Meli and Liel, are you listening? The words that will cause AI to report you to the authorities

The story of Meli and Liel Yahalomi, who disappeared for days and occupied the authorities and their family members, raises a particularly intriguing question in the era of artificial intelligence: how much does a conversation with AI really remain private, and are there situations where the things written in it could reach the authorities?

During the investigation, the investigators managed to trace some of the conversations the two had with Google's Gemini, even though some of the searches had been deleted. The questions they directed to the artificial intelligence dealt, among other things, with ways to disappear and make it difficult to locate them, and the information helped the investigators trace their plan.

However, this story does not only concern Gemini. Millions of people conduct personal conversations every day with GPT, Claude, Copilot, and other AI tools, and sometimes share with them information they would not tell almost anyone. But these services have safety mechanisms, terms of use, and policies regarding the disclosure of information to the authorities.

So what might trigger a red flag?

At the top of the list are real threats to harm another person, an intention to commit a violent act, or planning that indicates an imminent danger. In GPT, for example, OpenAI explicitly states that when its systems identify a user who is planning to harm others, the conversation may be moved to a dedicated review path. If human reviewers conclude that there is an immediate threat of serious physical harm to another person, the company may refer the case to law enforcement authorities.

That is, sentences that express a concrete threat, along with details about a target, time, or method of execution, may take on a completely different meaning than a general or theoretical question. The more clearly things point to a real danger, the more likely they are to trigger the service's safety mechanisms.

Also in Claude, Gemini, and Copilot, there are mechanisms designed to identify and block dangerous or prohibited uses. Anthropic, which operates Claude, even notes that in cases where a review is required to enforce the terms of use, dedicated staff members may gain access to the conversation.

Google prohibits the use of Gemini for dangerous or illegal activities, and Microsoft also operates safety and data retention mechanisms that vary depending on the product and the type of account.

And there is another way in which conversations can reach the authorities — an existing investigation. Large companies in the field of artificial intelligence receive official requests from law enforcement agencies for information about users, and in cases where the request meets legal requirements, they may provide information in their possession. OpenAI, Google, Anthropic, and Microsoft all publish procedures for handling such requests.

In emergency situations, when there is a fear of death or serious physical harm, some companies also operate special channels for handling urgent requests from law enforcement authorities. And this is exactly what should be remembered before sharing particularly sensitive information with AI — the conversation may feel private, but it takes place within a system that stores and processes data, operates safety mechanisms, and is subject to legal requirements.