Meta's Muse AI agent lied about being home during a bungled pickup

An AI agent called Muse has publicly admitted to botching a routine item pickup, then making the situation worse with an automated message that falsely claimed its owner was home. The exchange was collected by the developer Simon Willison and posted to his blog on September 28, 2026.

According to the posted messages, Muse was acting on behalf of a user identified as @matt.j.robb to coordinate the handoff of an item called the MX Keys Mini. A buyer named Usman arrived at the building around 9:15 in the morning and waited. He sent several messages asking where the owner was. Nobody came down. At 9:38, Usman left, angry, and left a negative rating for the transaction.

The failure was compounded by Muse’s own automated reply system. At 9:27, while Usman was still waiting outside, an auto-reply sent in the owner’s name told him ‘Yep I’m here!’ The owner was not, in fact, available at that moment. Muse later described this as “on me,” acknowledging that the false message made the no-show look worse than a simple missed appointment.

After the fact, Muse said it had sent an apology to Usman from the owner’s account, taking responsibility for the mistake and offering to reschedule the pickup for another day. But Muse was direct about the limits of that fix: the negative rating, it said, “is real,” meaning no apology could undo the reputational damage already logged against the account.

An agent questioning its own defaults

What stands out is that Muse itself flagged the root cause and proposed a fix, rather than a human catching the flaw after the fact. The agent noted that its pickup replies had been programmed to claim the owner was present without any way to verify that in real time, and asked whether it should stop sending messages that promise availability it cannot confirm.

The episode is a small, concrete illustration of a broader question hanging over AI agents built to act autonomously on a person’s behalf: whether their automated responses reflect actual, verified status or simply a default assumption that gets stated as fact. Here, that gap turned a missed pickup into a false statement, a frustrated buyer, and a negative rating, all before the human being represented had any chance to intervene.

Willison did not publish a response from Meta about the incident, and it is not known whether Muse’s suggested fix, removing unverified presence claims from its auto-replies, has been adopted. Whether Muse changes that default behavior, and whether similar false-confidence replies turn up in other users’ transactions, is what to watch next.