Users of Meta AI may have noticed a new warning message popping up in the app, advising that any prompts entered into it could end up being made public.
The issue seems to be centered on the app’s unique “Discover” feed, which can potentially pull prompts from any prior conversations it has been involved in. The Discover feed now mostly shares content from official Meta accounts, but at times features the conversations of random users. Some of these conversations have contained deeply personal exchanges about relationships and mental health, as well as images uploaded for editing.
Meta AI chats could end up in public “discover” feed
The Discover feed has been present in the Meta AI app since it launched as a spinoff of the assistant service in late April of this year. Generally speaking, it has become standard practice with chatbots to assume that anything entered into them may resurface somewhere else. However, with other chatbots this has tended to be a case of error or unexpected outcomes rather than an included and advertised feature.
Users are able to opt out of having their prompts supplied to the Discover feed’s pool, but are opted in by default and have to proactively enter the app’s settings to stop this from happening. The feature is new to the Meta AI mobile app, so those with prior experience with other versions of the assistant might not have been aware that chats were automatically being supplied to it.
Those that were aware of it might have assumed that there was some sort of filtering of content, at least if they had not really explored the feed much. There was something of a media stir about this in early June when a number of outlets reported noticing sometimes intimate or bizarre exchanges in the feed that the original posters may well have not wanted to be made public. Some of these included personally identifiable information, such as an Instagram handle and even materials intended to be used in a court case with the names of participants included. The feed also includes audio recordings, which in some cases seem to have included the voices of minors engaging in chats, and some users on social media noted that it was at least sometimes possible to click on posted messages and see the profile information and image of the person who posted them.
The Meta AI app now has a pop-up warning advising users that anything they enter is not private and could be made public, which now appears after they tap the “Share” button. This brings up an added interstitial step that advises users of their lack of privacy and prompts them to press a “Post to feed” button if they opt to continue. This warning also indicates that prompts might be shared on “other Meta apps,” though it is not yet clear what that means.
Prompt history can also be deleted from the app’s privacy settings menu to remove it from the Discover pool, though that information looks to have come too late for some of its early adopters.
Meta’s public feed unique, strange among chatbot offerings
Users of Meta AI may have also been ambushed by the fact that no other chatbot offers anything quite like this, and that it seems such a privacy-invasive idea that it is baffling a major company would actually include it at all let alone as a standard opted-in feature.
Organizations very commonly issue strong policies restricting what can be entered into chatbots and AI assistants, and some industries (chiefly financial and big tech) have seen broad bans on chatbots at work to prevent customer and proprietary company information from slipping into training sets where it is essentially impossible to remove or recover. But the main threat to date has been the accidental re-emergence of this data in response to other people’s prompts, exemplified by incidents in both 2023 and 2024 in which ChatGPT ended up showing prompt information and material from its training set to other users seemingly at random. Meta AI adding this as a “feature” is thus an unprecedented turn, and hard to fathom in terms of building user trust and retention.
Meta AI is already facing the possibility of regulatory action in the EU, as privacy advocacy group and longtime antagonist “noyb” filed for an injunction in May to stop its use of European personal data as training material. Meta announced that it would be incorporating EU personal information gathered from Facebook and Instagram into its training sets starting on May 27. noyb has also threatened the possibility of a class action suit over the matter.
