How ChatGPT identified a father's plan to kill his son and alerted authorities

Father is arrested in Espírito Santo after revealing his intention to kill his son to ChatGPT
The father arrested in Espírito Santo for planning the death of his 8-year-old son was denounced after talking about the case with ChatGPT. OpenAI, creator of the artificial intelligence assistant, reported what happened to the FBI, which passed the information on to Brazilian authorities.
The man was arrested on June 19, one day before the date he had chosen to commit the crime. Before, he wrote on ChatGPT that he had a gun, a rope and poison and that he would carry out attacks in public spaces.
Do you have any suggestions for a report? Send it to g1
Sought by g1, OpenAI said that its systems are designed to keep people safe and that the case shows the importance of these measures.
"When we identify conversations that indicate an imminent and credible risk of harm to others, we may notify the appropriate authorities," the company said.
Father who planned to kill his son in Espírito Santo told ChatGPT that he offered the gunman R$50,000
Reproduction/PCES
In the messages, the man wrote that he tried to hire a gunman for R$50,000 to kill his son, but that the proposal was refused after informing that the victim was a child.
"I wanted to know where this desire to kill people comes from. I like to see someone else suffer", he said, in one of the messages.
Assistants like ChatGPT monitor everything that is said in conversations, even though most of the review is done by automated systems, said Álvaro Machado Dias, professor at the Federal University of São Paulo (Unifesp).
"In the case of the father from Espírito Santo, the border was crossed because there was an identified victim, method, means and close date. In other words, the alert triggered was of maximum severity", he said.
See messages from the father to ChatGPT while he planned to kill the 8-year-old son
How ChatGPT detects suspicious cases
OpenAI says it begins the analysis through automated systems, which include AI models capable of categorizing content and carrying out more complex analyzes on the context of the conversation.
The stage also includes databases that bring together files that have already been flagged, as happens in cases of child sexual exploitation, and lists of terms prohibited in conversation.
If the conversation is flagged by these mechanisms, it is sent to human moderators. They analyze activity to indicate whether there has been a policy violation and whether the user may actually commit an act of violence.
Moderators can classify the activity as a low-risk case or deactivate the account and contact people and organizations better able to resolve the situation, as happened in Espírito Santo.
ChatGPT-4o was announced and launched by OpenAI on May 13, 2024
NurPhoto via Getty Images
The company was sued in 2025 on charges of contributing to the suicide of a 16-year-old in the United States, instead of adopting safeguards such as directing him to crisis support lines.
In a statement from April 2026, the company said it worked for the ChatGPT rejects requests to facilitate cases of violence, while seeking to make the service differentiate problematic cases from neutral questions for historical or educational reasons, for example.
"Some security risks only become evident over time: a single message may seem harmless, but the broader pattern across a long conversation - or across multiple conversations - may suggest something more concerning," said OpenAI.
Communication with authorities
Reports to authorities are common in cases identified on social networks, but they do not happen frequently when they appear on assistants like ChatGPT. This was only the third case of its kind in Brazil, delegate Breno Andrade told TV Gazeta, a Globo affiliate in Espírito Santo.
The trend is for the number to increase as more people use these systems as notepads or as partners to report what they are doing, said Álvaro Machado Dias, from Unifesp.
But the notification to the FBI, and not to a Brazilian authority, is inadequate, said the expert. According to him, the correct channel would be the Federal Police or the Ministry of Justice, although it is possible to understand the logic of OpenAI, based in the United States and in contact with American authorities.
"The crime would be in Brazil, the victim was in Brazil, the suspect was in Brazil, and the arrest depended on the Brazilian State. Whoever provides services here cannot treat Brazil as a consumer market and the United States as the natural forum for any conflict", he said.
See messages sent by a father to ChatGPT while planning to kill his 8-year-old son years in Espírito Santo
Reproduction/PCES
At the same time, Brazilian authorities need to offer a direct channel to receive this type of alert, which allows the investigation to be speeded up, he highlighted. "Speed is decisive because AI compresses the distance between intention, plan and execution."
For the professor, there is no doubt that communicating the case to authorities was the correct decision. "People talk to ChatGPT like someone talking to a therapist or a lawyer, but AI is neither one thing nor the other, and therefore carries no professional secret."
See messages sent by a father to ChatGPT while he planned to kill his 8-year-old son
How the investigation went
The FBI, the American equivalent of the Federal Police, forwarded the complaint to the Ministry of Justice and Public Security, which forwarded the material to the Civil Police of Espírito Santo.
The information was then sent to the Specialized Police Station for the Repression of Cybercrimes of Espírito Santo, which carried out the arrest in compliance with court warrants.
Delegate Ícaro Olímpio said that the man denied the accusation, but was detained based on the history of conversations with ChatGPT provided by OpenAI.
"We had enough elements to be able to prevent this serious crime that was about to happen", he stated. "We are convinced that we avoided what would have been acts of extreme violence."

Source: G1