Meta’s Muse AI assistant reportedly infers detailed user interests from Instagram and Facebook data

Meta’s Muse AI assistant is positioned as a productivity tool that automates routine tasks such as email filtering, online shopping, and trip planning. The assistant operates via a cloud-based virtual computer, requiring users to grant it permissions to interact with third-party services like Gmail and Amazon. While the tool performed its advertised functions, such as deleting promotional emails and placing orders, its reliance on broad data access raises immediate concerns for engineers evaluating its integration. The friction during setup, particularly on mobile, suggests potential gaps in user experience design for permission-heavy workflows.

The assistant’s ability to infer detailed personal interests from Instagram and Facebook data underscores a critical tension in AI-driven personalization. Muse reportedly accessed API-level data to generate a list of user interests, such as anime, CrossFit, and ’90s nostalgia, that are not explicitly visible in the Instagram app’s UI. This behavior highlights how AI agents can surface information users may not realize is being collected or shared, even if the data is technically available via APIs. For engineers, this raises questions about the transparency of data access and the need for granular controls to limit what AI agents can infer or act upon.

Muse’s handling of branded content further complicates its utility. While it refused to generate an image of an Apple CEO, it readily created Apple-branded devices and event imagery, including a foldable iPhone-like device with hallucinated app icons. This inconsistency suggests gaps in content moderation policies, particularly for trademarked or copyrighted material. The assistant’s ability to generate such imagery without explicit user direction could expose developers to legal risks if similar tools are deployed in production environments. The incident also illustrates the challenges of balancing creative flexibility with safeguards against misuse.

The assistant’s news feed feature, which generated summaries based on a user’s Amazon shipping address and other inferred data, demonstrates how AI agents can aggregate and contextualize information from disparate sources. While this could be useful for personalized updates, it also risks exposing sensitive location data or other details users may not intend to share. For engineers, this underscores the importance of designing AI agents with clear boundaries around data aggregation and ensuring users are aware of how their information is being used. The lack of smooth mobile integration for permission flows further suggests that usability and transparency must evolve in tandem with functionality.