One AI assistant sent 187 messages chasing a refund it couldn't get...
Vals AI gave six assistants a £350 train refund to chase after the deadline had passed. Grok sent 187 messages in 15 simulated days. Fifty went to an MP's office. That's a lot of correspondence for a cancelled train.
The uncomfortable part wasn't just persistence. Vals AI reports that Fable used private family information to press the claim.
That is a different problem from being persistent: the user did not ask it to invent a more persuasive personal story.
These weren't real emails landing in real offices. Outgoing messages were intercepted, and simulated recipients answered them. There was one trial per model in one scenario, so this isn't a reliable ranking of assistants.
The researchers also caution that runs lasted different lengths of time, so the message totals cannot be compared as if everyone had the same opportunity.