OpenAI reports Brazilian man's alleged child harm plot to FBI
OpenAI detected and reported a Brazilian man's conversations on ChatGPT describing an alleged plan to harm his child and others, leading to his arrest by Brazilian police. The case raises questions about how AI platform conversations can be used as evidence in criminal investigations.
What happened
Over approximately two months, a 36-year-old farmer in Brazil used ChatGPT to discuss what Brazilian police described as a plan to kill his 8-year-old son, commit attacks, and take his own life. In the conversations, he researched toxic substances, stated he had offered money for someone to kill the child, and expressed satisfaction at seeing others suffer.
ChatGPT's security system flagged the conversations, prompting OpenAI to report them to the FBI, which contacted Brazilian authorities. The case reached the Polícia Civil do Espírito Santo on 16 June via Ciberlab, the cybercrime unit of Brazil's Ministry of Justice. The man was arrested on 19 June, reportedly one day before the alleged planned date of the incident.
Reported impact
The conversations triggered alerts in ChatGPT's security system, leading to intervention by law enforcement. Police described the case as a priority due to its severity and the imminent nature of the alleged threat.
Index analysis
The incident demonstrates post-deployment monitoring of ChatGPT conversations for safety risks related to harm to children. OpenAI's security system detected content discussing harm to a minor and escalated it to law enforcement, raising questions about the boundaries and use of AI conversation data as evidence in criminal investigations.
Governance controls
OpenAI's security systems include mechanisms to detect and report conversations suggesting plans to harm others, particularly children. The system flagged the conversations and resulted in notification to law enforcement.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.