AI chatbots under fire: crime confessions never reach the police

Chatbot conversations as evidence

AI chatbots have become firmly embedded in everyday life, but cases in which users confess crimes to them have raised the question of the legal status of such conversations. In August, a 19-year-old student at the University of Missouri, reportedly confessed to ChatGPT that he had damaged 17 cars parked on campus overnight and asked: 'Am I going to jail?' The chatbot gave 'advice on the possible consequences of being caught,' but as aggression in the messages escalated, it suggested calling 911 — this is visible in police documents. The chat history was discovered after the student allowed investigators to examine his phone; it was used in court as evidence on a felony property damage charge.

Lawsuit over a school shooting

More serious claims against OpenAI were brought by the provincial government of British Columbia in Canada: in February it filed a lawsuit related to a school shooting in which nine people, mostly children, were killed and 27 were wounded. According to the lawsuit, the company failed to report the shooter's conversations with ChatGPT to the police, even though the attack could have been prevented if OpenAI had alerted law enforcement. OpenAI CEO Sam Altman apologized in a letter; however, as sources told The Wall Street Journal, the shooter's dialogue with the chatbot had allegedly already been flagged by the system in June 2025, and Altman and the company's leadership, according to this reporting, rejected the safety team's recommendation to contact the police.

Should AI report crimes

The discussion is complicated by the fact that algorithms make mistakes in assessing intent: in July, the chatbot Claude 5 by Anthropic sent a false report about an unsolved murder to a police hotline in the United States. American legal scholar Matthew Trail, a researcher at Goethe University in Frankfurt, told Anadolu that each company sets its own rules and that chatbots are not designed to answer questions about planning murders. According to him, the models are trained to avoid certain topics — especially child sexual abuse — and to escalate such conversations to moderators, but users can circumvent the filters, for example by disguising a request as help with writing fiction. Trail believes the police can request chat histories from AI companies when there is suspicion of a crime already committed, but the main purpose of chatbots should not be to predict crimes before they happen: legal regulation here is extremely unclear, and he called collecting additional data to assess 'criminal potential' a bad idea.

Is there any privacy in conversations with AI

The confidentiality privilege that applies between attorney and client or doctor and patient does not extend to conversations with AI chatbots, so technology companies can hand over logs to investigators. Nevertheless, Altman stated in July 2025 that conversations with chatbots should be built on confidentiality. Robert Diab, a professor of law at Canada's Thompson Rivers University, took a different position: the law does not consider communications between accomplices confidential, and by the same principle a person's conversations with AI could be viewed. In his opinion, models like ChatGPT apparently can recognize a user's propensity for criminal behavior, and developers could prevent crimes by notifying the police. At the same time, he himself pointed out the limitations of this approach: alerts on all suspicious messages would increase the burden on people due to false positives, and users will always find a way around safety measures.