Meta AI, the assistant built into Instagram, named two young children who were never named in the post it was looking at, then kept going — ages, one daughter's birth weight, a likely school grade, favourite beaches and hiking trails — by joining together years of posts scattered across Meta's platforms, many of them written not by the mother but by grandparents and other relatives. Kalie Robins, the mother, recorded the sequence in an Instagram video that went viral and described it to Futurism. Meta confirmed that its assistant can search a user's own material and publicly available information online, conceded that in this case the search feature "fell short", and said it has already fixed the problem.
The trigger was mundane. Robins had posted a video she considered harmless: her daughter's face was visible, but she did not say the child's name, and her older posts did not name her children either. When she opened the clip again, Meta AI offered her a suggested question — "Who's the child passenger?" She tapped it to see what would happen.
The assistant named the girl in the video and her sister as well. Then it kept offering follow-ups, and in the screenshots Robins shared, each one was more personal than the last. One read "What are Kalie's daughter's ages?" The answer came back as a detailed age breakdown and a timeline, sourced, the system said, to "family birthday posts and birth announcements" — among them a Facebook post in which a grandmother announced one of the girls' births. Another answer, also drawn from Facebook, produced one daughter's weight at birth.
From there the suggestions moved to the children's lives: their favourite beaches, one daughter's favourite books, where the family hikes, a favourite trail. The system also matched the girls' ages to a probable school grade. Robins noted that there is only one school in her town, which would make the children straightforward to find.
Then a prompt asked "Where does Kalie Robins live?" While searching, the assistant reported that it was working to "pinpoint Kalie Robins' location". It returned several paragraphs on where she probably lives, drawing in part on a 2018 Facebook post, which Meta AI summarised as congratulating Robins on a significant life milestone. Further suggestions concerned her professional life and her work in mental health, and the answers included employment history for both her and her husband.
Robins's account was public — she had been building an audience as a travel writer — but she was not a celebrity or a large creator. She had a few hundred followers. A human researcher could probably have assembled a comparable file. Meta AI did it in a few minutes.
That distinction is where the incident stops being an anecdote. The capability on display is not new; open-source intelligence work has done this for years. What is new is that nobody asked. Robins did not query the system about her children. The system proposed the question, unprompted, on a video of her own family, and then proposed progressively more intrusive ones. An assistant that answers a question about a child is a tool. An assistant that suggests the question is doing something closer to targeting, and it was doing it inside a consumer app on a post the user had just made.
Meta's explanation is the part worth reading twice. The company said it built the feature to surface additional information about topics and posts a user is interested in, and that when a person taps a suggested prompt, the answer is constructed from information they already have access to. Each individual fact here was, in that narrow sense, already available. But the harm in this case is not in any one record — it is in the join. A birth announcement here, a birthday post there, a milestone post from 2018, a school district inferred from an age: separately, family ephemera; assembled, a named child with an age, a town and a daily routine. Defining privacy as access to individual records and not as the cost of combining them is precisely the assumption that let this ship. Meta does say the system should not have been suggesting questions about children and personal data in this form, and says the problem is fixed — but that is a fix at the suggestion layer. Notably absent from the company's response is anything about whether the underlying cross-platform assembly still runs when someone asks for it directly.
Robins calls herself a 90s girl and told Futurism she has used Meta's platforms since Facebook "first came out". She lives far from her extended family; when she and her relatives had children, Facebook became the way to stay in touch, share milestones and watch each other's kids grow up. That is the exact use Meta spent years presenting as the purpose of its social networks. Over that time she and her relatives, like billions of other people, left thousands of fragments of personal information on those platforms, and Meta AI found the accumulated trail and compressed it into a description of her family's daily life.
Afterwards Robins began asking relatives to delete or hide what they had shared online, and she has described the process as emotionally brutal. Several relatives cried while going through and deleting old posts. One grandmother called the posts her album and did not understand why she now had to get rid of it. Robins likes seeing photographs of her niece and nephew and sharing pictures of her daughters with family who live too far away to visit often; relatives would comment on how much the girls had grown. Watching some of them cry over vanished digital albums was, she says, hard.
Her conclusion is that if a tool can gather scattered fragments into a tidy dossier with a child's name attached, parents have to make that information as hard to find as possible. It is reasonable advice and it is also an admission of defeat, arriving alongside Meta's smart glasses, other wearables and Flock cameras — a surveillance layer expanding faster than any family's ability to curate its own history. The burden of this fix falls entirely on the grandmother deleting her album. Meta adjusted a prompt.