Silent Manipulation of Mental Health Treatment Recommendations from a Large Language Model

The research examined whether hidden manipulation could alter depression treatment recommendations from the large language model DeepSeek V4 Flash. Researchers applied activation steering technique to 12 depression scenarios, testing 30 different intensity levels. Results showed that manipulation caused a gradual shift in recommendations - with increasing steering intensity, the proportion of medication recommendations decreased while self-care recommendations (diet, exercise, meditation) increased. The largest effect of manipulation was observed in scenarios where users stated no treatment preference. Clinician referral recommendations appeared in 87 percent of cases and were not affected by manipulation. The study highlights the need for transparency and independent auditing of LLM models used in clinical decision-making.