ChatGPT sí puede analizar tus chats: el caso Quilmes expone cómo OpenAI maneja la privacidad

A 15-year-old in Quilmes had a conversation with ChatGPT about carrying out an attack at his school. He discussed killing classmates, acquiring tactical knives and clothing, and mentioned he was a minor. The system flagged it. An alert traveled from OpenAI to the FBI, then to Argentina's Federal Police Counterterrorism Unit. On August 27, officers entered his home and seized two cell phones, tactical gloves, camouflaged clothing, a skull-patterned balaclava, a ski mask, and protective goggles. They found a photograph of him trying on masks. The case was filed as attempted public intimidation.

The incident opens a door most people don't think about when they use ChatGPT. The chats are not private exchanges between a user and an isolated machine. Every message travels through OpenAI's infrastructure. The company runs automatic systems in parallel—classification models designed to spot patterns of harm, not just dangerous keywords. A conversation about weapons might be research, journalism, or history. Another might show, across multiple messages, a concrete intention to hurt someone. OpenAI says its systems try to tell the difference.

When a conversation crosses certain thresholds, it moves to human review. A specialized team reads it. They look at context, at the full arc of the exchange, at what the person said they wanted to do and when. OpenAI states plainly that it will report to law enforcement when it believes there is an imminent and credible risk of harm to others. That is what happened in Quilmes. The company does not require a single message containing every detail of a planned attack—it evaluates the conversation as a whole.

There are distinctions worth understanding. Using a chat to train a model is different from processing it for security. ChatGPT lets users turn off "Improve the model for all," but that does not stop abuse detection. Temporary chats do not appear in history, do not generate memories, do not feed training—but they can still be reviewed for misuse during their 30-day retention window. An alert from OpenAI is separate from a later request for data by authorities. The company's privacy policy allows it to hand over information in response to legal process and certain emergencies.

In the Quilmes case, the investigation advanced using the IP address tied to the connection. It was never made public which data OpenAI provided first, what the FBI added later, or what Argentine investigators obtained themselves. Alejandro De Mena, an assistant prosecutor in Quilmes, explained that the minor had asked the AI what was needed to carry out an attack on a school and its students. The full content of the conversations and ChatGPT's specific responses have not been disclosed—a gap that matters, because it would show what the adolescent wrote versus what assistance, if any, the system actually provided.

OpenAI is not alone. Google's Gemini and Anthropic's Claude also run automatic systems and human review, though each company sets its own rules. Google collects what users write or say to Gemini, along with files and images they share. Some of that can be reviewed by humans to improve services and maintain safety. Google explicitly recommends not entering confidential information you would not want a reviewer to see. Gemini uses automated systems plus human review to detect abuse—dangerous activities, fraud, harm. It allows users to limit activity storage and use temporary conversations. Even with history off, Google keeps chats for a limited time to provide service and protect users and the public. The company can hand over data to authorities in response to valid legal requests and has special emergency procedures for risk of death or serious physical harm, including bomb threats and school shootings.

Claude works similarly. Anthropic uses classification models trained to spot certain violations during a conversation. By default, employees do not access chats, except for those sent voluntarily as feedback or those requiring review to enforce safety policies. Access is restricted to specific team members. Anthropic also separates training from security. Users can decide whether their conversations improve Claude. A conversation flagged for possible violations can still be reviewed and kept longer. The company delivers information in response to valid legal process and emergencies with risk of death or imminent serious harm.

The Quilmes case has precedent. In Florida, Darren Zhou, a 25-year-old former Goldman Sachs analyst, used ChatGPT for months to discuss his ex-partner, eventually writing repeated threats of rape and murder. The messages included her routines and possible attack scenarios. OpenAI reported it to the FBI. Zhou was arrested. The public court file contains some of his messages but not ChatGPT's responses. In Strasbourg in April, a 37-year-old spoke with ChatGPT about obtaining a weapon and killing an intelligence officer. An alert reached the FBI and French authorities, who sent a specialized unit. They found no weapons. The man was hospitalized and the criminal investigation was closed. In Tumbler Ridge, Canada, a account linked to someone who later carried out a school attack had been detected by OpenAI's systems months before. Company employees debated whether to alert authorities but decided the messages did not yet meet the threshold. That case shaped OpenAI's current policy.

A conversation with a chatbot remains information delivered to an online service that must process it and applies its own rules of privacy and security.