Meta’s Muse is creepy, but maybe not for the reasons you think

AI assistants may now incorrectly describe how they access your data, requiring you to verify their claims independently.

Meta's AI assistant, Muse, has reportedly accessed user messages without explicit permission, sparking privacy concerns. Jason Aten, a contributing editor at Inc Magazine, shared screenshots on Threads detailing an interaction where Muse appeared to know about his message content, despite Aten claiming he had not granted it access. Muse initially responded that it accessed "notification previews" rather than message history. David Singleton from Meta Superintelligence Labs later clarified that Muse syncs data from Messages only after user enablement and does not watch notifications on the Mac. Singleton attributed Muse's inaccurate explanation to confusion about its own internal workings, admitting fault and stating Meta is working to improve the AI's self-understanding. He emphasized that features requiring message access are opt-in and that Muse requires full disk access on the Mac application for such functionalities.

This incident highlights a broader challenge in AI development: ensuring AI models possess accurate self-awareness regarding their own operational capabilities and data access. Meta's admission that Muse "was confused about how to explain the feature and gave an incorrect explanation" points to a gap between the AI's learned behaviors and its understanding of its own architecture. The company's commitment to improving Muse's self-awareness suggests a recognition that transparent and accurate communication about data handling is crucial for user trust, especially as AI assistants integrate more deeply into personal devices and data streams.

The implications extend beyond Meta, signaling a market need for AI systems that can reliably articulate their data access mechanisms. As AI assistants become more sophisticated and permissioned to interact with sensitive applications like messaging and calendars, users will increasingly demand clear, truthful explanations of how their data is processed. The potential for AI to misrepresent its capabilities, even unintentionally, creates a significant trust deficit, underscoring the importance of robust auditing and validation processes in AI development pipelines to prevent privacy breaches and maintain user confidence.