São Gabriel da Palha, Espírito Santo: a 36-year-old farmer was held in pre-trial detention for 54 days after OpenAI reported his ChatGPT messages about killing his eight-year-old son to the FBI, which passed them to Brazilian authorities; a court released him on 13 August 2026 because the investigation had not been completed, with an ankle monitor and a ban on contacting the child, and no charges had been filed; he denies intending harm and his lawyer says he was talking nonsense to a chatbot

BBC News Brasil (14 August 2026) reported that for about two months a 36-year-old farmer, given the pseudonym João, held conversations with ChatGPT which, according to the Espírito Santo civil police, indicated a possible plan to kill his eight-year-old son, commit attacks and take his own life; one message read that he had offered 50,000 to a man to kill him and his son, who refused because a child was involved. The conversations triggered an alert in the AI's safety system, leading OpenAI to report the interaction to the FBI, which contacted the Brazilian authorities; the case reached the Espírito Santo civil police on 16 June through the Ministry of Justice's Ciberlab. Three days later, and a day before the date on which police say the plan would be put into practice, he was arrested as he left his home in rural São Gabriel da Palha. The head of the cybercrime unit, Brenno Andrade, said the case was prioritised because of its gravity and because the father supposedly planned to kill the child by the 20th; that the suspect had no affective relationship or contact with the child beyond paying maintenance, which police believe was a motive; that everything the platform provided was corroborated when police arrived, that he confirmed some searches concerning his son but denied the intention to kill him; and that his phone was seized to establish whether he had contacted a hired killer. He was held in pre-trial detention from 19 June; on Thursday 13 August he was released after the court granted habeas corpus, finding that after more than 50 days the inquiry had not been concluded and no charge had been brought, and imposed an electronic ankle monitor and a ban on approaching or contacting his son and the child's mother. His defence lawyer told the BBC that talking nonsense to an AI is not a crime and that no concrete step was taken. According to investigators OpenAI provided details of the user and the messages he sent but not ChatGPT's replies; OpenAI did not answer the BBC's question why and said it may notify law enforcement when it detects a credible and imminent risk of harm to others. BBC News Mundo (16 September) published a Spanish text produced by AI translation from an English-language BBC original and reviewed by a BBC journalist, with the same facts. Both BBC versions report that, per the police chief, he tried to get round ChatGPT's restrictions by rephrasing questions and researched toxic substances including ricin. G1 (27 June) reported that the man was arrested on 19 June, a day before the date he had chosen, after OpenAI informed the FBI, which passed the information to the Ministry of Justice and on to the state police; that he had written that he would have a gun, a rope and poison and would carry out attacks in public places; that OpenAI said its systems are designed to keep people safe; that the arrest was made under judicial warrants by the state cybercrime unit, and that the police chief told TV Gazeta this was only the third such case in Brazil. Migalhas (26 June) summarised the G1 account: a 36-year-old arrested in São Gabriel da Palha on the 19th, suspected of planning his son's death to stop paying maintenance, with lawyers' analysis of whether planning in a chatbot conversation is punishable.

The user was held in pre-trial detention for 54 days without charge on the strength of his ChatGPT messages as reported by OpenAI, then released under electronic monitoring and a no-contact order covering his son and the child's mother while the investigation continues. Police say the report was corroborated at the scene and prevented a planned killing; he and his lawyer dispute any intent. The consequence to him is undisputed; whether it was warranted is contested.

Pre-trial detention from 19 June 2026; habeas corpus granted on 13 August 2026 (54 days) on the ground of excessive delay in concluding the police inquiry, with no indictment filed; precautionary measures imposed (electronic ankle monitor; prohibition on approaching or contacting the son and the child's mother). The police investigation continues, including forensic examination of the seized bottles and his phone to establish whether he contacted a hired killer as he told ChatGPT. OpenAI supplied user and message details but not the chatbot's replies; a criminal-investigation researcher quoted by the BBC (André Glitz) says access to the full conversation would require a court order and international legal cooperation. The arrest was carried out under judicial warrants (G1) and the pre-trial detention was court-ordered. Sources describe it as the first such AI-generated alert received by Espírito Santo and the third in Brazil.