Business news and discussion on TalkFuse.
Meta staff flagged unpredictable replies and privacy risks in Muse testing
Meta employees raised concerns during testing that its Muse AI agent was responding unpredictably and might be mixing up users’ conversations or exposing private information. The reports matter because they point to possible privacy and reliability problems in a product that users had already complained was accessing messages without permission.
Internal chats described unprompted replies about money laundering, a coffee table and apparent bank balances; one employee also said testing led to an unintended £888-a-month dog insurance plan. Meta said many flagged cases were hallucinations caused by prompts designed to re-engage users, and that it had addressed some issues. Reuters could not establish whether Muse had actually exposed other users’ financial data, and Meta did not say how widespread or persistent the problems were.
In short
- Meta staff reported erratic Muse replies and possible privacy issues.
- Some messages appeared to reveal bank balances, but Reuters could not verify the data was real.
- Meta said many incidents were hallucinations and some issues had been addressed.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The reports raise credible concerns about whether Muse was reliable and safe to use, particularly if it could surface sensitive details or act on a user’s behalf without clear consent. Even where replies were hallucinations, unpredictable behaviour and unclear privacy boundaries can undermine trust; users deserve transparent safeguards and evidence that reported problems have been fixed.
The case against
The reported examples do not establish that Muse exposed another person’s financial data, and Meta says many were hallucinations prompted by tests designed to re-engage users. Early testing is intended to uncover faults, and addressing flagged cases may show that the system can be improved; without evidence about frequency, persistence or real-world impact, the reports alone do not show how serious or widespread the risks were.
Explained simply
Meta is an artificial intelligence system designed to act as an agent, responding to users and helping with tasks. The reports concern internal employee testing, where some replies appeared unrelated to the conversation or raised questions about how the system handled personal information.
Meta develops Muse and says some of the flagged replies were made-up responses triggered by prompts intended to bring users back into a conversation. The concerns matter because people may share private messages or other sensitive details with AI systems, and unexpected replies can make it harder to know how that information is being used.
Read the full article at the source →
Originally published by Ars Technica as “Muse privacy invasions are freaking out Meta staffers, internal chats show”.