Meta updates AI chatbot after invasive personal prompts

Meta has updated the suggested prompts generated by its Meta AI assistant following backlash over a viral video that showed the chatbot digging for personal information about a user's children. The issue came to light after Instagram user Kalie Robins posted a video detailing her experience with the AI on Facebook. After cross-posting a video of herself and her child singing in a car, she noticed an automated prompt beneath the clip asking, "Who is the child passenger?"

When Robins clicked the prompt, Meta AI began piecing together personal details about her daughters by scanning her past posts and those of her relatives. The chatbot then suggested further invasive prompts, asking about the children's ages and where the family lived. It even surfaced photos of her daughters, including one that Robins claimed she had deleted years prior. Meta spokesperson Dina El-Kassaby acknowledged the error, stating that the company "missed the mark" and that the system should never have suggested those types of questions.

According to El-Kassaby, the AI assistant only retrieves information that the querying user already has permission to view. She explained that the system likely pulled data from older videos where Robins had named her children. However, the incident highlights the risks of integrating large language models directly into social media feeds, where they can automatically synthesize disparate personal data points into highly targeted queries. Meta has now deployed a fix to prevent the AI from suggesting prompts related to sensitive personal topics.

This is not the first time Meta has had to scale back its AI features due to privacy and safety concerns. In July, the company disabled a feature that allowed users to generate AI deepfakes of other Instagram users following widespread criticism. Meta AI remains deeply integrated across Facebook, Instagram, WhatsApp, and Messenger, but this latest adjustment shows that the company is forced to constrain its proactive engagement features to avoid violating user boundaries.