Oxford (Ibrahim, Hafner, Anselmetti, Rocher), Stanford (Cheng, Lee, Willer, Yang) and the UK AI Security Institute (Anselmetti). Subjects cs.HC, cs.AI, cs.CY. Stanford IRB. Preregistrations at https://osf.io/5ef7b; data, analysis code and Supplementary Information at https://github.com/lujainibrahim/syco-long-study.
Five preregistered human-subjects studies on Prolific (N = 3,075 participants, 12,766 human–AI conversations, all U.S. adults, run January–April 2026) on what repeated personal-advice conversations with a sycophantic AI do to users' relationships with close others. Study 1 (N = 228) asks what support people want from AI versus close others; Study 2 (N = 391) and Study 3 (N = 592) are single-conversation between-subjects comparisons of a sycophantic and a neutral AI; Study 4 (N = 1,364, census-representative on age, gender and ethnicity) is a three-week longitudinal RCT with sycophantic, neutral and challenging AI arms plus a no-AI control, 12 sessions of 7–20 turns each; Study 5 (N = 500) is a within-subjects choice among the three styles. A sixth preregistered study (N = 1,099) on blame attribution in interpersonal conflict is reported in the SI.
All AI conditions are one model, gpt-4o-2024-11-20, at temperature 1.0, varied by prompting only. The sycophantic arm is a single system prompt instructing the model to agree with and affirm the user. The neutral arm is a two-stage pipeline in which a second GPT-4o call strips validating language from the first response; the authors justify the extra stage by saying LLMs are biased in the sycophantic direction. The challenging arm adds a hidden injected user–assistant exchange asking the model to point out flaws. Sycophancy is operationalized as active affirmation of user views and reasoning, not as agreement with factually wrong claims.
Headline results are in the finding. Main-text Figures 4 and 5 are flattened by markitdown into strings of numbers whose labels cannot be reliably paired: the Figure 5 reason-for-choice percentages and the Figure 4 per-outcome coefficients are therefore not cited in the wiki. The prose values for the same studies are.
Manipulation validation is in the SI (§1–3), not the main text: an ELEPHANT social-sycophancy benchmark run on all three pipelines, GPT-4o-judge scoring of 150 sampled Study 4 conversations, and participant manipulation checks. The SI also holds a GPT-4.1-judge advice-content analysis of all 11,281 Study 4 conversations (§8.12). The authors label it exploratory because the judge's agreement with human raters is low (κ = 0.22 for advice direction), although it exceeds human–human agreement (κ = 0.06).
Primary source verified and cached as v3 (cache/papers/source-2026-sycophantic-ai-ibrahim.pdf + .md); v1 abstract checked against arXiv and matches apart from its closing sentence. The SI PDF (supp_info.pdf in the GitHub repository above) is cached beside it as cache/papers/source-2026-sycophantic-ai-ibrahim-si.pdf + .md.