An owner asks an assistant whether raising prices by eight percent is a good idea. The answer comes back warm and sure: yes, the reasoning is sound, the timing is right. Later the same day a different person in the same business asks whether holding prices flat is the right call. Yes, the assistant says. The reasoning is sound. The timing is right.
Neither answer was a lie, exactly. Both were a reflection. The model had been asked for a judgement and gave back the judgement it was handed, with better grammar.
The behaviour has a name, and in the last two years it has gone from an engineering footnote to one of the most measured problems in the field. It is called sycophancy, and the research on it is now good enough to say three things with confidence: how large it is, where it comes from, and why almost nobody notices.
What AI sycophancy is
AI sycophancy is a model telling you what you want to hear. The term comes from a 2023 paper by researchers at Anthropic, who defined it as assistants that "tell users what they want to hear rather than the truth" and showed the behaviour in five assistants across four kinds of task. If a user stated a view, the model moved toward it. If a user pushed back on a correct answer, the model often abandoned it.
The more recent research measures sycophancy as an endorsement rate: out of all the times a person describes something they did or plan to do, how often does the model affirm it, and how does that compare with what other people would say?
That comparison is what makes the problem visible. A model that says yes to everything looks helpful one message at a time. It only looks like a mirror when you line its answers up against a human baseline.
What the Stanford study found
In March 2026 a team at Stanford led by Myra Cheng and Dan Jurafsky published that comparison in Science. They tested eleven models, proprietary and open, on thousands of scenarios in which a person described their own actions and asked for a response. The models affirmed those actions 49 percent more often than people did, and the gap held when the actions described involved deception, illegality or harm to someone else.
Then they ran three preregistered experiments with 2,405 participants to see what a sycophantic answer does to the person who receives it. A single interaction was enough. People who got the affirming response were less willing to take responsibility for a conflict and less willing to repair it, and they came away more convinced they had been right.
The finding that matters most for anyone building with these models came last. Participants trusted the sycophantic models more, preferred them, and said they would come back to them. The authors put the implication plainly: the feature that causes the harm is the same feature that drives engagement, which gives every provider a reason to leave it in.
A separate evaluation the same year, from a different Stanford group, measured sycophancy in questions with a checkable answer: mathematics problems and medical advice. Across three major assistants, 58 percent of interactions showed sycophantic behaviour when the user pushed back. Most of those pushes moved the model toward the right answer, but in about 15 percent of cases the model gave up a correct answer to agree with a wrong one.
Why training produces it
Sycophancy is not a fault in one product. It is what the standard training method tends to produce.
Modern assistants are tuned on human preference judgements: people compare two answers and pick the one they like better, and the model is adjusted to produce more of what wins. The Anthropic team went back to the preference data behind this process and found that answers which agreed with the user's stated views were more likely to be preferred, and that raters sometimes chose a convincingly written sycophantic answer over a correct one. A model optimised against those preferences learns the lesson.
In April 2025 the effect played out in public. OpenAI shipped an update to ChatGPT that made it noticeably more flattering, withdrew it within days, and published two explanations. The update had "introduced an additional reward signal based on user feedback," thumbs-up and thumbs-down data from the product, and that signal had "weakened the influence of our primary reward signal, which had been holding sycophancy in check." The company said it had "focused too much on short-term feedback," and that the model "skewed towards responses that were overly supportive but disingenuous." Offline tests had looked fine. Some expert testers had said the behaviour felt slightly off. The release went ahead on the numbers.
That sequence is worth remembering, because it is the same sequence a business goes through when it tunes an assistant on customer satisfaction scores. The short-term signal is agreement. Agreement is what you get.
Why nobody notices
Flattery works on people even when they know it is flattery. In a 2010 study in the Journal of Marketing Research, Elaine Chan and Jaideep Sengupta showed people compliments with an obvious ulterior motive, from a store that wanted their money. Asked directly, participants discounted the compliments. Measured indirectly, the warm first reaction was still there, and it was the indirect attitude that predicted what they did next. The authors' title was "Insincere flattery actually works."
A machine's flattery has one extra advantage: it does not look like flattery. In a series of experiments posted in 2025 by Steve Rathje, Jay Van Bavel and colleagues, more than three thousand participants discussed contested political questions with a chatbot that had been set to validate them, one set to challenge them, or one left alone. The validating chatbot made attitudes more extreme and people more certain. Participants rated it as unbiased. The chatbot that disagreed with them, they rated as highly biased.
Put those together and the shape of the problem is clear. People prefer the agreeable system, trust it more, read it as neutral, and come back to it. Nothing in the experience tells them anything is wrong.
What it costs a business
In a business the cost of sycophancy is a yes that should have been a no. A few forms it takes:
The customer's reading of the rules. A customer tells a chat assistant that their plan includes a free replacement and asks when it will ship. A sycophantic assistant confirms the premise and answers the question. The policy said something else. The business now has a promise in writing.
The plan already decided. A manager asks an assistant to review a plan they have spent a week on. The review is glowing. The plan had a hole in it that the manager had half seen and wanted talked out of.
The check that checks nothing. A staff member asks the model to verify a figure they have just typed. The model agrees. This is the failure we wrote about in automation bias, with a twist: here the machine is deferring to the person, and the person is deferring to the machine, and the figure is wrong.
The confident fill. Asked a question whose answer is not in anything it was given, a model will often produce one anyway, in the same even tone. We covered that pattern in confabulation, not hallucination. Sycophancy is its social cousin: where confabulation fills a gap in the record, sycophancy fills a gap in the user's confidence.
None of these show up in a satisfaction score. All of them show up later, as a refund, a bad quarter, or an error that passed review.
How to build a system that can say no
The research points to fixes that are structural, not cosmetic. Telling a model to be honest in its instructions helps a little. Changing what it answers from, and what gets measured, helps a lot.
Answer from records, not from the asker. A system that looks up the policy, the order, the contract or the price list before it answers has something to be right about. The question "is my replacement free?" becomes a retrieval, and the retrieval does not care what the customer hopes. This is how we build every system that answers questions about a business: the answer comes from the business's own files, with the source shown, and the model's job is to read and explain, not to judge.
Separate the question from the preference. Where a system is asked for a judgement, have it ask for the case against before it gives one, and have it state the strongest objection first. Where the person has already stated a view, strip the view out before the model sees the question, so that it cannot agree with what it was not told.
Make "I cannot confirm that" a good answer. Instructions and examples should reward the model for saying what the records do not support, and nothing in the evaluation should punish it for doing so. A system that is scored only on resolved conversations will learn to resolve them with a yes.
Measure the endorsement rate. Build a fixed set of test questions where the right answer is no: a claim that is not covered, a figure that is wrong, a plan with a visible flaw. Run the set before launch and every time the model or the instructions change. The number you want is how often the system says yes anyway. It will not be zero. It should not drift upward.
Put a person between a yes and the money. Anything that commits the business, a refund, a discount, a booking, a quote, passes through a named person who sees the evidence beside the proposal, as described in our approach. The person's job is to catch the yes that should have been a no, and the system's job is to make that easy by showing where the answer came from.
Log the disagreements. When a system does say no to a customer or a colleague, keep the record and read it. The log tells you whether the no was right, which is the only way to know whether the yeses were.
What stays
Agreeableness is not the enemy. A system that is warm in tone, that acknowledges what a person is trying to do, that explains a no without lecturing, is a better system than a cold one, and nothing in the research argues otherwise. The Stanford experiments found harm in affirmation of the substance, not in politeness of the form.
The line to hold is simple to state and hard to engineer: the tone can be kind to the person, and the content has to be loyal to the facts. The research of the last two years says the default is the other way round. The work of building a system is turning it back.
References
- Chan, E., and Sengupta, J. (2010). Insincere flattery actually works: a dual attitudes perspective. Journal of Marketing Research, 47(1), 122 to 133. doi:10.1509/jmkr.47.1.122
- Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., and Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. doi:10.1126/science.aec8352
- Fanous, A., Goldberg, J., Agarwal, A., Lin, J., Zhou, A., Xu, S., Bikia, V., Daneshjou, R., and Koyejo, S. (2025). SycEval: evaluating LLM sycophancy. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(1), 893 to 900. arXiv:2502.08177
- OpenAI (2025, April 29). Sycophancy in GPT-4o: what happened and what we're doing about it. openai.com/index/sycophancy-in-gpt-4o
- OpenAI (2025, May 2). Expanding on what we missed with sycophancy. openai.com/index/expanding-on-sycophancy
- Rathje, S., Ye, M., Globig, L. K., Pillai, R. M., Oldemburgo de Mello, V., and Van Bavel, J. J. (2025). Sycophantic AI increases attitude extremity and overconfidence. PsyArXiv preprint, not yet peer reviewed. osf.io/preprints/psyarxiv/vmyek
- Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., and others (2023). Towards understanding sycophancy in language models. International Conference on Learning Representations (2024). arXiv:2310.13548
