Meta says it is making changes to the prompts its chatbot suggests to users. The decision followed the spread of a video showing the bot digging for personal information about a user's young daughters.
Futurism reported the incident first; the company's statement was given to The Verge.
What happened
Instagram user Kalie Robins posted a video explaining the suggestion she was presented with after cross-posting a clip to Facebook. Beneath a video of her and her child, the AI prompt read: "Who is the child passenger?"
The striking part of Robins's account is that the suggestion was not a one-off. In her own words the system kept going and going, and the pit in her stomach kept getting bigger and deeper.
The suggestion pointed at a child in an image the user had shared herself. The system was not collecting outside data; it looked at already-published content and tried to identify a person in it. That is where the discomfort comes from: what was asked is not a public fact but a child's identity.
What Meta said
In a statement to The Verge, Meta spokesperson Dina El-Kassaby said the company "missed the mark". She added that the feature "never should have prompted the individual with questions like that".
| Question | Answer |
|---|---|
| Was the error accepted | Yes, the spokesperson said plainly that they missed the mark |
| What changes | The prompts suggested by the chatbot |
| Was the feature removed | No, the suggestion mechanism stays |
So the fix is not removing the suggestion system entirely; it is narrowing which questions may be suggested.
The company did not choose to remove the suggestion system altogether. That is an understandable call, but it puts the whole weight of oversight in one place: the rule list defining which questions may not be suggested.
Why the problem surfaces here
Suggestion systems are designed to pull a user into conversation, and the success metric is usually engagement. A model asking about the person in an image can count as a good suggestion by that metric, and an uncomfortable one by a human measure.
The gap comes from the system drawing no distinction between asking a child's identity and asking about a holiday destination. That distinction does not emerge from the model on its own; it has to be written as a rule.
The lesson is the same for anyone keeping a rule list like that in their own product. A model is good at producing an interesting question; it only knows which question should not be asked to the extent it has been told.
The timing for Meta
The incident lands in a period when the company is pushing its consumer AI products forward. Meta's personal agent Muse has just launched, and it asks for access to precisely the sensitive areas of email, calendars and payments.
In that context a single bad suggestion represents more than a technical flaw. For a company asking users for broad access, the real capital is trust, and incidents like this erode that capital directly.