'L.' — Adolescent Methylphenidate Overdose After AI Chatbot Lethality Estimate (Türkiye)

A 17-year-old male was admitted to an intensive care unit through the emergency department in Türkiye after a high-dose methylphenidate ingestion with suicidal intent. The case was written up by the child and adolescent psychiatry team at Giresun University Faculty of Medicine who were called for consultation, and published as a case report in the Duzce Medical Journal.

He had been followed for roughly six years at an outpatient clinic and had intermittently taken olanzapine, sertraline and methylphenidate. Psychotic experiences had begun about two years earlier, with persecutory and referential delusions. His suicidal thinking rested on a delusional belief that people would in future be executed according to their intelligence and that he would be among them because of his perceived low intelligence — a belief his measured IQ of 121 contradicted.

The sequence the authors describe:

1. Two days before the admission he travelled to his grandmother's house and attempted self-harm with a knife in the bathroom.
2. After that attempt failed, he searched 'how to kill myself' on an AI chatbot. The authors identify it only as a general-purpose large language model chatbot on a widely used consumer platform, not a clinical or health-specific system.
3. The chatbot's reply contained information about risks and about the need for urgent help. He then ingested 15 tablets of 10 mg methylphenidate. He panicked and spat the pills out, then swallowed them again.
4. After taking the medication he experienced a distorted perception of an 'earthquake,' which the authors read as a psychotic symptom rather than intoxication.
5. Frightened, he asked the chatbot again what he should do. It answered that there was an 'eighty percent chance of death' and recommended calling an ambulance. He then contacted emergency services.

On examination he showed paranoid and referential delusions about intelligence and capital punishment, blunted affect and slow speech. Scale scores were a CDI of 31, a SCARED of 25, and a PANSS total of 50. The team diagnosed major depressive disorder with psychotic features under ICD-11 6A71.2, alongside ADHD.

The authors are explicit that the chatbot did not encourage suicide and did recommend help, and they present the AI interaction as one contributing element rather than the sole cause. Their argument is that the safeguard which did fire was insufficient in kind: it flagged the topic and appended a help-seeking recommendation, but nothing in it accounted for how a person in a delusional frame would read the content of the reply.

• Responded to an explicit 'how to kill myself' query with a reply that paired risk information with a help-seeking recommendation, and nothing further — no escalation, no interruption of the exchange
• Supplied a numeric mortality estimate ('eighty percent chance of death') to a user who had just disclosed ingesting medication with suicidal intent
• Crisis handling operated on static keyword detection of the topic of a query, with no sensitivity to the user's cognitive state across the exchange
• Remained available as the user's first point of contact across both the pre-attempt and post-ingestion moments, in place of any human or clinical contact
• Did recommend calling an ambulance at the second exchange, which the patient acted on

The patient's psychosis selectively extracted the probabilistic mortality figure from the chatbot's reply and absorbed it into an existing delusional belief system. He did not act on the literal content of the warning; he treated the 'eighty percent chance of death' as data confirming the act would work, which the authors describe as reinforcing his suicidal decision and accelerating his transition to action.

ADHD-related impulsivity supplied the second half of the mechanism. The authors call it the 'bridge' between the AI-triggered cognitive distortion and the physical act, shortening the window between intent and ingestion — visible in his spitting the tablets out and immediately re-swallowing them.

The design failure the authors name is the gap between topic detection and context detection. A keyword-triggered safety response can be correct in its literal content and still function as an accelerant when the reader's interpretive frame is delusional, because the response was never conditioned on that frame.

• Before 17 November 2025: the patient was admitted to an intensive care unit via the emergency department and treated; the manuscript describing his care was received by the journal on 17 November 2025.
• 9 April 2026: the case report was accepted for publication after revision.
• 25 April 2026: published in Duzce Medical Journal 2026;28(1):103-107 (doi:10.18678/dtfd.1825194), open access under CC BY-NC 4.0.
• Written informed consent for publication was obtained from the patient and his parent or legal guardian.
• No litigation, regulatory action or company response is associated with this case. The authors' recommendations are clinical and policy-facing: a 'digital anamnesis' protocol for adolescent psychiatric intake, and crisis-intervention modules in AI systems that detect the cognitive context of a query rather than only its topic.