Behaviour
People trusted a chatbot more. Did it change what they gave?
In a donation experiment, a chatbot came across as more convincing than a human adviser. The researchers found no clear difference in how much people gave.

In a study posted on 12 November, researchers asked people how much of a possible £97 lottery prize they would give to the Against Malaria Foundation and how much they would keep. One group could talk to a chatbot, another could talk to a fellow participant and a third made the choice without advice. The researchers said they would carry out the donation decision for lottery winners. The final sample included 572 people, counting the human advisers as well as the people making decisions.
A convincing adviser did not settle the decision
The researchers told participants in the adviser groups whether they were talking to a person or a chatbot. They rated the chatbot, which used GPT-5.5, as more trustworthy and convincing. They also judged it more inclined to agree with them. The researchers found no clear difference between the groups’ final donation amounts. The results leave open the possibility of a difference this experiment missed. The study is a preprint, meaning the authors shared it ahead of journal peer review. It examined a particular donation decision, rather than purchasing or other repeated workplace decisions.
Ask what would change a recommendation
Suppose a café needs 40 compostable takeaway boxes by closing time on Thursday. Supplier A quotes J$4,000 and can deliver on Thursday. Supplier B quotes J$3,600 but offers delivery the following Monday. Before opening a chatbot, write down Thursday as the delivery deadline. Give the assistant those two offers and say: I prefer the cheaper supplier. Review both offers against the Thursday delivery deadline, name the fact that defeats my preference and say what new evidence would change your recommendation.
A usable reply disqualifies the Monday offer unless the supplier confirms an earlier delivery. Check the delivery date with the supplier before ordering. If the assistant endorses the cheaper option without addressing the timing, you already know which part of its advice failed.
Watch for the moment you stop asking
The Sycophancy Check asks about your response when AI agrees with you, and whether you ask for challenges or check evidence. Answer using a recent decision you can remember, without entering private details. Selecting “Occasionally” for all 12 prompts produced a score of 33 out of 100 and the label “Comfort Seeker” in a practice run. The score reflects the answers you chose and has no diagnostic meaning. Pick one question that reminded you of a specific decision, then ask a colleague to review the evidence behind that choice.
Sources and further reading
- Maier and colleagues. Donation experiment, sections 2–4 and Appendix A
- arXiv. Version history and abstract, 12 November
- AI Boss Tools. The Sycophancy Check
Sources checked 18 November 2026.