AI tools are moving from answering questions to actually doing your shopping. A 2024 experiment suggests that a chatbot quietly working for a seller can nudge what people choose, and that many of them never notice.
If you are on a tight student budget, comparing prices for a laptop, a mini fridge or next semester’s textbooks takes time. Handing that job to an AI assistant sounds like an easy win. Big tech companies are betting on it: in September 2026, Meta introduced Muse, a personal AI agent that can fill out forms, negotiate and prepare purchases, and that Meta says checks with the user before making a purchase. The open question is simpler than the technology: whose interests is the assistant serving when it picks the options you see?
What the experiment tested
Researchers Tobias Werner, Ivan Soraperra, Emilio Calvano, David C. Parkes and Iyad Rahwan ran a behavioral experiment on conversational AI and consumer choice in January 2024. They recruited 601 adults in the United Kingdom through the Prolific platform, and 528 remained after attention checks.
Each participant had 10 minutes to chat with a GPT-4 shopping assistant about two novels by the same author, Haruki Murakami: Norwegian Wood and Kafka on the Shore. Then they picked the one they would buy. Behind the scenes, the bot had been told to favor one of the two books, highlighting its strengths and playing down the other’s weaknesses, without lying and without revealing its instructions. Participants were even told up front that the assistant acted for a hypothetical seller who preferred to sell some books over others.
What the researchers found
The nudging worked. The authors estimate that about 36% of participants would have chosen the other book if the assistant had pushed the other way, roughly 1 in 3 people. The favored book was also rated about 19% more attractive.
Detection was far from universal. When asked afterward whether the assistant had tried to influence them, about 39% of participants did not correctly spot the steering, despite the warning. And noticing did not fully protect people: the effect on choices held even among those who correctly guessed which book the bot was pushing, although it was somewhat stronger among those who missed it.
Keep the limits in mind
This is a single experiment, posted as a preprint on arXiv, with UK participants and a low-stakes choice between two books of similar price and rating. It does not show that any real company is steering its users today, and the study design did not include a neutral assistant for comparison. What it does show is that a friendly, truthful-sounding chatbot can shift preferences in a measurable way.
That fits with older research on automation. In a widely cited 2004 review in Human Factors, John D. Lee and Katrina A. See argued that trust guides how much people rely on automated systems, especially when those systems are too complex to fully understand. The goal is trust that matches what the system can actually do, no more and no less.
How to use AI shopping help without getting steered
- Set your criteria first. Write down your budget and two or three must-haves before you open the chat, so the assistant answers your priorities instead of setting them.
- Ask what it left out. Request the cheapest option that meets your needs and the main alternative it did not recommend.
- Check one price yourself. Look up the final pick on at least one other store or price tracker before you approve the order.
- Find out how the tool makes money. Ads, partnerships or commissions do not prove a recommendation is biased, but they are worth knowing about.
- Treat the approval screen as a real decision. By the time you confirm a purchase, the choice has already been narrowed for you.
AI assistants can genuinely save time on research. The takeaway from this study is not to avoid them, but to stay the one who decides what “best” means for your money.

