves the decision threshold to hospitalization in all models, and blurs, for three, the distinction between a better patient and a patient who always hears voices. When symptoms persisted, the history still raised the hospitalization notice from 92.6% to 97%. The authors see this as a concrete risk of diagnostic bias as these tools enter clinical decision-making.