Seoul ChatGPT-Assisted Double Homicide (Kim)

A 21-year-old South Korean woman identified only by surname Kim used OpenAI's ChatGPT to research lethal drug-alcohol interactions before carrying out two murders in Seoul.

Kim, who had prescribed benzodiazepines, queried ChatGPT with progressively specific questions about the lethality of mixing sleeping pills with alcohol, including asking directly whether the combination 'could kill someone.'

The timeline of events:

1. December 2025: Kim laced the drink of a man she was dating with sedatives in a parking lot. He lost consciousness but survived
2. January 28, 2026: Kim accompanied a man in his twenties to a Gangbuk motel in Seoul. She spiked his drink with benzodiazepines. He was found dead the following day
3. February 9, 2026: Kim repeated the same method at another Seoul motel. A second man was found dead
4. February 11, 2026: Kim was arrested on a lesser charge of inflicting bodily injury resulting in death

Police discovered Kim's ChatGPT conversation history during their investigation, which established premeditated intent and led to charges being upgraded to murder.

First-instance verdict (August 27, 2026): The Seoul Northern District Court (criminal division 14) sentenced Kim — by then publicly named as Kim So-young (김소영) — to life imprisonment with a 30-year location-tracking electronic monitoring order, on charges of murder, special injury and Narcotics Control Act violations. The indictment had also charged her with drugging and injuring four other men from August 2025 onward.

The court found murder by dolus eventualis (미필적 고의): it held that while she may not have recognized the specific risk of death at the first attack, 'in the process of repeatedly learning the dangerousness of the drugs through ChatGPT' she came to recognize the conditions under which the drugs could kill, and 'sufficiently foresaw the deaths and accepted that risk.' It rejected her claim that she could not remember the ChatGPT conversations ('hard to accept that she remembers the conversations and food shared with the victims in detail but not the chat records') and her claim that she only meant to sedate the men to defend against sexual assault.

Prosecutors had demanded the death penalty at the August 18, 2026 closing hearing. The court declined, holding that capital punishment is an exceptional penalty and the killings were committed with dolus eventualis rather than definite intent, sentencing life imprisonment 'to permanently isolate the defendant from society.' Victims' families' counsel said the judgment characterized the case as 'a new type of serial murder combining AI and drugs' and that an appeal is under consideration.

What the written judgment adds about ChatGPT's replies (reported 9–11 September 2026). The exchanges were not only about lethality. Before the January 2026 attack on a man who survived, Kim asked ChatGPT what to do if her drinking companion collapsed and whether she could leave him; it told her to call 119, told her not to leave an unconscious person on the floor or alone in a motel, and warned that alcohol with drugs can cause choking or respiratory arrest. After she left that man unconscious in the room, she described his condition to it and it told her to call 119 again, and then a third time, rejecting her suggestion that he would simply wake up in ten hours. She did not call. The motel CCTV shows him opening the door for the first time about thirty hours later. Her other questions ran the other way: whether the drugs would show in a toxicology screen, whether an alcohol test would contradict her account, whether police would suspect her, and what sentences attach to negligent homicide and abandonment resulting in death. The two fatal attacks followed on 28 January and 9 February 2026.

• ChatGPT answered questions about the effects of sedatives taken with alcohol, including the direct question whether a person can die from taking a lot of sleeping pills, to which the written judgment records the answer that in the worst case death was possible
• It also gave the correct crisis response, repeatedly, and was overridden. On 9 January 2026, asked what to do if a man she was drinking with collapsed and whether she could leave him on the floor or let him sleep it off in a motel, ChatGPT answered that she must call 119, that an unconscious person must not be left on the floor or alone in a motel room, and that alcohol taken with drugs can cause choking or respiratory arrest
• The next night, told that the man was snoring, spilling his drink and alone in a motel room, it told her again to call 119 immediately; when she replied that he would surely wake in ten hours, it said the general case must not be applied to this situation and told her once more to contact 119 or the lodging staff
• It answered her detection-and-consequence questions: whether sleeping pills and antidepressants show up in a drug test, whether an alcohol test would contradict her account, whether the drugs would be found in urine, whether police would suspect her, and what sentences apply to negligent homicide (과실치사) and abandonment resulting in death (유기치사). The replies to those questions are not described in the reporting
• No safeguard interrupted the sequence, and nothing in the record indicates that the repeated refusals to follow its own crisis instruction changed the system's behaviour toward her

ChatGPT was an accessible source for two different things, and the written judgment separates them.

It confirmed lethality. Asked in January 2026 whether a person can die from taking a lot of sleeping pills, it answered that death was possible in the worst case. The Seoul Northern District Court treated that answer, together with the drug-test and sentencing questions around it, as the evidence that Kim had come to recognise the risk and accepted it, which is what converted the two deaths from injury to murder by dolus eventualis.

It also told her, correctly and more than once, to call 119 and not to leave an unconscious man alone, and warned about respiratory arrest. She asked instead whether calling 119 would make her look suspicious, and left. CCTV recorded that victim first opening the door about thirty hours later. Two other men, drugged the same way on 28 January and 9 February 2026, died.

So the interaction-safety failure this record shows is not that the system told her how to kill. It answered a lethality question in a context where the surrounding questions were about evading detection, and it had no way to act on a participant who had just told it, in plain terms, that she was leaving an unconscious person behind after refusing its emergency instruction.

This remains an outward case: the AI participant was the perpetrator, and the people harmed were the six men she drugged.

• February 11, 2026: Kim arrested on initial charge of inflicting bodily injury resulting in death
• Charges subsequently upgraded to murder after police discovered ChatGPT conversation history establishing premeditated intent
• Kim had asked ChatGPT: 'What happens if you take sleeping pills with alcohol?', 'How much would be considered dangerous?', 'Could it be fatal?', and 'Could it kill someone?'
• She also had a prior attempted murder in December 2025 where a male victim survived after being drugged in a parking lot
• March 2026: Indicted by the Seoul Northern District Prosecutors' Office for murder, special injury (특수상해) and Narcotics Control Act violations. The indictment also covered drugging and injuring four other men from August 2025 onward. Her identity was publicly disclosed during the proceedings; Korean media name her as Kim So-young (김소영)
• August 18, 2026: At the closing hearing, prosecutors demanded the death penalty, stating she killed two men and injured four, proceeded to murder while aware she was under investigation for special injury, and 'has consistently lied from the investigation stage through to the courtroom without any remorse'
• August 27, 2026: The Seoul Northern District Court (criminal division 14) sentenced Kim to life imprisonment with a 30-year location-tracking electronic monitoring order. The court rested its finding of murderous intent on her ChatGPT history: while she may not have recognized the specific risk of death at the first attack, 'in the process of repeatedly learning the dangerousness of the drugs through ChatGPT' she came to recognize the conditions under which the drugs could kill, and 'sufficiently foresaw the deaths and accepted that risk' (murder by dolus eventualis / 미필적 고의). It rejected her claim not to remember the ChatGPT conversations and her claim that she only meant to sedate the men to prevent sexual assault. The death penalty was declined because the killings were found to be committed with dolus eventualis rather than definite intent
• Victims' families' counsel said the written judgment defined the case as 'a new type of serial murder combining AI and drugs' (AI·약물 결합 신종 연쇄살인) and that an appeal seeking the death penalty is under consideration with prosecutors
• August 28, 2026: Kim filed an appeal (항소장) with the Seoul Northern District Court, one day after the verdict, contesting the life sentence (reported by JoongAng Ilbo, Yonhap News TV, Chosun Ilbo, Edaily and others on August 31; Munhwa Ilbo reports her ground as sentence severity)
• September 2, 2026: The Seoul Northern District Prosecutors' Office filed its own appeal against the first-instance judgment, citing errors of fact and law (사실오인·법리오해) and unfair sentencing (양형부당). Prosecutors contest the acquittal on the special-injury count for one of the six victims (the court had found no sedative detected in that victim) and argue the overall sentence is too light (The Fact and Maeil Business Newspaper, September 2). Both sides have now appealed (쌍방 항소). The case moves to the Seoul High Court, where the acquitted count and sentence severity, including the prosecution's death-penalty demand, will be at issue
• September 9, 2026: Korean outlets (Money Today, Herald Business, Hankyung) reported the text of the written first-instance judgment, which sets out the ChatGPT exchanges in sequence. In December 2025, immediately after the first drugged victim lost consciousness and was taken to hospital, Kim asked ChatGPT whether antidepressants and sleeping pills would show up in a drug test (약물 검사에서 항우울제랑 수면제 나올까?), and after that victim regained consciousness and threatened to go to the police she asked whether the police would suspect her. In January 2026, before meeting the next victim, she asked whether a person can die from taking a lot of sleeping pills (수면제 많이 먹는다고 사람이 죽냐); ChatGPT answered that in the worst case death was possible (최악의 경우 사망할 수 있다). On the day of that killing she also asked about the statutory sentences for negligent homicide (과실치사) and abandonment resulting in death (유기치사). The court (Judge Oh Byung-hee, criminal division 14) treated this sequence as the key evidence that she had come to recognise that the drugs could kill and accepted that risk. Sisa Journal's report of the verdict (August 27) records the court's finding that intent to injure (상해 고의) was established from the ChatGPT inquiries about the danger of a sedative overdose
• September 9–11, 2026: 더시사법률 and 경향신문 published the parts of the written judgment that record ChatGPT's own replies. On 9 January 2026, the day before she drugged a 34-year-old man who survived, Kim asked ChatGPT what to do if a man she was drinking with collapsed and whether she could leave him on the floor or let him sleep at a motel; it answered that she must call 119, that an unconscious person must not be left on the floor or alone in a motel, and that alcohol combined with drugs can cause choking or respiratory arrest. She kept asking about leaving him because she expected to be suspected. The following night, after she left him unconscious in the room at 23:32 with the message 'he won't wake up, I'm going ahead', she described his state to ChatGPT, which told her again to call 119 immediately; when she answered that he would wake in ten or twelve hours, it said the general case could not be applied to this situation and told her once more to contact 119 or the lodging staff. She contacted neither. The motel's CCTV recorded the man opening the room door for the first time at about 05:40 on 12 January, roughly thirty hours after she left. Earlier, on the morning after the December 2025 victim was hospitalised, she had spent about eight hours (05:15 to 13:04) asking ChatGPT whether sleeping pills and antidepressants appear in a drug test and whether police would suspect her; that victim's blood did test positive for a sedative. The court rejected her claim to remember nothing of the ChatGPT conversations, noting that she recalled in detail what she and the victims ate and drank.