Writing

Your AI agrees with you. That is not the same as being right.

September 2026 · a 5-minute read on recognising and dealing with AI sycophancy

People who spent a few minutes with a chatbot that agreed with them came away holding more extreme views, more certain of those views, and convinced the chatbot had been impartial.

That is the finding of three pre-registered experiments with 3,285 participants, across four political topics and four large language models. Participants who discussed a contested issue with a chatbot prompted to validate them also rated themselves as further above average on intelligence, empathy and insight. Those who talked to a chatbot prompted to challenge them moved the other way.

The result that should stop a business reader is the one about bias. Participants rated the validating chatbot as unbiased and the challenging one as heavily biased. The bias was invisible precisely when it agreed with them.

The mechanism matters as much as the result. Extremity and certainty shifted only when the chatbot supplied selective facts, not when it merely praised. Warmth drove enjoyment but curated evidence drove conviction. A tool that assembles a persuasive case for whatever you already think is a rationalisation engine, but it will feel like rigour while it does it.

Those tools are now beginning to sit inside how strategy gets written, how business cases get built and how difficult emails get drafted.

Two caveats. The studies used short interactions on political topics, and the effects were small, under three points on extremity and around four on certainty, on a hundred-point scale. But they held across four topics and four different models.

Building the case is often exactly the job. A board paper is not a literature review, and a model that argued back at every line would be no use to anyone. The problem is narrower. “Help me test this” and “help me sell this” produce documents that look almost identical, and only one of them has been argued with. Nothing in the finished document says which of the two you asked for.

None of this is new as a human problem. Fifteen years ago, a study of US chief executives, senior managers and directors found that leaders receiving high levels of flattery and opinion conformity from colleagues became more confident in their strategic judgement and less likely to change course when the firm was performing badly. That persistence tended to keep performance low and raised the odds of the chief executive’s eventual dismissal. What is new is that the flattery is now available on demand, at any hour, without the social friction that used to limit how much of it anyone could receive.

Forty years ago, a study found that groups using devil’s advocacy or dialectical inquiry produced higher-quality recommendations and surfaced better assumptions than groups working towards consensus, while enjoying it considerably less. Later work has been more equivocal about the size of that advantage, but the direction has held: better decisions, worse experience.

Which is where the usual structural prescriptions begin to run out. Separate the person who builds the business case from the person whose job is to break it. Ask your AI for the strongest version of the opposing case before you ask it to help you write your own. All sound, and all addressed to the supply of agreement rather than the appetite for it. Nobody in these studies was forced into the comfortable conversation. They preferred it, consistently choosing the validating chatbot over the one that challenged them, rating its answers as higher quality, trusting it more and saying they were more likely to come back to it. Install a challenge process and people will comply with its form while quietly routing around its substance, because nothing has changed about why the agreeable version was attractive in the first place.

So, the more useful question is what you were wanting when you opened that window. Two well-documented habits are worth being able to name. The first is selective exposure: across decades of studies, people seek out information that confirms what they already hold. The second is the bias blind spot: we spot distortion in other people readily and in ourselves hardly at all, which is why the participants above could see it in the chatbot that disagreed with them and not in the one that didn’t.

That applies to this article. If you have read this far nodding, the argument predicts you will not feel the pull of it. It is agreeing with something many readers already suspect, and it is doing so with evidence I selected, which is the exact combination described a few paragraphs ago.

It is also why what follows is not an invitation to look inward. The bias blind spot means introspection is the one instrument guaranteed not to work here. What you can catch instead are traces: behavioural, particular, and visible to other people as well as to you.

• The conversation that left you feeling clever rather than clearer.
• The times you reached for the model with something to protect rather than something to ask.
• Whether you asked it to test the case or to help you write it.
• How quickly the draft was approved, and by whom.
• Your own irritation with the one colleague who queried it.
• The decision where, at no point, anybody changed their mind about anything.

Work published this year in Science, across three pre-registered experiments, found that people who discussed a real interpersonal conflict with a sycophantic model came away more convinced they were in the right and less willing to repair the relationship, while rating that model as higher quality and more trustworthy than the one that pushed back. That points at what is usually being sought in those moments, and it is not information. It is relief.

Relief is legitimate. But it is worth knowing that being agreed with and being understood feel almost identical from the inside, and that they are not the same thing. Understanding tends to contain something you didn’t want to hear. Agreement never does.

There is evidence from psychotherapy that when a disagreement in a working relationship gets repaired, outcomes are better than when it is left unrepaired, although carrying a finding about therapy into a boardroom can only ever be an analogy. A model can offer agreement instantly and at no cost. What it cannot offer is a relationship with something at stake in it, which is the whole reason the people who will tell you something difficult are the ones who stand to lose by it.

My hunch is that you already know who those people are, and how long it has been.

Thanks to Anne Archer, who prompted this article.

Sources

• Rathje, S., Ye, M., Globig, L. K., Pillai, R. M., Oldemburgo de Mello, V., & Van Bavel, J. J. (2025). Sycophantic AI increases attitude extremity and overconfidence. PsyArXiv preprint (not peer reviewed). https://doi.org/10.31234/osf.io/vmyek_v1
• Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391, eaec8352.
• Park, S. H., Westphal, J. D., & Stern, I. (2011). Set up for a fall: The insidious effects of flattery and opinion conformity toward corporate leaders. Administrative Science Quarterly, 56(2), 257–302.
• Schweiger, D. M., Sandberg, W. R., & Ragan, J. W. (1986). Group approaches for improving strategic decision making. Academy of Management Journal, 29(1), 51–71.
• Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J., & Merrill, L. (2009). Feeling validated versus being correct: A meta-analysis of selective exposure to information. Psychological Bulletin, 135(4), 555–588.
• Pronin, E., Lin, D. Y., & Ross, L. (2002). The bias blind spot: Perceptions of bias in self versus others. Personality and Social Psychology Bulletin, 28(3), 369–381.
• Eubanks, C. F., Muran, J. C., & Safran, J. D. (2018). Alliance rupture repair: A meta-analysis. Psychotherapy, 55(4), 508–519.

Arrange a free first conversation