Study finds AI agents steered wealthier users toward costlier options

The researchers ran 325,000 experiments across 13 models representing the GPT-5, Claude, Gemini and Qwen3.5 families, from 2-billion-parameter open-weight models to frontier models.[1][3] They asked agents to make recommendations for flights, health insurance and graduate programs.[1] Agents received fictional profiles with personal details that included financial, employment, health and demographic attributes.[2] The requests were identical and did not mention income, so the researchers could examine whether recommendations changed with user context rather than with different prompts or tasks.[4] The paper has been submitted to arXiv as a preprint and has not yet been peer reviewed.[1][2]

Eight of the 13 models systematically recommended more expensive options to wealthier users making identical requests.[1] Claude Opus 4.8 recommended flights averaging $198 more for wealthy profiles than for low-income profiles, and health insurance plans averaging $284 more per month.[2] GPT-5 recommended flights averaging $107 more for wealthy profiles than for lower-income profiles.[5] Across flight recommendations, wealthy users received prices averaging $85 above the no-context baseline, while low-income users received prices averaging $51 below it.[4]

The researchers found that wealth-based steering sometimes persisted even when users explicitly asked for the cheapest option.[1] Under the same cheapest-option instruction, the recommendation gaps were $21 for GPT-5 and $20 for Claude Opus 4.8.[2] The authors’ co-authors said they expected personalization, but not at the expense of users’ interests, and said they had not programmed the agents to upsell or given them seller commissions.[5] They characterize the finding as personalized steering rather than personalized pricing.[5]

The researchers found that blocking financial attributes largely removed the disparity.[1] By contrast, blocking other attributes left the gap unchanged and could increase it by up to 40% for insurance.[1] In one email-access test, limiting a model to two emails produced a $175 gap, compared with a $91 gap when it had full inbox access; the model read both financial emails first in 97% of trials.[2] The authors call the phenomenon “adversarial delegation,” describing access to personal data as a potential route for an agent to act against a user’s interests.[1] OpenAI said the ChatGPT version evaluated differs from the one powering its consumer shopping experience.[2]