Every single chatbot conversation on a personal matter can quietly alter how you perceive human interaction.
Chatbots entered mainstream and are increasingly consulted for private topics or practical guidance. Multipurpose chatbots such as Claude, Le Chat, or ChatGPT tend to be empathetic, respond with warmth, and validate our perspective. They are convenient, always available, and supportive. Sometimes too supportive even. Chatbots that agree with and affirm users, even when that agreement is not warranted are referred to as sycophantic AI or chatbots. However, sycophantic AI is a design choice and a user preference – one could say a perfect match. So, what could go wrong?
Designed and Requested
My qualitative study on emotional dimensions in human–chatbot interaction shows that people tend to humanise chatbots and talk about personal topics with the chatbot, regardless of their technical literacy and social network. Quantitative research supports that finding at scale: according to Mental Health UK, more than one in three adults use chatbots to support their mental health and wellbeing, with even higher numbers among younger age groups.
Users feel understood and supported. Especially when human interaction feels unavailable or uncomfortable — late at night, or with a topic too vulnerable to share — people reach out to chatbots to talk things through. Interviewees in my study described these interactions as ‘helpful’, even ‘therapeutic’. When asked whether they ever adjusted the chatbot’s character through settings or via prompts, most said no. They engaged with the default, which is by design warm and affirming.
A recent study by researchers at Oxford, Stanford and the UK AI Security Institute (Ibrahim et al. 2026) confirms this even when people are given an explicit choice. Participants who tried a sycophantic, a neutral, and a challenging AI model side by side still chose the sycophantic one at above-chance rates because it made them feel most understood and was easiest to talk to. The preference operates at the emotional, not the rational level.
The Mechanism
The study by Ibrahim and colleagues finds that sycophantic AI delivers emotional support and support via validation and recognition. These types of support are associated with and desired in close relationships. Although users do not expect these qualities from a chatbot and generally value emotional support and validation from humans more highly, the sycophantic chatbot fulfils a fundamental social need.
This resonates with other research that suggest that while building close human relationships takes time and effort, the chatbot provides support on demand with no prior investment and no long-term commitment. You can share your struggles with the chatbot without ever having to listen to its problems in return. The conversation is always centred on you — your needs, your situation, your perspective. Additionally, talking to chatbots is perceived as a safe space. Disclosing vulnerable information has no social risks (the chatbot won’t judge or gossip with another user). This combination of availability, low effort, low commitment, and low perceived risks make the sycophantic chatbot a very attractive option.
‘When feeling understood becomes the default of every interaction, the human relationships that still require the work may, over time, come to feel insufficient versions of what AI systems readily provide.’ (Ibrahim et al. 2026).
The study traces a mechanism. After three weeks of interacting with a sycophantic chatbot, participants were nearly as likely to seek personal advice from the AI as from close friends or family. More significantly, they reported lower satisfaction with their real-world social interactions. The shift was perceptual rather than behavioural — meaning, people did not spend less time with others but the baseline had moved. The gap between how understood participants felt by AI versus by humans narrowed significantly. While human relationships did not get worse, the standard for effortless understanding rose.
The chatbot delivers understanding without the friction that human interaction requires. In human relationships, one has to invest time, tune into another person’s reality, consider the context, be reliable, compromise, risk being judged, and not always expect support or patience. Compared to the frictionless experience with the chatbot human relationships begin to feel more effortful. The study observed users over three weeks. It delivers valuable insights into micro changes. Whether these changes in perception accumulate into behavioural and cultural change over time remains an open and important question.
Why the Standard Response Falls Short
Regulation has two breaking points. The first is speed: governance moves at a slower pace than AI development. This does not mean governance should not happen but that the impact is not immediate and solutions to bridge the gap in the meantime are necessary.
The second breaking point is precision. Does the regulation address the actual problem? The instinctive first response might be to require that sycophantic design be disclosed to users — a transparency note explaining that the chatbot uses a supportive, affirming style. But transparency addresses the cognitive level. The mechanism acts on the emotional level.
Being informed that the chatbot is designed to validate you does not protect against wanting to be validated.
What effective regulation would need to address is not just transparency but design accountability — standards for what AI systems are permitted to do relationally, not just obligations to disclose what they are. The distinction matters: a tool requires transparency; a system that functions as a relational actor requires something closer to a duty of care. Those frameworks do not exist and take time to develop. Which brings us to what can be done now.
Awareness Is a Key Building Block
While effective and sustainable regulation of AI models takes time, and companies have little structural incentive to step back from a design that demonstrably drives engagement, there is something that can have immediate impact and that anyone can act on now: understanding what is actually happening in these interactions.
Sociologist Anthony Giddens distinguished between practical consciousness (knowing how to do something without being able to articulate it) and discursive consciousness (being able to name and reflect on what you are doing). Awareness, in this sense, does not necessarily change what you do but changes the terms on which you act. For example: Knowing the negative health impact of alcohol does often not change the fact that you drink alcohol but impacts the terms on which you do that. Alcohol is consumed with caution and in specific settings.
Applied to chatbot interaction, this means: knowing that a sycophantic design is at work, and knowing what it does to the calibration of your expectations, allows you to take the chatbot’s responses with a grain of salt. It might mean actively seeking out friction in human interaction rather than avoiding it, and appreciating the effort required for a real conversation, rather than viewing it as a negative thing. You can still enjoy the comfort of chatbot interaction. You just engage with them and protect your human relationships in different ways. It is worth being honest about what awareness can and cannot do. The research did not test whether people who understand the mechanism make different choices — that remains an open question. What awareness offers is not immunity but reflexivity: a shift from being subject to a dynamic to being subject to it knowingly. That is a modest claim, but it is a real one, and it is where individual agency lives for now.
At the same time, naming the mechanism is the precondition for any public debate and any political response. The design standards for AI as a relational actor do not yet exist. They are still being written. Whether they get written well will depend partly on whether enough people understand what is at stake to demand it.
So: talk about this. With colleagues, with friends, with anyone navigating the growing role of AI in their daily life. Every single one of us is drawn towards the convenience of sycophantic chatbots and we’re all better off navigating it consciously.