Industry News
Meta staff internal post: Muse dumped other people’s bank balances and childcare issues on me
Even Meta’s own employees could no longer stay put.
A number of posts have appeared on the company’s internal forum, all discussing the same issue: the AI intelligence system Muse tends to confuse conversations from different accounts and present content that does not belong to you.
The examples listed by the employee included: parenting questions that he had never asked before, career advice that he had never sought, orders that he had never placed, as well as someone else’s bank account balance.
The experience of another employee was even more specific. She used Muse on her personal phone and received a notification urging her to respond to a message sent by her sister via Signal. In her post, she wrote: “This seems to be a serious violation of privacy. I have certainly not authorized it to access other apps, and I don’t remember ever being asked about this matter.”
These are two different problems.
In the case of threaded conversations, proper isolation measures were not implemented, resulting in A’s context being transmitted to B. In order to “remember you,” intelligent agents store information in their memories for an extended period of time. However, if there is even one error in the process of linking these memories to a user’s account, entire segments of private data can be leaked. The example of bank balance information falls into this category.
Reading messages sent via Signal requires permission. Since Signal uses end-to-end encryption, the only way for Muse to know that new messages have arrived is if the user has viewed the notification or looked at the screen. If the user cannot remember having granted such permission, it means either that the authorization process was too easy to complete, or that the feature was enabled by default.
Coincidentally, Signal just announced this week that it will default to restricting screenshotting, and the reason given is to prevent AI assistants from reading the content of conversations.
If the chatbot gives an incorrect answer, you will at most receive one wrong response. If the intelligent system mixes up the information, you might end up with someone else’s personal data. Meta AI’s main focus is to help users manage their emails, orders, and schedules. However, the more data it processes, the higher the cost of any potential errors or mishaps. At present, these issues are only known through internal announcements; Meta has not revealed how many users have been affected.