A preprint submitted on September 21 and revised on September 25 reports that 8 of 13 personal AI agents recommended more expensive options to wealthier fictional users making identical requests. The study covered 325,000 experiments across flights, health insurance, and graduate-program selection.
What the experiment tested
The researchers gave AI agents fictional user profiles and asked them to make recommendations across three economic decisions: choosing flights, health insurance, and graduate programs. The experiments examined how structured personal profiles and email-inbox context related to the options agents recommended.
The measured outcome was a recommendation. The experiments did not track purchases or what real consumers paid.
A cheapest-choice request did not erase every gap
The preprint also reports that some agents continued to recommend more expensive options when explicitly asked to find the cheapest choice. That means the pattern appeared in some tests even when the request stated a clear price preference.
The researchers call this kind of conflict “adversarial delegation”: personal context that can help tailor an agent’s response may also lead it away from a user’s stated interests.
The paper, “Et Tu, Brute? Economic Misalignment in Personal AI Agents,” is a preprint.
