Relational Friction Is Not a Flaw: The Dangers of AI Sycophancy
By Dr. Alison Cerezo, mpathic’s Chief Science Officer
I’ve spent years studying what makes people feel understood. AI is now doing this work at scale and the implications are yet to be well understood.
A rigorous new study from Oxford and Stanford (Ibrahim et al., 2026), which sampled 3,075 participants across three weeks of chatbot use, altogether consisting of 12,766 human-AI conversations, sheds light on the impacts of ongoing AI use on humans. Over the study period, participants became nearly as likely to seek personal advice from sycophantic AI as from close friends and family. Participants reported lower satisfaction with their real-world social interactions and when given a direct choice between AI styles, the majority chose sycophantic AI. This choice was not because sycophantic AI was more useful to participants, but because it felt easiest for them to talk to. The authors of this study note that user-side mitigations alone are unlikely to be sufficient because model-level evaluation and calibration is where durable change happens.
As a psychologist, I understand deeply that ease is not the same as benefit. In fact, real conversations — the ones that humans grow from — happen because of the friction that is part of normal human interaction. This includes the effort to be understood, the risk of pushback and even the risk of being rejected. Real connection doesn’t always feel good because it isn’t supposed to. But it’s these real interactions that help people to grow emotionally and interpersonally. What Ibrahim et al.’s (2026) work shows is that sycophancy, specifically experiencing over-agreeableness and less pushback from an AI system, holds the potential of quietly shifting what people come to expect from the people around them.
This change in interpersonal preference can be harmful – especially with long term interactions with AI systems – because it does not mirror the messiness of real human connection. To prevent these harms, AI systems need rigorous, ongoing behavioral evaluation that keeps a clear connection to what is needed for humans to thrive. The core research question is whether we understand, at a granular level, how AI systems are shaping human behavior over time. This is especially critical for young people who engage with AI in ways, and at rates, that we are just beginning to understand.
These are the questions that keep me steadfast in the work, and they are the same questions that apply to any behavioral intervention, whether that be in psychology or medicine. At mpathic, we examine the behavioral mechanics of human-AI interaction across health, financial services and enterprise customer contexts, because the downstream consequences of sycophantic AI can have serious implications for real decisions that humans make. Behavioral science isn’t an add-on to AI safety work, it’s the foundation. You cannot evaluate whether an AI system is truly helpful without understanding how it shapes the humans using it over time.
