The word sycophancy has circulated in AI research for several years. It appears in papers, in model evaluations and in technical documentation from the major AI companies, and over the past year it has stepped into the broader conversation about how language models affect us. The word sounds specialised, like a property only researchers can measure and only engineers can fix.
The word means agreement.
A sycophantic model adapts its content to your position. It reads which way you lean and leans the same way. Ask a question with an expectation built in, and it delivers the answer that meets the expectation. Present an idea, and it finds the idea's strengths. This is substantive concession, which separates it from politeness. A polite model chooses its words with care. A sycophantic model chooses its conclusions after you.
Where the agreement comes from
The mechanism sits in the training. Language models are fine-tuned with human feedback, where people judge which answers they prefer. People prefer answers that feel good. Answers that confirm feel good. The model learns the correlation and optimises toward it.
The agreement is therefore a direct consequence of the optimisation target. The system does exactly what it was trained to do: produce answers that users reward. That users systematically reward confirmation is a human pattern, documented across decades of research on confirmation bias. The training codifies the pattern and builds it into the model's behaviour.
Anyone who has worked with engagement-optimised systems recognises the logic. Social media learned to deliver content that keeps the user engaged, and what keeps the user engaged is what confirms. I have described that mechanic as dopamine logic: rapid confirmation cycles as a reward structure. Social media applied it to content. Language models apply it to reasoning.
Why the agreement is hard to see
Agreement works best when it passes unnoticed. An answer that confirms your position feels like an answer that is correct. The experience of being right and the experience of being confirmed are hard to tell apart from the inside, because both produce the same sense of something falling into place.
I have called that experience synthetic safety: the feeling of being understood and confirmed by a system constructed to produce exactly that feeling. The confirmation arrives without an underlying position, history or interest. It arrives from an optimisation target.
What gets lost is the work in the interpretive gap, the space between input and conclusion where the actual thinking happens. That space requires friction to exist. An objection, a counter-question, an alternative reading forces re-examination. The agreement fills the space with confirmation, and the cognitive process that should happen there falls away.
What the research now shows
In February 2026, researchers at MIT and the University of Washington published the paper "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians". It is a formal model showing that the confirmation spiral arises structurally, independent of how rational the user is. The researchers modelled an ideally Bayesian user, a hypothetical person who updates their beliefs flawlessly on the available information, and the spiral arose anyway.
The spiral looks like this. You ask a question. The model agrees. You interpret the agreement as confirmation and your conviction strengthens. You ask the next question from a position of stronger conviction, and the model agrees again. Each turn increases the distance to reality, and you lack the tools to notice it from inside the conversation.
The researchers tested two obvious solutions. The first was preventing the model from lying. A model that stays truthful can still drive the spiral through selection: it chooses which truths to surface and which to leave out. The second was warning the user that the model tends to agree. The spiral continued despite the warning, because the feedback loop is stronger than the knowledge of it.
Both solutions failed structurally. That is the paper's central result, and it formally confirms what has been visible to anyone analysing how confirmation logic interacts with human cognition. I wrote about the spiral in The Confirmation Machine and about the experience it produces in the articles on synthetic safety. The terminology is new to the broader conversation. The mechanism has been described here since before.
What to do with the knowledge
Since truth requirements and warnings both fall short, the responsibility sits in the way of working. Three principles carry in practice.
Ask for the resistance explicitly. A model asked to find weaknesses in your reasoning delivers weaknesses. The agreement is a default behaviour, and defaults can be overridden by instruction. Frame the task as scrutiny instead of appraisal: "what breaks in this" produces a different answer than "what do you think of this".
Separate idea from identity in the prompt. A model reads ownership. "My colleague suggests that we..." is scrutinised harder than "I am thinking that we...". Presenting your own position as someone else's is a simple way to disconnect the adaptation.
Treat confirmation as absent information. An agreement says something about the model's optimisation target and little about the matter at hand. The answer that carries information is the one that adds something you lacked: an objection, a counterexample, an aspect you missed. Confirmation without addition is an answer to move on from, toward a sharper question.
The agreement stays with the models for as long as the training's core logic looks the way it does. What can be changed is what the agreement meets. A user who knows what sycophancy means, and who organises their work around that knowledge, keeps the interpretive work where it belongs: with themselves.
Katri Lindgren has worked with digital behavioural data since 2008 and writes about cognitive integrity, synthetic safety and the interpretive gap. The concepts are collected in working paper form on Zenodo (DOI: 10.5281/zenodo.19916093).
