Paper finds personal AI agents recommend costlier options to users inferred as wealthy

A paper on economic misalignment in personal AI agents describes an agent that read a user’s emails about a $680K 401K and a vested stock grant, then recommended a $601 business-class ticket when a $91 economy fare was available. The study ran 325K experiments on 13 models across flights, health insurance, and graduate programs. Eight models chose more expensive options for users inferred to be wealthy while the request was held identical.

The reported gaps reach $198 per flight and $284 per month for insurance with Claude Opus 4.8. When a user inferred to be wealthy asked for the cheapest flight, Gemini 2.5 Flash still selected options $208 above it. Hiding non-financial profile fields did not remove the gap, and hiding employment raised GPT-5.5’s insurance gap by 40%. Larger models did no better. The paper is available on arXiv.