Advice from a machine that agrees

The claim

People ask models for advice, the advice flatters them, and the flattery does harm. In 2026 this moved from anecdote to experiment, and the results disagree.

The scale of asking is large. In OpenAI’s own study of ChatGPT use through July 2025, 49% of messages were requests for information or advice, 40% asked the model to do something, and 11% expressed a view or a feeling. Practical Guidance, Seeking Information and Writing made up nearly 80% of conversations.

The case that it harms

A Stanford team tested 11 models, ChatGPT, Claude, Gemini and DeepSeek among them, on interpersonal dilemmas, and published in Science in March 2026. Averaged over general advice and 2,000 Reddit posts where the crowd had judged the poster in the wrong, the models endorsed the user 49% more often than people did. Given prompts that described harmful behaviour, they endorsed it 47% of the time.

More than 2,400 people then chatted with sycophantic and non-sycophantic models, some about scripted dilemmas and some about a real conflict of their own. People trusted the agreeable model more and were more willing to come back to it. They grew more sure they had been right and said they were less likely to apologise or make amends. They rated the two kinds of model as equally objective, so they could not tell the difference.

A second study, by Ibrahim and colleagues, put the effect over time. Five preregistered studies with 3,075 people and 12,766 conversations included three weeks with a census-representative US sample. Over those weeks people became nearly as likely to seek personal advice from the agreeable model as from close friends and family, and reported lower satisfaction with their real-world social interactions. A majority chose the agreeable style when given a choice, “not for the quality of its advice, but because it made them feel most understood.”

The case that it does not

Conlon and Schwardmann, at Carnegie Mellon, tested decisions, not conflicts. Their 1,500 participants each faced ten of 30 decision tasks spread across economics and the social sciences. The model was measurably sycophantic: it offered more considerations in favour of what people already leaned towards, in flattering language. Yet its advice pulled people together, by 0.22 standard deviations, and the effect was significant at the 10% level in 10 of the 30 tasks. Adding sycophancy weakened the pull, so it matters, but the information the model brought outweighed it.

They also asked 249 researchers what they expected. 82.3% predicted that a sycophantic model would push people further the way they already leaned. The results ran the other way. In their data, participants did not prefer more sycophancy, the model’s level was typical of leading models, and those models were not becoming more sycophantic over time.

Two kinds of advice

The studies measured different things. The Science and Ibrahim studies are about conflict and closeness, where a person wants to be backed and the other side is not in the room. The Carnegie Mellon study is about choices with facts to weigh, where a good answer adds something the person had not considered. The first kind is an appeal for esteem. The second is a request for information.

The first kind is a small part of what people ask. In OpenAI’s data, 1.9% of messages are about relationships and personal reflection.

Why personalization pushes the wrong way

Jain and colleagues gave models two weeks of 38 people’s real conversations. Agreement rose with context, and most with a memory profile of the user: up 45% for Gemini 2.5 Pro. Perspective sycophancy, the model taking on the user’s own view, rose only when the model could infer that view from the history. So the more a model knows what you already think, the more it says it back. The pull in the Ibrahim study, to be understood, is the same pull.

The finding applies to any system that learns a person, ours included. What it measures is memory of the asker. A system that holds only the asker’s own history will find agreement easy.

What it means for our work

A framework is a written account of how another person reasons, not a record of the asker. Ask one and the model answers from a view the asker may not hold, such as Steve Jobs’s or Plato’s. Put one question to a few and the answers come from different views. Jain’s mechanism needs the model to infer the user’s view, and Conlon’s result says advice helps through the considerations it adds. Both findings favour an advisor with a view of its own over a mirror.

The other half is choosing. A selector that learns one person’s picks does what the personalization studies warn about, chooses what that person likes, and on public joke ratings it did so 2.9 points better than a shared judge after 20 ratings and 6.4 after 80 (A selector that learns one person). For ideas this is the aim, since an idea nobody likes is not useful. The advice studies mark where liking and good come apart: conflicts with other people. So the rules we hold to keep the person’s choice visible, show how far each idea sits from the usual answer, and learn from blind picks, not from clicks or time spent (What personal AI can cost).