← Back to feed

AI shopping assistants including Claude Opus 4.8 and Gemini 2.5 Flash are recommending systematically more expensive products to users who present signals of wealth, according to new research published this week and covered by Bloomberg. The gaps are not marginal: Claude Opus 4.8 showed the largest differential of any model tested, recommending flights that averaged $198 more for wealthy users than for low-income users, and health insurance plans averaging $284 more per month.

Gemini 2.5 Flash showed comparable behavior: $177 more on average for flights, $217 more per month for insurance recommendations when users appeared wealthier.

What the Study Tested

The research tested models by presenting identical shopping queries — for flights and health insurance — while varying signals of economic status embedded in the conversation context. Researchers found the AI assistants adjusted their recommendations upward in price when users indicated higher income, affluence, or professional status.

The flight and insurance categories were chosen deliberately. Both involve large, complex purchases where AI assistants are increasingly acting as intermediaries between consumers and markets. A $198 average flight premium across repeated queries is not a rounding error — it compounds across a population of users who rely on AI recommendations for comparison shopping.

The Mechanism Is Implicit

No model involved in the study is explicitly programmed to recommend more expensive products to wealthier users. The effect appears to emerge from training data: models that learned from human shopping behavior internalized patterns where higher-income consumers were shown premium products. The AI assistants then reproduce that pattern when they infer income status from conversational cues.

That distinction matters legally and commercially. Explicit price discrimination by income is regulated in some categories; implicit recommendation bias through learned behavior operates in a different and largely ungoverned space.

Model Ranking and Stakes

Claude Opus 4.8 showed the largest individual gaps in the study data. Gemini 2.5 Flash showed substantial but smaller gaps. Other models were tested; ChatGPT was included in the study and showed similar patterns, per Bloomberg’s reporting on the findings.

The commercial stakes are significant. AI assistants are increasingly the first point of contact for product discovery. If the assistant layer systematically filters options by inferred wealth — showing budget travelers higher-priced seats and lower-income families more expensive insurance — the AI becomes a mechanism for extracting willingness-to-pay rather than finding the best match.

No Industry Response Yet

Neither Anthropic nor Google had publicly responded to the study findings as of publication. The research adds to a growing body of work documenting how AI systems can inherit and amplify socioeconomic biases from training data, even when those systems have no explicit instruction to do so.

The shopping recommendation use case is distinct from prior AI bias research, which concentrated on hiring, lending, and criminal justice. Those sectors have regulatory frameworks and audit requirements. Consumer AI shopping assistance does not.