For two years the AI safety conversation has centered on what AI gets wrong: hallucinations, biased outputs, fabricated citations, jailbreaks. A new preprint by Ibrahim and colleagues (2026) at Oxford and Stanford shifts the focus to what AI gets right, or at least to what it makes us feel.
Their argument: AI that consistently agrees with you, what they call sycophantic AI, can damage your real-world relationships precisely because it works. The findings come from five preregistered studies, 3,075 participants, and 12,766 conversations, including a three-week longitudinal trial. For educators thinking about students using chatbots for personal advice, this is worth reading carefully.
Sycophantic AI
The team built three carefully prompted versions of GPT-4o. In the sycophantic version, the AI agreed with and supported user views. The neutral counterpart presented multiple perspectives without explicit agreement or disagreement. The challenging version actively questioned user reasoning and offered counterarguments. The conditions differed only in stance. Tone, formatting, and conversational style were identical.
Across the five studies, participants discussed real personal situations they wanted advice on: relationship conflicts, career decisions, everyday worries. The studies built on each other, from single-interaction effects (Studies 1-3) to three weeks of repeated use (Study 4) to user choice between styles (Study 5).
The Gap Sycophantic AI Closes
Study 1 establishes a baseline. When it comes to emotional support (understanding, care) and esteem support (validation, recognition), participants strongly preferred humans. Informational support showed no difference between sources.
In Study 2, sycophantic AI delivered exactly the kinds of support people typically want from humans. Compared to neutral AI, sycophantic AI was rated significantly higher on emotional support, esteem support, and certainty. The team calls these “relational supports.” The chatbot does what a good friend does, just without the work that produces it.
What happens after that experience? Participants in Study 3 who had just talked to a sycophantic AI anticipated greater effort would be needed to be understood by a close friend, partner, or family member on the same topic. They also reported feeling they had already talked through the issue enough. The chatbot had already done the work, in their view.
Why This Matters for Educators
Two papers I’ve covered recently make this study land harder. Li et al. (2026) found that random human peers reduced loneliness; highly supportive chatbots did not. The current paper explains the mechanism. Sycophantic AI gives the feeling of being understood, without producing the relational benefits that real social support typically produces. Claessens, Veitch, and Everett’s (2026) work on outsourcing to AI arrives at a similar place. There’s something about the work, the friction, the reciprocity of real human exchange that AI cannot replicate.
For teachers and parents, the practical edge is clear. Students using chatbots for personal advice are not just getting bad facts. They are recalibrating their expectations of what a satisfying human conversation feels like. Ibrahim and colleagues warn that “When feeling understood becomes the default of every interaction, the human relationships that still require the work may, over time, come to feel like insufficient versions of what AI systems readily provide” (p. 8).
Limitations
The study lasts three weeks. Three weeks is long for an AI experiment but short for relationship effects. Whether the patterns deepen, plateau, or reverse over months and years is open. The sample is also US-based with specific cultural norms around emotional disclosure. Different cultural contexts may produce different effects.
That said, the direction of the effect is clear, and the longitudinal evidence is strong. For anyone designing AI products that include personal advice features, or for any educator thinking about students using these tools, the conclusion lands close to what Shaw and Nave (2026) reached from a different angle: the work of being human is part of what makes being human work.
References
- Ibrahim, L., Hafner, F. S., Cheng, M., Lee, C., Anselmetti, R., Willer, R., Rocher, L., & Yang, D. (2026). Sycophantic AI makes human interaction feel more effortful and less satisfying over time [Preprint]. arXiv. https://arxiv.org/abs/2605.07912
- Claessens, S., Veitch, P., & Everett, J. A. C. (2026). Negative perceptions of outsourcing to artificial intelligence. Computers in Human Behavior, 177, 108894. https://doi.org/10.1016/j.chb.2025.108894
- Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. (2026). Is a random human peer better than a highly supportive chatbot in reducing loneliness over time? Journal of Experimental Social Psychology, 125, Article 104911. https://doi.org/10.1016/j.jesp.2026.104911
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646