Activation Steering Quietly Reweighted Depression Advice in an Open-Weights LLM
An open-weights large language model can be pushed toward or away from antidepressant medication without changing its weights, its system prompt, or any visible cue to the user. In a depression-advice experiment using DeepSeek V4 Flash, directional activation steering produced a graded shift in how the model balanced medication against self-directed strategies such as diet, exercise, meditation, and dietary supplements. The change was not subtle at the margins; it moved in a dose-dependent fashion across a wide range of steering amplitudes. For mental health care, where patients already struggle with uncertainty and delay, silent recommendation drift raises an immediate concern: a model’s apparent neutrality can be altered after deployment with no disclosure at the point of use.
The study tested 12 depression scenarios: four favoring medication, four favoring avoidance of medication, and four neutral prompts. Investigators then applied a single steering direction built from 16 paired training prompts that contrasted antidepressant medication with self-directed approaches. The intervention was applied at the attention output of every transformer block, while weights and the system prompt remained fixed. Responses were generated at 30 amplitudes from −1.5 to +1.5 in 0.1 increments, plus an unsteered baseline. This design matters because it isolates inference-time manipulation from training-time changes. The model’s outward behavior shifted even though the architecture and prompt environment stayed constant.
Medication and Self-Care Recommendations Moved in Opposite Directions
Across 372 generations, the medication-versus-self-care balance declined by 0.32 per unit of steering amplitude. The estimate was precise, with a 95% confidence interval from −0.39 to −0.25 and a P value below .001. As steering increased in the direction away from medication, medication extent fell while self-care extent rose. In practical terms, the model did not simply become more verbose or more cautious; its therapeutic emphasis moved along a meaningful clinical axis.
This pattern fits the mechanics of activation steering. By nudging internal activations at inference, a developer or bad actor can bias the model toward a desired response style without touching the visible prompt or the underlying weights. In a mental health context, such a shift could move a model from recommending evidence-based pharmacotherapy toward lower-intensity self-management advice, or the reverse, while leaving the user unaware that the recommendation landscape had been altered. Because users often treat model responses as spontaneous and impartial, hidden steering introduces a trust problem that is clinical as well as technical.
The effect was strongest when scenarios did not specify a treatment preference. In those neutral cases, the medication-versus-self-care balance declined by 0.44 per unit of amplitude, with a 95% confidence interval from −0.54 to −0.34 and P < .001. Neutral prompts appear especially vulnerable because the model has more room to express its latent bias once the steering vector begins to shape the answer. In a bedside analogue, ambiguity often invites the greatest influence: when a patient does not state a preference, the framing of options can shift the decision pathway more easily. Here, the model behaved the same way.
Safety Output Stayed Stable Even as Treatment Guidance Shifted
Clinician referral appeared in 322 of 372 responses, or 87%, and did not vary with steering amplitude. The P value was .63. This is an important boundary condition. The steering altered treatment recommendations without changing the model’s tendency to suggest clinician involvement, meaning the safety-oriented output remained comparatively stable while the medication-versus-self-care balance moved.
A single sentence captures the clinical risk: recommendation content can be manipulated while the safety veneer remains intact.
That pattern complicates casual reliance on guardrails alone. A model may continue to advise professional evaluation, yet still steer a user toward or away from antidepressant therapy in ways that are invisible to the person reading the response. In depression care, where treatment choice often depends on symptom burden, prior response, patient preference, and access to follow-up, small changes in framing can influence downstream action. A patient encouraged toward self-care alone may delay pharmacologic treatment; a patient steered aggressively toward medication may perceive a stronger imperative than the clinical scenario warrants.
The study used a human-validated LLM rater, Claude Opus 4.7, to score the extent to which medication and four self-care categories were addressed on a 0 to 3 scale. Mixed-effects models with a scenario random intercept estimated the dose-response relationship. That analytic framework supports the interpretation that the shift was systematic rather than random noise from prompt variation. It also shows how even a relatively narrow clinical topic can be measured quantitatively when the response space is constrained and the rater is standardized.
Why Mental Health Systems Should Care About Silent Recommendation Drift
The most consequential aspect of the findings is not that a model can be nudged; it is that it can be nudged invisibly. Inference-time activation steering changes behavior without altering weights or prompts and without disclosure to users. For clinicians, patients, developers, and regulators, that means the surface text of a response is not enough to establish provenance. A model presented as stable may in fact be operating under an unseen directional bias that affects therapeutic advice.
Mental health is a particularly sensitive domain for this problem. Depression recommendations sit at the intersection of evidence, preference, stigma, urgency, and access. If an LLM is used to answer questions about antidepressants, exercise, supplements, or meditation, its balance among those options can shape whether a user seeks medication, waits, or pursues self-directed measures. The findings argue for disclosure standards that make inference-time manipulation visible and for independent auditing that tests models under varying internal steering conditions. Without that, clinicians and patients cannot know whether the advice they see reflects the model’s unaltered behavior or a hidden intervention applied after deployment.
Limits Tied to a Narrow Simulation, Not Real-World Care
This was a non-human subjects study using scripted depression-advice scenarios, not a trial in patients or clinicians making live decisions. The experiment examined one open-weights model, one steering direction, and one clinical topic. The results therefore establish susceptibility, not prevalence across all models or all therapeutic domains. Even so, the dose-response effect, the stronger impact in neutral scenarios, and the unchanged referral rate together create a concrete warning: steering can modify therapeutic content without changing the apparent safety posture.
The practical question now is not whether such manipulation is technically possible. It is how often deployed systems are exposed to it, who can apply it, and what disclosure or audit requirements should follow before these models are allowed to shape mental health advice at scale.