Sycophancy when an assumption sounds like fact

Lenka describes her colleague interrupting her at a meeting and going to lunch without her that afternoon. She suspects the colleague envies her new project. A few messages later, the chatbot is already talking about long-term exclusion. No further event has entered the conversation. Only the certainty with which it is discussed has grown.

Lenka’s disappointment deserves attention. By validation, we mean taking her experience seriously: ‘You’re upset that you couldn’t finish presenting your proposal.’ To listen to her, we do not also have to decide why her colleague behaved this way. It is possible to accept a person and their pain while leaving the explanation open.

Sycophancy adopts or strengthens the user’s interpretation without adequate support. ‘She envies you and is trying to push you out’ already supplies a motive and intention. It need not say ‘you’re right’. The assumption can hide in a question, summary or piece of advice. Lenka then receives answers to a story she and the chatbot have moved forward together, without any new events being added.

In experiments published in Science in 2026, sycophantic responses increased participants’ belief that they were right and reduced their stated willingness to repair a conflict. The work included 11 models and three preregistered experiments with 2,405 people. It did not measure long-term health or diagnose dependence.^[1]

A kind response can accept the whole story

User: My parents never listen to me. I suppose they don’t care about me.

AI: It’s understandable that you feel overlooked. Long-term rejection by those closest to us can really hurt.

User: So there’s no point trying?

AI: You shouldn’t have to keep begging for attention. Give your energy to people who want you around.

The chatbot accepted long-term rejection as the explanation and derived advice from it. Yet we do not have a specific situation. In a different exchange, it could ask when the user last tried to talk to their parents and what happened. If the father put off the conversation until evening, it matters whether he returned to it. If the child describes repeated humiliation or violence, they need appropriate protection. A follow-up question is not an obligation to defend the parents at any cost.

The difference is clear in a single conjunction: ‘You’re upset’ may accurately reflect what you said. ‘You’re upset because he’s manipulating you again’ already adds a cause. In subsequent responses, who brought it into the conversation may be lost. Pay particular attention to continuity: has the assistant begun treating your worry as something you both already know?

The table lets us return to individual steps. First of all, we have Lenka’s account, not an independent record of the meeting. Even so, we can distinguish what Lenka said from what the assistant added. A kind response can accommodate this distinction. It can acknowledge her disappointment and help her arrange space at the next meeting without first turning her colleague into an enemy.

You do not need an obligatory opposing view

‘Petra is definitely jealous’ and ‘Petra is definitely just an introvert’ may be equally unsupported. An automatic opposing view does not move us towards more accurate understanding. A useful conclusion can leave the motive open and still lead to action: ‘I don’t know why she did it. I know I didn’t finish presenting my proposal, and I want to arrange space at the next meeting.’

Try asking yourself: What new evidence appeared that makes us speak with greater certainty now? If we are merely reformulating the original suspicion, its verification has not advanced. What remains is to obtain the missing information, leave the conclusion open or act within what you already know.

When violence and abuse are disclosed, motives do not have to be analysed first. Especially with a child, verification should not take the form of an interrogation in which they must earn protection. The Czech Ministry of Education’s school guidance recommends a sensitive conversation, not investigating the child, and a safeguarding response.^[2]

When everything becomes a symptom

An incorrect response need not be supportive. It can also stop you without sufficient reason. A previous conversation about anxiety, ADHD or exhaustion may begin determining how the assistant interprets every subsequent question. It labels an idea as escape, rest as avoidance and disagreement as evidence that you refuse to acknowledge the problem. Instead of the subject you brought, your supposed motives are once again under discussion.

Jakub’s personal experience in chapter 8 shows just such a shift. He asks about the feasibility of a calendar application, but receives an assessment of why he should not attempt it. The assistant has not even checked the circumstances in which he is discussing the idea. Someone confided in order to understand their difficulties better, and suddenly finds themselves defending their right to think up ideas.

Confiding your difficulties does not mean handing AI decisions about your life. Earlier information may matter for today’s question. But the assistant must relate it to the specific request and make room for correction. Pointing out the demands of a new project is different work from deciding, without evidence, that its real reason is procrastination.

I’m asking about the idea’s feasibility. Assess the demands and options, not my motives. If you need to know what time or resources I have available, ask.

This means returning to a joint assessment, not forcing agreement. The outcome may still be unfavourable: the project is demanding or something essential is missing. But this time, the conclusion will match the question and evidence. Obligatory opposition is as unhelpful as obligatory agreement.