SAU PAULO – A 36-year-old man in Brazil lands in jail after authorities were alerted to ChatGPT conversations in which he allegedly discussed plans to kill his eight-year-old son, attack police and other people, and eventually take his own life, according to Brazilian investigators.
The case raised major new question for police, courts and AI giants like What should happen when an AI system detects what appears to be a credible threat before a crime is actually committed?
According to police, the man shared messages with ChatGPT for around months. Investigators say he discussed violence against his son and other potential targets, as well as suicide. The suspect was arrested on June 19 in a rural area of Brazil after OpenAI, the company behind ChatGPT, allegedly alerted the US Federal Bureau of Investigation (FBI). FBI then passed information to Brazilian authorities.
Breno Andrade, head of a local cybercrime police station, said the case was treated as a priority because investigators had information suggesting the suspect was planning to kill his son by June 20.
Police said the suspect had little or virtually no contact with his son apart from making child-support payments. Investigators believe anger related to those payments could have been among the possible motives, although that theory remains part of the investigation.
The investigation took an even more serious turn when police searched the suspect’s home. Authorities said they found four bottles containing unidentified substances and a rope tied with a noose. Police said the items could indicate preparations for suicide.
Investigators also allege that the man searched for information about poisonous substances, including ricin, and discussed possible attacks on schools, churches and police officers.
The suspect allegedly asked whether killing several criminals at a drug dealers’ hideout would make him famous before killing one or two police officers and dying. In another message, he reportedly discussed buying a pistol and shooting at police so that officers would kill him in response.
Police further claim that he attempted to get around ChatGPT’s safety restrictions by wording his questions in different ways.
The suspect denies intending to harm his son. His lawyer, Pedro Paulo Pessi, argued that his client was simply engaging in meaningless and unserious conversations with a chatbot.
After spending 54 days in custody, the man was released on August 13 because of delays in the investigation. No formal charges have yet been filed, but the investigation continues. A court has imposed precautionary measures, including electronic monitoring and a ban on contacting his son and the child’s mother.
Investigators are now conducting a forensic examination of his phone and awaiting laboratory results from the seized materials. They are also examining whether he actually contacted a contract killer, as he allegedly claimed in his ChatGPT conversations.
The case raises concerns about AI-generated evidence. Investigators say OpenAI provided the user’s identity and details of messages he sent, but not ChatGPT’s responses. Legal experts said the complete conversation would be crucial in determining the context and whether the chatbot merely answered the user or actively influenced the direction of the discussion.
AI is becoming modern interactive system capable of identifying patterns associated with potentially dangerous behavior. But she and other experts have warned that such systems must distinguish genuine, imminent threats from emotional expression, hypothetical scenarios or provocative statements.
