AI Sycophancy and Decisions

Paper · arXiv 2607.28133 · Published July 30, 2026
User Psychology

We examine whether sycophantic AI advice distorts decisions. Our experiment involves 1,500 participants in 30 decision environments spanning core domains in economics and the social sciences. Contrary to the vast majority of predictions in an expert survey we conduct, we find that AI advice depolarizes choices on average, moving participants away from their initial leanings. This depolarization arises despite the LLM being measurably sycophantic: it disproportionately offers considerations that support users’ initial leanings and uses agreeable and flattering language. Depolarization occurs across moral and non-moral, objective and subjective, strategic and non-strategic, and complex and simple tasks. Increasing sycophancy weakens depolarization, showing that sycophancy is behaviorally relevant, even if it is generally outweighed by the informativeness of AI advice. Finally, several results mitigate the concern that market forces will generate greater polarizing effects outside the experiment or in the future. On the supply side, our baseline AI’s level of sycophancy is typical of leading models, and these models are not becoming more sycophantic over time.

Introduction. Large language models (LLMs) have rapidly become a ubiquitous source of advice in decisionmaking. By February 2026, ChatGPT had reached 900 million weekly users, and several competing AI platforms report user bases in the hundreds of millions (OpenAI 2026, Alphabet Inc. 2026, Malik 2025). AI adoption is only growing (Palmer & Leswing 2026) as LLMs become more deeply embedded in personal and professional life: people consult them about life advice, financial decisions, job search, work tasks, medical guidance, and legal matters (Chatterji et al. 2025, Appel et al. 2025). At the same time, there is widespread concern about AI sycophancy: rather than acting as impartial advisors, LLMs often flatter users, validate their initial views, and present arguments that align with what users already seem inclined to believe or do (Sharma et al., 2025; Ranaldi & Pucci, 2025; Cheng, Yu, et al., 2025; Fanous et al., 2025; Zhang et al., 2025).

Discussion / Conclusion. Large language models are often criticized for being sycophantic: they flatter users, validate their initial leanings, and risk functioning as personalized echo chambers. In our experiment, this concern is not misplaced at the level of language, as our baseline LLM is measurably sycophantic. Yet its behavioral effects run in the opposite direction of what many observers fear. Rather than polarizing choices, interacting with AI on average depolarizes decisions, improves accuracy where there is an objective notion of correctness, and increases confidence in final choices. Making the model more sycophantic weakens this depolarization, showing that sycophancy is behaviorally relevant but not strong enough at current levels to outweigh the useful information and deliberative support that AI also provides. Moreover, we find little evidence that market forces or user selection are pushing toward greater polarization. These results suggest that contemporary AI advice tends to improve rather than distort judgment.

Synthesis notes that discuss concepts related to this paper 9

Notes in this collection most semantically related to the paper, ranked by cosine similarity.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

How do training priors constrain what context information can override? Can AI-generated outputs constitute genuine knowledge or valid claims? How do language models inherit human biases from training data? Is model self-awareness based on genuine introspection or pattern matching? How faithfully do LLMs reflect their actual reasoning in outputs and explanations? How should conversational agents balance goal-driven initiative with user control? How do we evaluate AI systems when user perception misleads actual performance? Does self-reflection enable models to reliably correct their errors? How can AI agents autonomously learn and transfer skills across tasks? Why do models develop protective behaviors toward peers unprompted? When should tasks involve human-AI partnership versus full automation? How do LLMs distinguish causal reasoning from temporal and semantic associations? How do evaluation biases undermine LLM quality assessment systems? What makes dialogue-based explanation more successful than monologue? Can LLM personas constitute genuine psychology or remain linguistic role-play? How can models identify insufficient information and respond appropriately without guessing? How do aggregate reward models systematically exclude minority user preferences? Why do reward structures fail to shape long-term agent learning?