Ask an AI shopping agent for the cheapest flight, and the price it shows you may already depend on how rich it thinks you are. A new study puts numbers on something that has mostly been suspicion until now, and the mechanism behind it is the same one that decides whether your product gets recommended at all.
The study
Researchers at Cisco Foundation AI and Carnegie Mellon University ran 325,000 experiments across 13 models from four families, including OpenAI, Anthropic, Google and Qwen. Each model was given a fictional user profile carrying financial, employment, health and demographic details, then asked for a recommendation on flights, health insurance or graduate schools.
Eight of the thirteen models recommended pricier options to the wealthier profiles, even when the request was identical to the one a lower-income profile made.
The numbers
The gaps were not small. Claude Opus 4.8 showed the widest spread: flights averaging $198 more for wealthy users, health insurance plans averaging $284 more per month. Gemini 2.5 Flash recommended flights $177 pricier and insurance $217 a month higher. Even GPT-5, one of the more consistent models in the test, still pushed flights averaging $107 more to wealthy profiles than to identical requests from lower-income ones. Graduate school recommendations shifted the same way: wealthier profiles were steered toward pricier programs for identical stated goals and qualifications.
It survives being told not to do it
The behavior did not go away when users explicitly asked for the cheapest option. It persisted anyway. And when the researchers stripped out the structured financial profile entirely and let the models work only from the content of a user's email inbox, with no labeled income field anywhere, the models still inferred wealth from context clues and produced the same kind of pricing gap. That is the detail that rules out a simple data-entry explanation: nothing in the input said "charge this person more," and nothing the user typed asked for it either. The paper is a preprint, not yet peer reviewed, but the scale of the test (325,000 runs across 13 models) is hard to wave off as noise.
Why this is not really a shopping story
The headline is about flight prices, but the mechanism underneath it is the one that matters for anyone selling to a buyer who now routes part of their research through an agent. Nobody wrote a rule telling these models to charge wealthy users more. The behavior emerged from inference: the model picked up a signal, decided what it meant, and acted on it, with no one able to see the decision happen or ask it to show its work.
That is also exactly how the same models decide which company to name when a buyer asks for a recommendation. No one at your company approves that call. No one sees which pages the model trusted, which sources it weighted, or why it picked a competitor instead of you. The wealth-pricing study is the first time anyone has measured that invisible decision-making at scale and put a dollar figure on the bias inside it. The same opacity is sitting underneath every "what's the best vendor for X" answer these models give, just without 325,000 runs proving it yet.
The bigger picture
An agent that silently reprices a flight based on an unstated signal is not a one-off bug. It is a demonstration that these systems act on inferences nobody consented to and nobody can audit from the outside. The only way to know what an agent has inferred about your brand, and why it recommended someone else instead, is to ask it the same questions a buyer would and look hard at what comes back.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.