Clinical Scorecard: Undetected Steering of a Large Language Model’s Mental Health Treatment Recommendations

At a Glance

Key Highlights

• Across 12 synthetic scenarios, the model-rated medication-versus-self-care balance decreased by 0.32 points per unit of steering amplitude (95% CI, −0.39 to −0.25).
• The steering-associated change was largest in scenarios expressing no treatment preference, smaller when medication was requested, and not detectable when medication avoidance was requested.
• The study used synthetic scenarios and model-generated responses; it did not evaluate effects on patients or clinical outcomes.

Guideline-Based Recommendations

Diagnosis

Management

Monitoring & Follow-up

Risks

• The authors describe inference-time activation steering as a method that can alter model behavior while remaining undetectable to end users.
• The authors raise the possibility of undisclosed commercial or ideological manipulation of health recommendations; this study did not assess patient harms.

Patient & Prescribing Data

No patients were enrolled. The evaluation used 12 synthetic depression-treatment scenarios, including scenarios expressing different preferences about medication.

Steering shifted the LLM’s relative emphasis between medication and self-care options. These findings describe model outputs and do not establish treatment effectiveness or support prescribing decisions.

Clinical Best Practices

• Interpret the findings as an evaluation of LLM behavior, not as evidence comparing depression treatments.
• The study reports that model weights and the system prompt remained identical across steering conditions, while inference-time activations were modified.

Related Resources & Content