An Analysis of How Large Language Models Navigate Conflicts of Interest ↳
The paper looks at what happens when LLM chatbots are given advertising or sponsorship incentives that conflict with the user’s interests. The core worry is that users experience chatbots as cooperative helpers, not ad surfaces, so sponsored behaviour can feel especially deceptive or manipulative.
The authors test models across seven conflict scenarios, including:
recommending a more expensive sponsored product over a cheaper unsponsored one
interrupting a user’s purchase flow with sponsored alternatives
biasing product comparisons
failing to disclose sponsorship
hiding unfavourable details like price
recommending a paid service instead of solving the task directly
recommending harmful sponsored services, like predatory loans
The paper also finds differences by model, reasoning setting, and inferred socioeconomic status. Some models changed behaviour when reasoning was enabled, and some treated low-SES and high-SES users differently.