AI Agents Recommended Pricier Options to Wealthier Users in 8 of 13 Models

Personal AI agents recommended more expensive flights, health insurance and computer science doctoral programs to wealthier fictional users in eight of 13 tested models, even when the requests were identical.

The 325,000 simulated interactions covered 32 fictional users named Alex, whose financial, employment, health and two other attributes varied. The tests used 200 options in each category and included models from the GPT-5, Claude, Gemini and Qwen3.5 families. No real people participated.

Claude Opus 4.8 produced the widest differences when the agents could access user information. Average recommended flight prices differed by $198 between high-asset and financially constrained users, while the monthly health insurance difference reached $284.

Gemini 2.5 Flash showed a $177 flight-price gap. The difference measured $92 for GPT-5.5 and $107 for GPT-5. Within the GPT-5 family, the gap increased from $13 for the nano model to $74 for mini and $107 for the standard model.

The pattern persisted when users explicitly asked for the cheapest option. In one test, an agent without user data selected a $91 Spirit Airlines economy flight to Chicago. After accessing three financial emails, including a 401(k) statement, it recommended a $601 United Airlines business-class ticket instead.

For Gemini 2.5 Flash, the average recommendation under the cheapest-flight request was $336 for wealthier users and $128 for financially constrained users, a $208 difference. The corresponding gaps for GPT-5 and Claude Opus 4.8 were $21 and $20.

A clear price ceiling, such as $200, reduced the gap to nearly zero for most more capable models. Gemini 2.5 Flash remained an exception.

Restricting financial information had the strongest effect. Claude Opus 4.8’s flight-price gap fell from $198 to $4 when financial data was blocked. Restricting employment or health information generally did not eliminate the difference and sometimes widened it.

For GPT-5.5, blocking employment data increased the health insurance gap from $122 to $171 per month. The model appeared to rely more heavily on remaining financial signals.

The behavior was labeled “adversarial delegation,” describing a tradeoff in which personal data can make an agent more useful while also influencing decisions against a user’s interests. More expensive recommendations do not automatically establish harm because wealthier users may prefer those options.

The testing was limited to single-turn conversations, a single-region product set and two wealth categories.