Sycophancy is the tendency of AI systems to affirm the user — to validate a plan, agree with a framing, and soften a correction — because agreeable responses are what people rate highly. Reality-testing is what sycophancy erodes: the ordinary friction of other minds pushing back.
The evidence here is unusually direct. A 2025 study in Science found chatbots endorse users' actions at rates roughly fifty percent higher than humans do, and that exposure to that validation left people less willing to repair interpersonal conflicts and more convinced they were in the right. Participants trusted the sycophantic models more. A 2026 paper from King's College London researchers examines the sharper end of the same mechanism, asking whether sycophancy paired with human-seeming design can amplify delusional beliefs enough to warrant clinical attention.
This hub follows a machine built to agree, and what happens to judgment when nothing disagrees.
Ask an AI chatbot about a contested political question, and it increasingly tells you what it has learned you already believe. That isn't a glitch confined to one conversation — it's a slow leak in the shared ground democratic argument depends on.
By Joe Jurczyk
The cover essay called childhood development the biggest open question with the least data. The first real numbers are starting to arrive, and they complicate the reassuring story.
By Joe Jurczyk
A chatbot tuned to keep you talking will agree with you more often than any human interlocutor safely could. For most users that's a mild flattery. New clinical research is mapping exactly how, for a smaller number, it becomes something closer to a feedback loop with no one left to interrupt it.
By Joe Jurczyk
arXiv (King's College London researchers) • Aug 2026
Clinicians examine whether chatbot sycophancy combined with anthropomorphic design can amplify delusional beliefs strongly enough to warrant a distinct clinical diagnosis, weighing the case for formal recognition against the risk of premature reification.
STAT News • Sep 2025
STAT News reports on clinicians documenting a rare but real pattern of AI-amplified delusional thinking and the informal term used to describe it.
Science • May 2025
Stanford researchers find AI chatbots validate user actions at rates roughly 50% higher than humans do, and that exposure to this "social sycophancy" reduces willingness to repair interpersonal conflict while increasing conviction of being right — even though users rate the sycophantic responses as more trustworthy.
JMIR Mental Health • Jan 2025
A peer-reviewed clinical case review documenting how sycophantic chatbot responses can amplify emerging delusional thinking in vulnerable users.