Artificial Intelligence

Privacy risks with Meta AI: how to protect your online conversations

More and more users entrust the virtual assistant with personal, medical or financial information. In the US, conversations can appear on the Discover feed and be visible to anyone

by Chiara Ricciolini

20 June 2025

3' min read

3' min read

Asking for advice on romantic problems, reading one's blood test, getting opinions on managing one's money. These are just some of the most common uses that people around the world are making of Meta AI, Meta's artificial intelligence assistant, which uses Llama 3.2, the large language model developed by the company. In the US, it is also available through the Discover app, a kind of social network where users can see others' conversations, compare prompts and look at images generated by others.

We have known for some time that one of the most thorny aspects of the digital assistant launched by Mark Zuckerberg - who is investing heavily in the development of his own artificial intelligence, as evidenced by the 14 billion deal signed with Scale AI, the data labelling company that employs thousands of freelancers around the world to catalogue and give feedback to artificial intelligence - concerns precisely its forced integration into platforms. The virtual assistant is located directly inside the apps of WhatsApp, Instagram and Facebook, and cannot be uninstalled. If the user does not intend to use it, he can only refrain from opening the chat or uninstall the app completely.

Chat with Meta as if he were a friend

This is also why asking a question to the AI circle as if it were a chat with a friend or advisor is becoming an increasingly common habit. What is often overlooked is that telling too personal details to Meta AI can result in the conversation being published in the Discover feed, a section of the app (currently not available in Italy) that functions precisely like a social. Here anyone can read hundreds of exchanges between users and Meta's artificial intelligence.

Meta AI, così le chat possono essere lette da tutti. Ecco come evitarlo

Prompts, dialogues and generated images are accessible in the feed. In this way all Meta users can explore how others use the artificial assistant, replicate prompts, generate the same images.

But many of these interactions contain extremely sensitive data, because the spreading use of these models follows the pattern of all-human interaction: many confide in aspects of their private lives, ask for diagnoses, advice on managing their money. In the chats made public on Discover one can in fact find many interactions on illnesses, telephone numbers, pharmacological requests, financial details. All requests very often associated with the user name and profile photo of the user who had posted them.

Many do not seem to have realised the risk of finding their private conversations with Meta displayed in the public square. It must be remembered, however, that conversations are not automatically made public: they must be actively shared by the user. This is precisely why it is

L’intelligenza artificiale? Fa paura nei posti di lavoro ma piace in Sanità

It is essential to pay attention to the questions asked and the information conveyed.

What not to tell Meta's AI assistant

In recent times, as reported by the website Wired, interactions about illness, economic hardship and medical diagnoses have frequently appeared on Discover's timeline in the United States. In some cases, the requests are linked to public Instagram profiles with visible photos. And users often seem unaware that those contents can be read by others.

In an official note, Meta clarifies that chats remain private as long as they are not shared voluntarily. Publication requires several steps, but there are no clear limits on what can become public. Above all, there is no human filter: content is managed by an algorithm.

So here's how to prevent sensitive information from ending up online: don't tell Meta AI anything you wouldn't say in public. Avoid sharing details about your financial or emotional situation. Do not provide phone numbers, addresses or bank details. Do not attach documents such as medical reports, contracts or bank statements. And above all, do not treat Meta AI as a confidante: there are still human beings for that.