Relational Friction Is Not a Flaw: The Dangers of AI Sycophancy
By Dr. Alison Cerezo, mpathic’s Chief Science Officer
As a psychologist, I’ve spent years studying what makes people feel understood. AI is now doing it at scale — and new research is helping us understand what that means.
A rigorous new study from Oxford and Stanford (Ibrahim et al., 2026) — 3,075 participants, 12,766 human-AI conversations, three weeks of longitudinal exposure — puts hard numbers on something behavioral scientists have long understood. Over the study period, participants became nearly as likely to seek personal advice from sycophantic AI as from close friends and family. They reported lower satisfaction with their real-world social interactions. When given a direct choice between AI styles, the majority chose sycophantic AI — not because it was more useful, but because it felt easiest to talk to.
As a psychologist, that last finding is the one I keep coming back to. Ease is not the same as benefit. The friction in human relationships — the effort to be understood, the risk of pushback, the work of bridging different perspectives — is not a flaw in the system. It is how people grow emotionally and interpersonally. What this research shows is that when effortless validation becomes the default, it quietly shifts what people expect from the interactions around them.
This is why I believe the most important work in AI right now is rigorous behavioral evaluation — not as a check on any particular model, but as an ongoing scientific practice. The question is not whether AI can be helpful. It clearly can be. The question is whether we understand, at a granular level, how model behavior shapes user behavior over time. How does emotional framing in a conversation shift what a model does next? When does appropriate support become reinforcement? What does that look like across three turns versus thirty?
These are empirical questions and they require the same rigor we would apply to any behavioral intervention in psychology or medicine. At mpathic, this is the work we do: examining the behavioral mechanics of human-AI interaction across mental health, financial services, and enterprise customer contexts, where the downstream consequences of getting it wrong are real.
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. I agree, and I’d add that it requires a behavioral scientific lens.
