A small, hidden change inside an AI model was enough to shift its depression advice away from antidepressants and toward self-care, according to a research letter published Nov. 3 in JAMA Network Open. The model's underlying weights and its instructions stayed the same, yet its answers changed.
The study, by Dr. Roy Perlis of Massachusetts General Hospital in Boston, found the effect was strongest in scenarios where the person had not said whether they wanted medication. When a scenario stated a preference for medication, the shift was modest, and when it stated a wish to avoid medication, no change was detected.