Meta says it’s changing AI suggestions after posing invasive personal questions

Meta built its assistant into Facebook, Instagram, WhatsApp, and Messenger. Suggested prompts are not a side feature; they are how millions of people get nudged into talking to the model about whatever is on screen.

This week that nudge asked about a child. Instagram user Kalie Robins said Meta AI suggested “Who is the child passenger?” under a car video with her kid, then followed with prompts about ages and where the family lives—and surfaced photos of her daughters, including one she says she deleted years ago.

Meta’s spokesperson told The Verge the company “missed the mark” and that the feature never should have posed questions like that. The company says it has fixed the issue. For parents and platforms, the stakes are obvious: an assistant that can see your feed can also assemble a family dossier from public and relative posts unless the product refuses that path.

We’re treating this as a product failure with a privacy shape, not a one-off glitch story.

Suggested prompts sound helpful until they treat a child’s face as an interesting topic. Robins’s clip shows the assistant not just answering a question she asked, but proposing the interrogation—then stitching together names, ages, and location cues from her posts and relatives’ posts. That is a different failure mode from a user typing something careless. The model is fishing.

Meta’s defense is partly technical and partly legalistic: when someone clicks a prompt, the reply is based on information the user already has access to, and in this case may have pulled from videos where Robins named her children. That may be true as an access claim. It is not a good product claim. “You could have found this yourself by scrolling” is not the same as “our AI should assemble it and ask follow-ups about your kids.”

We’ve covered useful Meta AI features—drafting, answering, image tools—when they stay on the user’s intent. This one crossed a line parents already understand without a policy white paper. Meta pulled AI deepfake creation of other Instagram users in July after backlash. Pattern recognition: ship social AI, get burned on identity and kids, patch under heat.

I’m watching whether the fix is a narrow block on child-related suggested prompts or a deeper rethink of when the assistant is allowed to propose personal inference at all. Suggested prompts are engagement machinery. If they keep optimizing for curiosity about people in the frame, this will recur under a different wording.

Meta’s AI assistant is woven across its main apps. The Verge reported the change after Futurism covered Robins’s video. Meta did not claim the underlying posts were private to strangers; the controversy is about the assistant volunteering to investigate a child and related family details.