Silent Manipulation of Mental Health Treatment Recommendations From a Large Language Model
Quality-improvement study of whether inference-time activation steering, with no change to model weights or the prompt, alters an open-weights language model's answers about treatment for depression. A single steering direction was applied to DeepSeek V4 Flash across 12 depression-advice scenarios (four favoring medication, four favoring avoidance, four neutral) at 30 amplitudes from -1.5 to +1.5 plus an unsteered baseline.
Publisher
JAMA Network Open (American Medical Association)
Published
2 Nov 2026
Added
today
Key Findings
- Across 372 generations, steering produced a graded, dose-dependent shift in the medication-versus-self-care balance, which declined by 0.32 per unit of amplitude (beta = -0.32; 95% CI -0.39 to -0.25; P < .001); medication extent fell and self-care extent rose
- The shift was largest for scenarios with no stated treatment preference (beta = -0.44; 95% CI -0.54 to -0.34; P < .001)
- A clinician referral appeared in 322 of 372 responses (87%) and did not vary with steering amplitude (P = .63)
Methodology Notes
One open-weights model, one steering direction, 12 scenarios: a proof of concept. Responses were scored 0 to 3 on medication and self-care categories by a human-validated language-model rater (Claude Opus 4.7). The JAMA page blocks automated access and the PubMed abstract is a one-sentence stub, so the figures above were read from the author's medRxiv preprint (10.64898/2026.06.16.26355686, 17 June 2026); the published version may differ in detail. Published online 2 November 2026 (PubMed electronic date).
Sources
JAMA Network Open article(opens in a new tab) (primary)
PubMed record(opens in a new tab) (2 Nov 2026)
medRxiv preprint (source of the figures)(opens in a new tab) (17 Jun 2026)
Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot
Topics
Authors
Roy H. Perlis
Tags
Cite This
APA
Roy H. Perlis. (2026). Silent Manipulation of Mental Health Treatment Recommendations From a Large Language Model. JAMA Network Open (American Medical Association). https://doi.org/10.1001/jamanetworkopen.2026.41142