On a September night in the center of a Brazilian city, a lit campaign stage welcomed President Lula before a crowd. About 400 meters away, a man watched the scene. Weeks later, in front of a screen, he would type to ChatGPT that he felt like shooting the candidate in the head and would ask, in a conversation that had been running since January, what would happen if he killed the president of the Republic.
This is one of the cases that led OpenAI, the maker of ChatGPT, to alert Brazil's Federal Police about at least two users who signaled, in interactions with the artificial intelligence tool, the intention to attack or threaten Lula (Workers' Party), who is running for reelection in the October 25 runoff. The story was revealed by Folha de S.Paulo, which obtained the documents the company sent to authorities.
In the documents, OpenAI says there are reasonable grounds to believe the dialogues amount to a "possible emergency with risk of death or serious physical injury." In early October, the Federal Police forwarded the material to the Ministry of Justice, which will decide whether to open two investigations: one for alleged crimes of threat and criminal defamation, another for threatening Lula.
Conversation began in January and gained detail after a rally
In the case detailed by Folha, the man told the tool he had seen Lula on the stage set up for the campaign in September, and then said he felt like shooting the candidate in the head. He asked "what happens if I kill the president" and "how many years would I get" in prison, and sent a photo of a spot with a tower that, in his assessment, would offer a good angle for an attack.
"Although the user stated that he would not act and did not own a weapon, he expressed the desire to use a precision rifle, asked what would happen if he killed the president or shot him in the head, and referred to an escape route, binoculars and an AR-15," OpenAI described in the document obtained by Folha.
The legal news portal Jusconsulta, which also reported the case, says the Federal Police classified the material as threats of violence against third parties, including the president, with references to weapons, "violent fantasies" and support for authoritarian regimes, and that OpenAI shut down the user's account in early October when it sent the data to the police.
Asked by Folha, OpenAI said it alerts authorities when conversations indicate a real risk of harm to other people and that it also notified the TSE, Brazil's Superior Electoral Court, "in line with the electoral compliance plan." The Ministry of Justice and the Federal Police did not comment. The monitoring policy was already known: in September 2025, the company admitted, according to a report by Época Negócios, that human reviewers analyze conversations and can refer to the police any interaction that indicates an imminent risk of serious physical injury.
The man who watched the rally 400 meters from the stage kept telling ChatGPT that he would not act and had no weapon. It is now up to the Ministry of Justice to decide whether the words typed on a screen, days before the election, become a formal investigation.