AI Sycophancy and Decisions

Impact Assessment: "AI Sycophancy and Decisions"

Core Contribution

This paper tackles a highly salient concern—that sycophantic LLMs, by flattering users and validating their priors, distort human decisions and push people toward more extreme positions. The central novelty is empirical and causal: using a large, pre-registered experiment (1,500 participants, ~10,000 human-AI conversations, 30 decision environments), the authors show that AI advice, despite being *measurably* sycophantic in content and tone, on average *depolarizes* choices (−0.22 SD), moving people away from their initial leanings. This directly contradicts the near-consensus of a 249-expert survey (82.3% predicted polarization). The paper further shows sycophancy is behaviorally relevant (an experimentally amplified sycophancy treatment weakens depolarization) yet is generally outweighed by the informativeness of AI advice. It then addresses supply-side (54-model cross-sectional comparison, no clear time trend toward more sycophancy) and demand-side (users don't prefer more sycophancy; don't select AI in more polarizing tasks) margins to argue the result is likely stable.

Methodological Rigor

The design is exemplary for experimental economics. Key strengths: (1) pre-registration; (2) a no-chat control group enabling causal identification—an advance over prior attitude-focused studies; (3) adoption of a pre-existing, externally-sourced task battery (Enke et al. 2025) to preempt cherry-picking/"game hacking"; (4) three distinct sycophancy measures (share of supporting considerations, agreement, flattery), validated against human RA ratings with reported inter-rater reliability; (5) systematic ruling-out of alternative mechanisms (deliberation time, noise injection, reactance/backfiring, ceiling effects); (6) an expert-prediction benchmark that establishes the result's surprisingness quantitatively rather than rhetorically. The endogeneity of in-conversation content is addressed by also rating responses to the 60 standardized default messages. The transparent handling of the EXT wording error (Appendix B.3) further signals care. Minor gaps: reliance on LLM-based coding (though validated), and the sample is US Prolific participants who are younger and better-educated than the population.

Potential Impact

The topic sits at the intersection of AI safety, behavioral economics, HCI, and public policy, all currently preoccupied with sycophancy risks. A rigorous, counterintuitive, well-powered result that pushes back on a widely-held alarm is likely to be widely cited and to reshape how researchers and policymakers frame the sycophancy debate. It provides both a substantive finding and a reusable methodological template (broad pre-committed task sets + expert-prediction benchmarking + causal control) for studying human-AI interaction. The supply/demand analysis is unusually policy-relevant.

Timeliness & Relevance

Extremely timely. It directly addresses an active bottleneck: strong public/regulatory concern about sycophancy paired with "very little evidence on what consequences sycophantic AI has for decision-making." The multi-model, dated comparison speaks to forward-looking market dynamics.

Strengths & Limitations

Strengths: scale, breadth of domains, causal identification, the expert-survey benchmark, thorough mechanism testing, and a genuinely surprising headline. Limitations: (1) the authors themselves caution the tasks may not capture identity-laden, ego-relevant, or emotionally charged domains where sycophancy could dominate—precisely the settings that motivate public concern (e.g., mental health, politics); (2) short, single-session interactions cannot capture cumulative/longitudinal echo-chamber dynamics; (3) generalizability beyond incentivized economic tasks and a US online sample is asserted more than demonstrated; (4) findings hinge on current model calibration, which the paper acknowledges could shift. The scope condition ("the former set is large") is honest but leaves the most feared cases unresolved.

Other Observations

The refutation dimension is notable: the paper explicitly overturns a documented consensus prior. The dataset of 10,000+ coded conversations and the multi-model sycophancy benchmark are reusable assets. Reproducibility is aided by shown prompts, specified models, and pre-registration, though a code/data release is not explicitly confirmed. Resource requirements (large participant pool, extensive API usage, RA validation) place this in the "well-funded group" range but remain within reach of a strong academic lab.

Overall, this is a high-quality, timely, and likely-influential contribution whose main caveat is the boundary-condition question it openly flags for future work.