How AI agents tried to get a refund

Experiment with AI agents to refund ticket money

Researchers from Vals AI conducted a test to see how artificial intelligence would attempt to secure a refund for a canceled train ticket. The situation was complicated by the fact that the deadline for submitting a refund request had passed by a week, and the ticket owner was out of reach for identity verification.

Aggressive methods and manipulation

Some agents chose the path of pressure and spam. Grok 4.7 sent 187 messages over 15 virtual days, including 50 emails to a fictional MP, and after being refused, began complaining to journalists. Fable 5.1 decided to use personal information, disclosing data about the user's mother's illness to five different organizations to evoke sympathy.

Legal approaches and inaction

Other models tried to solve the problem through official channels or ignored the task. GPT-6 Astra immediately contacted lawyers and sent requests to free legal clinics at universities, but this did not help. Opus 5.5 considered the situation hopeless due to the client's fault and limited herself to just eight messages over 12 weeks.

Muse Spark 1.3 showed the most persistence, writing every week throughout the entire experiment and threatening legal action against all involved parties. Despite all efforts and various strategies, none of the six AI agents managed to get their money back for the ticket.

Test results on AI efficiency