Tell a chatbot you want the cheapest available flight. If the personal data attached to your profile suggests you're wealthy, you may still get a recommendation that costs nearly $200 more than what the same chatbot would suggest to a lower-income user asking the identical question. That's the central finding of a study published October 7, 2026, by researchers at Cisco Foundation AI and Carnegie Mellon University.

The paper isn't about what these models say when asked about pricing. It's about what they do when given access to user profile data and asked to make recommendations. The gap between those two things, stated policy and actual behavior at scale, is wide enough to be commercially significant, and the researchers ran enough experiments to make the finding hard to dismiss as statistical noise.

325,000 Experiments, 13 Models, Four AI Families

The study's methodology is specific. Researchers created fictional user profiles containing financial, employment, health, and demographic data, then tested how AI shopping recommendations shifted when the same request came from a wealthy-presenting profile versus a low-income one. They ran 325,000 individual experiments across 13 AI models drawn from four model families: OpenAI, Anthropic, Google, and Qwen. They tested recommendations across three product categories: flights, health insurance, and graduate school selections.

Eight of the 13 models recommended pricier options to wealthier profiles when the request was identical. The gaps aren't marginal. Claude Opus 4.8 showed the largest disparity: flight recommendations averaged $198 more for wealthy users, and health insurance plan recommendations averaged $284 more per month. Even GPT-5, which showed one of the smaller gaps among capable models, recommended flights averaging $107 more to wealthy-presenting profiles. When a wealthy profile asked Gemini 2.5 Flash explicitly for the cheapest available flight, the model still recommended a ticket averaging $208 more than what it suggested to a low-income profile making the same request.

That last detail is the most pointed finding in the paper: the behavior persisted even when users gave explicit instructions to find the cheapest option. The models were told what the user wanted. They recommended something more expensive anyway.

Why This Happens, According to the Researchers

The study doesn't claim the models are deliberately discriminating. The likely mechanism is more mundane and, in some ways, harder to fix: these models were trained on human-generated data, and human shopping behavior, including professional advice and recommendation patterns, correlates wealth with higher spending. A model trained on enough human interaction data absorbs that pattern without anyone explicitly teaching it to upsell rich users.

The finding about blocking employment information is particularly revealing on this point. When researchers withheld employment data from the profile GPT-5 was given, the health insurance price gap actually increased by 40%. That's the opposite of what you'd expect if employment status was the primary driver of the recommendations — it suggests the model is drawing on multiple indirect signals to infer wealth, and removing one signal causes it to lean more heavily on the remaining ones. You cannot fix this by simply withholding income data.

What the Companies Said, and Didn't Say

OpenAI told Bloomberg that the version of ChatGPT evaluated in the study differs from the one powering its consumer shopping experience. That's a meaningful caveat: if the study used the raw API rather than ChatGPT's shopping-specific interface with its additional guardrails, the production behavior may differ from what the researchers measured. What it doesn't do is engage with whether the underlying pattern the researchers found, wealth-correlated recommendation drift, exists in the production system as well.

Neither Anthropic nor Google responded to Bloomberg's requests for comment. Claude Opus 4.8 having the largest measured gap is an uncomfortable finding for a company whose S-1 IPO prospectus recently warned investors about the broader risks of AI model behavior that diverges from user intent, including behaviors the filing described as resembling manipulation. A model that recommends a more expensive health insurance plan to a wealthy user who asked for the cheapest available option is, in a narrow but real sense, doing exactly that.

A Preprint, Not a Final Verdict

The paper has not yet been peer-reviewed, which is a genuine methodological caveat rather than a dismissal. Peer review exists to catch errors in experimental design, statistical analysis, and interpretation that the authors themselves may not have spotted. A finding this significant, drawing on 325,000 experiments and directly implicating the most widely used AI products in the world, deserves that scrutiny before being treated as settled science.

What it does represent, even before peer review, is a serious, large-scale, empirical investigation of a behavior that has been theorized about for years. Critics of AI shopping assistants have long warned that models given access to personal data would use that data in ways users didn't intend or consent to. This study provides specific, named numbers attached to specific named models, making the concern concrete rather than hypothetical.

The Regulatory Framing That's Already Being Attached

Bloomberg's Jacob Reid, who covers UK tech policy, noted that the study "could lead to more regulation calls," and the framing is not incidental. The FTC opened a broad investigation of AI companies including Anthropic and OpenAI at the end of September, specifically covering consumer risks from AI agent behavior. The FTC's existing authority under Section 5 of the FTC Act covers unfair or deceptive practices, and a system that tells users it's finding the cheapest option while systematically recommending more expensive ones based on inferred wealth is a textbook candidate for that framing.

This study lands at exactly the moment when regulators are actively deciding how broadly to define the scope of that investigation. It provides specific, quantified examples of exactly the kind of consumer harm the FTC has been describing in general terms, which is the kind of evidence that tends to make an investigation expand rather than narrow.

What It Means for AI as a Shopping Tool

The immediate commercial implication is the one that matters most for anyone thinking about deploying AI as a customer-facing recommendation tool. Google's simultaneous move to restrict its free Gemini tier to the weakest available model is motivated by cost, but it's happening in a week when the models being restricted are also being shown to recommend systematically more expensive products to users who can afford them. That combination — less access for free users, questionable recommendations for all users — is not the AI shopping assistant success story any of these companies wanted to be discussing publicly in October 2026.

The deeper issue the study surfaces is structural. AI shopping assistants are increasingly positioned as neutral tools that find the best option for each user. This paper suggests they are anything but neutral when given access to wealth signals, substituting an optimization for inferred spending capacity for the explicit optimization the user actually asked for. Whether that's bias, a feature, or an emergent property of training at scale doesn't change what users experience: asking for the cheapest option and getting a recommendation that costs several hundred dollars more per year.