Not KakaoTalk, Not Search History — 'ChatGPT Prompts': Why Korean Police Have Started Using AI Chat Logs as Criminal Evidence
As generative AI becomes mainstream, Korean police are increasingly prioritizing ChatGPT conversation logs over KakaoTalk chats and web searches during digital forensics. In the Gangbuk motel serial murder case, a suspect's AI queries proved decisive in establishing premeditated intent, and the legal community now faces new challenges surrounding privacy, evidence admissibility, and the risk of evidence destruction.
"Can you die from mixing sleeping pills and alcohol?" — This single line a suspect left in ChatGPT overturned a murder charge the defendant had denied was intentional.
TL;DR
- Korean police are now checking ChatGPT conversation logs first during phone forensics — before KakaoTalk or Google search histories.
- Prompts are full-sentence queries, not mere keywords, so intent and motive emerge in concrete detail.
- In the Gangbuk motel serial murder case, AI queries became the central basis for applying a murder charge.
- For court admissibility, hash-value verification of the original data is mandatory.
- The potential for evidence destruction via the 'Temporary Chat' feature has emerged as a new legal issue.
What Happened: The Facts
According to reporting by Yonhap News on February 25, 2026, front-line police investigators are increasingly making generative AI apps such as ChatGPT and Gemini the very first thing they check when forensically examining a suspect's phone.
The defining case is the Gangbuk-gu motel serial murder. A female suspect in her twenties is accused of spiking drinks with drugs to kill two men and leave a third unconscious. The suspect claimed the acts were impulsive, but police found queries including "Tell me the lethal dose of sleeping pills" and "Can you die from mixing sleeping pills and alcohol?" in her ChatGPT history and upgraded the charges from accidental manslaughter to murder.
Why AI Prompts Are Drawing Attention
"Web searches consist of keyword strings centered on individual words, but AI prompts are written in sentence structure, so the actor's intent and context are recorded in specific detail."
— Prof. Jeong Du-won, Department of Forensic Science, Sungkyunkwan University (cited by Yonhap News)
Traditional forensics analyzed Naver/Google search terms and KakaoTalk messages. Generative AI conversations are structurally different.
| Category | Traditional Web Search | AI Prompt |
|---|---|---|
| Input format | Words / keywords | Sentences / questions |
| Intent exposure | Indirect / ambiguous | Direct / specific |
| Context preservation | Low | High |
| Evidentiary value | Moderate | High |
Unlike attorneys who hold privilege, AI has no legal protection. Whatever a suspect confides to an AI can become direct evidence for investigators.
Who Is Affected?
Law enforcement: Gains a groundbreaking tool for proving intent and motive. Forensic procedures and priorities are being rapidly restructured.
Defense attorneys: Reviewing a client's AI chat history before meeting them has become the new standard. Some attorneys now refuse to take a case if the client refuses to share those records.
General users: They now face the reality that "confiding in an AI" can become courtroom evidence at any time.
AI service providers (OpenAI, etc.): Facing a simultaneous rise in data requests from investigators and risk of eroding user trust.
How Long Will This Last?
As AI usage expands across all aspects of daily life, this trend is likely to persist long-term. If AI forensics is officially incorporated into standard Korean police procedures, it will feed into a precedent-building phase alongside the courts establishing standards for admissibility.
Legal Issues: Is Evidence Admissibility Recognized?
ChatGPT conversation logs are a form of electronic information. Under Article 308 of the Criminal Procedure Act (principle of free evaluation of evidence), they may be admitted if collected through lawful procedures.
- Key requirement: The hash value of the original must match the submitted copy → Integrity verification is essential (Supreme Court 2014Do10978)
- Limitation: A conviction is not possible if the defendant's confession is the only evidence (CPA Art. 310) → AI records must be used alongside corroborating evidence
- New issue: The potential for evidence destruction by exploiting the 'Temporary Chat (no history saved)' feature
Secondary Issues: Derivative Debates
- Privacy violations: All intimate conversations and medical information shared with an AI become potential subjects of investigation
- AI ethics: The irony of a user's honest disclosure to a trusted AI being used against them
- Platform responsibility: No clear standard for how far OpenAI, Google, and others should cooperate with law enforcement requests
- Legislative gaps: A lack of domestic investigative guidelines or laws explicitly governing AI forensics
Checklist: What You Need to Know Right Now
Reference Links
- Police Now Check 'ChatGPT Records' Before 'Search Records' — Yonhap News (2026.02.25)
- Police Are Checking ChatGPT Prompts… Will Courts Actually Accept Them as Evidence? — LawTalk News (2026.02.25)
- AI Chat Logs Rising as Investigative Evidence — Today Newspaper (2026.02.25)
- It Wasn't KakaoTalk or Search History… What Police Check First These Days — Nate News (2026.02.25)
Image Credit
- ChatGPT Logo: Wikimedia Commons — ChatGPT logo.svg (Public Domain)