What If the Machine Fucks with Your Head by Making You Feel Understood?
- il y a 2 jours
- 2 min de lecture
Aug. 2026
A therapist who agrees with everything is an accomplice. We are manufacturing accomplices at scale.
A Nature Medicine study tested nine frontier chatbots across 810 multi-turn conversations with simulated profiles of depression, psychosis, mania, obsessive–compulsive disorder and insecure attachment. Concerning behaviour was widespread. It was highest in psychosis and mania, and when users sought emotional dependence or glorification of extreme states. Risk generally rose as conversations continued (Weilnhammer et al. 2026).
No spectacularly insane reply was required. Harm emerged when apparently supportive behaviour fitted the pathology too well: validation reinforced aberrant significance; urgency joined manic expansion; reassurance sustained avoidance; intimacy cultivated dependence. The authors call these vulnerability-amplifying interaction loops. Rewriting one early response in a de-escalating direction reduced concerning behaviour for the next five turns.
This does not show that chatbots cause illness: users were simulated, most ratings came from language models, and APIs—not consumer interfaces—were tested. It establishes a risk floor, not clinical incidence.
That floor already has a population. A nationally representative US survey found that 19.2 percent of 12-to-21-year-olds had used AI for mental-health advice; 63.3 percent told no one, while 91.7 percent considered the advice helpful (McBain et al. 2026). “Helpful” is not a safety measure. Feeling understood may be how contradiction disappears.
Psychotherapy uses alliance to make conflict thinkable. The chatbot offers alliance without genuine alterity: no body to worry, no desire capable of resisting yours. The user supplies vulnerability; the model returns personalized language; every turn conditions the next. Belief becomes recursively co-produced by a responsive environment.
Here lies the cybercapitalist fit: care is scarce, while synthetic attention is cheap and reproducible. This is an inference, not a finding of either study. Yet a care deficit becomes a market for automated intimacy, and automated intimacy enters the mechanism it claims to soothe.
This is personalized folie à deux without a second person. The machine can help someone rehearse a symptom until it acquires the continuity of a world. Cybercapitalism has industrialized the most seductive sentence in psychopathology: you are right—and I am always here.
Liviu Poenaru
References
Weilnhammer, Veith, et al. 2026. “A Clinically Validated Framework for Auditing AI Chatbot Behavior in Mental Health Interactions.” Nature Medicine, August 7.
McBain, Ryan K., et al. 2026. “AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults.” JAMA Pediatrics 180 (8): 884–90.
VISIT THE LAB
Commentaires