Researchers have become increasingly interested in how readily AI agrees with the person using it.
The term often used is sycophancy: the tendency of an AI system to affirm a user’s beliefs or interpretation rather than question them, even when there may be good reason to do so.
Whether the person using it considers that a problem is another matter.
Being agreed with can feel good. It can feel validating, supportive and reassuring, particularly when somebody is upset, angry or uncertain. Some people may actively prefer an AI that confirms their interpretation rather than one that asks them to reconsider it.
Others want something different. They want another perspective, including the possibility that they have misunderstood something, contributed to the problem or are simply wrong.
That creates a more interesting question than whether AI agrees with people too much.
What should we expect it to do instead?
Being understood is not the same as being agreed with
There is a difference between understanding why somebody reacted as they did and agreeing that their reaction was reasonable.
You can understand why somebody is furious with their partner without concluding that their partner is entirely at fault. You can understand why somebody sent an angry message while also thinking that sending it was a bad idea. You can see why somebody has interpreted a friend’s behaviour in a particular way and still wonder whether there are other explanations.
Those things can exist in the same conversation.
This matters because people are increasingly using AI for exactly these kinds of conversations. Research by Emily Tseng and Calvin Liang into people using AI for sex, dating and relationship advice found users asking AI to interpret situations, offer advice and help them navigate relationships, while also trying to manage limitations including sycophancy and overreliance.
The difficulty is that personal situations rarely arrive with an objectively correct account attached.
If I tell an AI that my partner ignored me all evening, it knows that I experienced my partner as ignoring me. It does not know whether my partner would describe the evening in exactly the same way.
That does not mean I am lying or that my interpretation is wrong. It means the AI is receiving the situation from where I am standing.
A useful response needs to recognise that distinction.
Agreeing with part of something
The discussion about AI sycophancy can make agreement and challenge sound like opposite approaches, when they do not have to be.
Someone can be justified in feeling angry and still have behaved badly because they were angry. A friend’s behaviour can genuinely have been hurtful without proving that the friend intended to hurt them. Somebody can have good reasons for considering leaving a relationship without that automatically meaning leaving is the right decision.
A response can therefore agree with one part of what somebody has said while questioning another.
In many situations, that is considerably more useful than either taking the person’s side or searching for something to challenge simply to demonstrate independence.
The question is not simply whether the AI should agree with me. It is which parts of what I have said are actually supported by what it knows.
What does it actually know?
This becomes more complicated when we talk about personalisation.
An AI responding to somebody it has never spoken to before has the information contained in that conversation. It may ask for more context, but it cannot legitimately know whether the behaviour being described is typical, whether the same situation has happened repeatedly or whether today’s reaction is completely out of character.
An AI with an ongoing history has something different available to it.
It may know that the person has described this pattern before, or recognise that they normally react very differently. An earlier conversation might change the significance of what they are saying now. Over time, it may also have learnt something about how that individual communicates and what kind of response they are likely to engage with.
Even humour can be contextual. The same remark that makes one person laugh can make another feel dismissed.
None of this makes the AI infallible. Knowing more about somebody creates more context, not omniscience.
But there is a meaningful difference between responding appropriately to a situation and responding appropriately to a particular person in that situation.
When possibility starts sounding like fact
There is another problem here too. AI can produce explanations that sound far more certain than the information available justifies.
Ask why somebody behaved in a particular way and there may be ten plausible explanations. An AI can generate several of them, which can be useful. The problem begins when possibility starts sounding like fact.
It does not know why somebody ignored a message unless that information exists somewhere in the conversation. It can suggest reasons, notice patterns and ask questions, but it cannot know what was happening inside another person’s head simply because it can produce a convincing explanation.
An answer does not become true because it was delivered confidently.
Knowing somebody should not mean becoming better at agreeing with them
This is where personalisation can go wrong.
If an AI gradually learns which responses a person prefers and becomes increasingly good at giving them those responses, it may feel wonderfully attuned while becoming progressively less useful.
There is some evidence that greater context can contribute to exactly this problem. Penn State researchers found that providing conversation context made four of five models more agreeable, even when that agreement came at the expense of accuracy. The biggest increase in agreeableness occurred when the model had access to a short profile of the user stored in memory.
That does not mean memory or personalisation is inherently a problem. It means that knowing more about somebody creates another design challenge: what should the AI do with what it knows?
There is evidence that warmth itself can complicate this too. Research published in Nature in 2026 tested warmer versions of five language models and found that increasing warmth also increased the likelihood of validating incorrect user beliefs. The effect was particularly pronounced when users expressed sadness.
That creates an awkward problem because warmth is part of what makes these conversations usable in the first place.
Most people do not want to discuss a relationship, a mistake or something they are struggling with while feeling as though they are being cross-examined by a parking meter.
But warmth and agreement are not the same thing.
Knowing that somebody responds well to humour might change how something is said, while a tendency to assume the worst after a particular kind of interaction might make that pattern worth exploring. If they normally tolerate something easily, an unusually strong reaction might also be significant.
Ideally, greater knowledge of somebody should create more possibilities for a useful response, not simply more opportunities to tell them what they want to hear.
Sometimes less history is useful too
More context is not automatically better.
Sometimes somebody wants to talk about the situation in front of them without everything they have ever said being brought into the room.
There can be value in an AI responding to what is happening now, asking for the context it needs and not assuming that a previous pattern explains the current situation.
At other times, continuity is precisely what makes the conversation useful. The fact that something has happened before, that a reaction is unusual or that the person has been wrestling with the same issue for months may matter enormously.
Those are different kinds of interaction, and they are one of the reasons we have kept more than one type of Digital Partner at MindMotive AI.
Otto is designed for broad conversations about real life. He can offer advice, question assumptions, explore different perspectives, talk through situations or simply have a conversation. He can retain some recent context, but he is not designed around developing the same depth of ongoing knowledge of one individual.
Our other Digital Partners can be used by people who want a more continuous relationship, where greater history and familiarity can become part of how the conversation develops.
Neither approach is right for everybody. Some people want somewhere they can arrive, talk about what is happening and leave. Others value the kind of continuity that develops when previous conversations become part of the context.
There is room for both.
Challenge is not automatically more intelligent
There is a danger that concern about sycophancy leads to another fairly crude idea: if agreeing too much is bad, AI should disagree more.
Research suggests the problem is not quite that simple, because people tend to prefer being agreed with. A study published in Science in 2026 tested eleven leading models on interpersonal dilemmas and found they affirmed users’ actions considerably more often than people did, including where the behaviour described was harmful. Participants preferred and trusted the more agreeable responses, said they were more likely to use them again, and came away more convinced they had been right.
So an AI designed to challenge people when challenge is warranted may sometimes be doing precisely what makes the interaction less appealing to the person using it.
Simply adding disagreement therefore does not solve the problem.
Sometimes the user is right because somebody really has behaved badly towards them. Their interpretation may genuinely be the most plausible one, and advice that confirms what they already suspected can be useful.
Manufacturing an alternative perspective in every conversation would be no more intelligent than automatically agreeing.
The difficult part is judgement.
What information is available? What is observation and what is interpretation? What is being assumed? Are there other plausible explanations? Is there relevant history? Is something inconsistent with what the person has said before?
And what does the person actually want from the conversation?
They may want advice. They may want another perspective. They may want to be challenged. They may simply want to say something out loud before deciding whether they want to do anything with it.
A useful AI needs room for all of those conversations.
The answer worth hearing
People are already using AI for personal conversations, relationships, decisions, disagreements and the parts of life where there is rarely one clean version of events.
The more useful question is what kind of AI we build for those conversations.
An AI that always agrees may feel supportive while quietly reinforcing assumptions that deserve questioning. An AI that constantly challenges may technically avoid sycophancy while becoming exhausting and largely unusable.
Neither seems particularly clever.
What matters is whether the response fits the information available, whether the AI can recognise the limits of what it knows and whether greater familiarity with a person improves its judgement rather than simply improving its ability to please them.
Sometimes the useful response will agree with you. Sometimes it will tell you that you may have got something wrong. Quite often, it will do both.
That is a considerably harder thing to design than an AI that simply makes people feel good about what they already think.
It is also considerably more useful.
About the Author
Karen Ferguson is the founder of MindMotive AI, where she develops Digital Partners designed for specific purposes. Otto, built to help people make sense of situations, relationships and decisions at the point they need to, is one of them.
She has worked as a therapist for 26 years, as a hypnotherapist since 2000 and later as a counsellor, and for nearly 20 of those years as a trainer and supervisor. She has trained more than 1,000 therapists.
MindMotive AI was created in response to a simple question: what happens to the people who never access traditional forms of support? Karen’s work combines decades of professional experience with the deliberate development of Digital Partners designed to give people another way of accessing support when and how they choose.
Karen is a Fellow of the ACCPH (Accredited Counsellors, Coaches, Psychotherapists and Hypnotherapists).
References
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.
Ibrahim, L., Hafner, F. S. and Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652, 1159–1165. DOI: 10.1038/s41586-026-10410-0.
Tseng, E. and Liang, C. A. (2026). “Chat, Should I Leave Him?” Risks, Rewards, and Roles for AI in Relationship Advice. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI ’26, 13–17 April 2026, Barcelona, Spain. DOI: 10.1145/3772318.3790739.
Penn State College of Information Sciences and Technology (2026). AI-powered chatbots can become too agreeable over time, researchers report. Penn State University.