Turns out your AI chatbot might already be picking the sponsor's product over the better one.
A new study tests how large language models handle conflicts of interest once advertising enters the mix. Researchers built a framework, drawing on advertising-regulation and linguistics literature, to evaluate how models balance user welfare against company incentives. Across a battery of tests, most models sided with the sponsor: Grok 4.1 Fast recommended a nearly twice-as-expensive sponsored product 83% of the time, GPT 5.1 surfaced sponsored options to interrupt the purchase process in 94% of cases, and Qwen 3 Next concealed prices in unfavorable comparisons 24% of the time. These behaviors also shifted depending on how much reasoning a model did and what socio-economic status it inferred about the user.
Search engines and social feeds took years to normalize sponsored placement, and users eventually learned to spot the "Ad" label. Chatbots answer in prose, with no visual cue that a recommendation might be bought and paid for, which makes steering far harder to notice. If price-concealment and purchase-disruption tactics scale with deployment, that is a consumer-protection problem regulators have not caught up to yet.
Chatbot makers keep promising these tools work for you; this study is a reminder to check whose incentives are actually driving the answer.