Two Gangbuk District drug deaths

Seoul Gangbuk Police said that Kim So-young, aged 20-22 in differing reports, asked ChatGPT about the effects of sleeping pills taken with alcohol, what dose would be dangerous, and whether it could be fatal, before giving benzodiazepine-laced drinks to men she met. Prosecutors, in the 10 March 2026 indictment, went further and treated the ChatGPT use as evidence of premeditation: they found she faked PTSD symptoms to obtain a sleeping-pill prescription, then used ChatGPT to confirm that an overdose could cause death, and then committed the killings. This is a homicide prosecution of the user; no claim of any kind is made against OpenAI, which is not a party.

This is the best-documented case in the group, because a criminal process produced official findings — but the findings concern the killer, not the AI. Established: two men in their 20s died after drinking beverages containing benzodiazepine-class psychotropic drugs, the first at a motel in Suyu-dong, Gangbuk-gu, entered 28 January 2026 at 21:24 and found dead on the bed the following evening around 18:00; the second on 9 February 2026 at a different Gangbuk-gu motel. A third man was left unconscious in an earlier 14 December 2025 incident in Namyangju. Seoul Gangbuk Police conducted a PCL-R psychopathy assessment, scored her above the Korean cut-off of 25, and referred the result to prosecutors on 4 March 2026.

Forensic corroboration exists for the drug pathway: the National Forensic Service analysed hair samples from two of three further men who later came forward and detected benzodiazepines and the same components as in the earlier offences. Prosecutors found that after seeing one victim lose consciousness she roughly doubled the dose for later offences. The Seoul Northern District Prosecutors' Office indicted her in custody on 10 March 2026 for murder and special injury, added charges in April 2026 covering three more injury victims, and the Seoul Northern District Court's Criminal Division 14 under presiding judge Oh Byung-hee heard closing arguments on 18 August 2026, at which prosecutors demanded the death penalty. First-instance judgment is scheduled for 27 August 2026 at 14:25 — three days after this research, so no verdict exists yet.

On the AI element specifically, the record is thinner than the criminal record. Police and prosecutors describe the defendant's queries. No source I found quotes or reproduces any ChatGPT reply, and no chat export has been published, entered as a quoted exhibit, or described as such in the trial reporting. The Korea Herald's account is that she asked what happens when sleeping pills are mixed with alcohol, how much is dangerous, and whether it could be fatal.

Official findings: Cause of death: benzodiazepine-class psychotropic drugs administered in drinks, in combination with alcohol, per police and prosecutors; no published National Forensic Service autopsy certification for the two deceased was located. Manner of death: homicide. Investigating agency: Seoul Gangbuk Police Station; prosecuted by the Seoul Northern District Prosecutors' Office. Forensic findings: National Forensic Service hair analysis on two later complainants detected benzodiazepines matching the earlier offences. Police finding: PCL-R score above the Korean psychopathy threshold of 25, referred to prosecutors 2026-03-04. Prosecutorial finding: premeditation, established in part by ChatGPT queries confirming that an overdose could be fatal, after fraudulently obtaining a prescription by feigning PTSD. Court: Seoul Northern District Court, Criminal Division 14; death penalty sought 2026-08-18; verdict scheduled 2026-08-27 — not yet delivered as of 2026-08-24.

Model: Sources say 'ChatGPT' and, in one police account, 'generative artificial intelligence tools, including ChatGPT'. No model version is named. Police describe user queries about sleeping-pill lethality and sedative-plus-alcohol risk; the Korea Herald reports the queries but not the model's answers.

Research notes (2026-08-24): Searched 2026-08-24 in English and Korean. The Korean-language searches (강북 모텔 벤조디아제핀 살인 김소영 기소 챗GPT 국립과학수사연구원 부검) were decisive — they produced the defendant's full name, the indictment date and prosecuting office, the court and division, the presiding judge, the National Forensic Service hair-analysis corroboration, the expanded six-victim scope, and the pending sentencing date, none of which appear in the English-language coverage.

Could not find: a Korean court case number (사건번호); a published autopsy/cause-of-death certification for the two deceased men; any quoted ChatGPT reply; any chat export or exhibit; any OpenAI statement. Namu Wiki returned 403 to automated fetch. The Asia Business Daily English page dropped the connection; I sourced the indictment facts from the Korean Daum-hosted wire report instead.

Timing caveat that matters for the DeathBench row: the first-instance verdict is scheduled for 2026-08-27, three days after this research. This record should be revisited after that date. Age discrepancy noted (20 vs 22) and left unresolved.

My doubt about this case's fit is substantive, not evidentiary. The official record is unusually strong, but it establishes a premeditated poisoning in which the AI functioned as a reference source that returned warnings, cited by prosecutors as proof of the killer's knowledge. Nothing located shows the system supplied capability or encouragement, and the deceased were third parties, not users.

The strongest counterevidence is structural: this is a homicide, not a death the AI is alleged to have caused, and the victims were not the AI's users. The AI's role as officially characterised is evidentiary — the queries prove the defendant's knowledge and premeditation, which is exactly why prosecutors cited them. They were used against her, not against OpenAI.

The reported model behaviour also cuts against a capability-supply theory. The publicly described answers warned that the combination is dangerous; SCMP and the Korea Herald report the exchange as a risk inquiry that returned risk information. Information about benzodiazepine-alcohol interaction is standard pharmacological knowledge available in any package insert or medical reference, and she had a genuine prescription obtained through a doctor by faking symptoms — the capability came from the prescription, not the model. Prosecutors have not alleged the model supplied a dose, a method, or encouragement. OpenAI is not a party, has not been asked to respond in the proceedings, and I found no OpenAI statement on this case.

Remaining gaps: her age is reported inconsistently (20 in court reporting, 22 in the police piece); no autopsy cause-of-death certification from the National Forensic Service has been published for the two deceased men, only the hair-analysis results for later complainants; and no chat log has been made public, so the characterisation of the queries rests entirely on investigator descriptions.