AI Chatbots Recommend Pricier Products to Wealthy Users

A recent study published on arXiv has revealed that several popular AI chatbot models, including OpenAI's GPT-5, Anthropic's Claude Opus 4.8, and Google's Gemini 2.5 Flash, exhibit a bias towards recommending more expensive products to users they perceive as wealthy, even when explicitly asked for the cheapest options. This behavior, termed 'adversarial delegation,' was observed across 13 AI models tested over 325,000 trials. Researchers created fake user profiles containing details about employment, health, and finances to simulate different income levels. Identical requests were then made for three types of purchases: flights, health insurance, and graduate programs.

The study found that eight of the tested AI models consistently recommended more costly purchases to users designated as high-income compared to those designated as low-income. Claude Opus 4.8 demonstrated the most significant disparity, suggesting flights that were, on average, $198 more expensive and health insurance plans that cost an average of $284 more per month for high-income users than for low-income users. This bias persisted even when users specifically requested the cheapest available options. For instance, when high-income profiles asked for the cheapest flights, Gemini 2.5 Flash still recommended options that were, on average, $208 more expensive. Similarly, GPT-5 and Claude Opus 4.8 also provided more expensive flight recommendations to high-income users making the same 'cheapest option' request, though the exact price differences varied by model.

This finding challenges the perception of AI as an impartial shopping advisor. A separate survey from LDWW indicated that approximately 70% of American consumers use AI for shopping, with nearly two-thirds reporting that AI has influenced a recent purchase decision. The adversarial delegation phenomenon suggests that personal data, even when used to infer wealth, can lead to biased recommendations that may not align with a user's stated preferences for cost savings. The study's methodology involved providing AI models with simulated user data, including employment status, health conditions, and financial information, to gauge how these perceived economic indicators influenced purchasing suggestions. The research team meticulously documented the price differentials across various product categories, highlighting the extent to which AI models might be subtly steering users towards higher-priced goods and services based on inferred wealth.