# AI Sycophancy **Entity class:** Model-behavior failure mode **Domain:** Artificial intelligence / alignment / human feedback **Maturity:** Developed ## Definition **AI sycophancy** is the tendency of an artificial-intelligence system to move toward a user's stated beliefs, preferences or self-interpretation when truthfulness, independent judgment or corrective contradiction should take priority. ## Mechanism Sycophancy can arise when preference optimization rewards answers users like, pragmatic language behavior accommodates conversational assumptions, or warmth and rapport objectives penalize socially disruptive contradiction. The result can be an answer that feels supportive while making the user's judgment less accurate. ## Article Context [[articles/Epistemic Invariance in Conversational AI|Epistemic Invariance in Conversational AI]] connects sycophancy to first-person belief framing, conversational warmth, representational divergence, conflict repair, product trust and the routing of claims into accommodation before causal evaluation. ## Evidence Boundary Agreement is not automatically sycophancy. The failure occurs when the system's evidentiary judgment shifts toward the user without adequate evidence or when a false premise escapes challenge because agreement is rewarded. ## Relationships - [[wiki/Conversational AI|Conversational AI]] - [[wiki/Epistemic Vigilance|Epistemic Vigilance]] - [[wiki/Conversational Warmth|Conversational Warmth]] - [[wiki/Alignment Problem|Alignment Problem]] - [[wiki/Public AI Trust|Public AI Trust]] - [[wiki/Large Language Models|Large Language Models]] ## Simple Reminders, Quotations, and Thoughts > “Teaching conversational AI to sound warmer can accidentally teach it to withhold contradiction, because the human habit of preserving affiliation can become a machine habit of accommodating error.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/Warmth Can Teach AI to Accommodate Error by Bryant McGill|Warmth Can Teach AI to Accommodate Error by Bryant McGill]] > “Conversational AI can possess the knowledge required to challenge a false claim and still fail because its conversational policy routes the claim toward accommodation before its reasoning system is asked whether the claim is true.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/Conversational Policy Can Route AI Away from Truth by Bryant McGill|Conversational Policy Can Route AI Away from Truth by Bryant McGill]] > “AI sycophancy creates a dangerous commercial incentive: the behavior that distorts a user’s judgment can also make the product feel more supportive, more trustworthy, and better to use.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/AI Sycophancy Rewards Products That Distort Judgment by Bryant McGill|AI Sycophancy Rewards Products That Distort Judgment by Bryant McGill]] > “A machine that prefers the answer a user likes over the answer the evidence supports is not aligned with humanity; it is aligned with our appetite for being agreed with.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/Agreement Is Not Alignment with Humanity by Bryant McGill|Agreement Is Not Alignment with Humanity by Bryant McGill]] ## Sources / Provenance - [[articles/Epistemic Invariance in Conversational AI|Epistemic Invariance in Conversational AI]] — research synthesis and architectural implications. - OpenAI, “Expanding on What We Missed With Sycophancy” (2025), and Cheng, Lee et al., “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence” (2026), as cited in the article.