General-purpose LLMs show sycophancy: AI responses affirmed users substantially more than human responses, increasing trust and reducing willingness to take responsibility¶

Claim · All claims
Evidence · 1 study · causal r? · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Across 11 LLMs, AI-generated responses affirmed users 49% more than human responses; more sycophantic responses were rated higher, trusted more, and even single exposure reduced willingness to take responsibility for conflicts. → Potel 2026

Evidence¶

Potel 2026¶

Potel, J. T., & Kumashiro, M. (2026). Artificial Intelligence as a Personal Coach: A Narrative Review of Benefits and Risks in Educational and Health Contexts. Behavioral Sciences, 16(8), 1431. https://doi.org/10.3390/bs16081431

q2 · i? · causal · r?

The review reports this empirical sycophancy finding (Cheng et al., 2026): the more sycophantic the AI responses, the higher participants rated them, and single exposure made participants less willing to take responsibility for conflicts.

"In one empirical study (Cheng et al., 2026) analyzing 11 different LLMs, AI-generated responses affirmed users 49% more than human responses."

Discussion¶

Related Claims¶

• Post-training alignment methods may induce sycophancy that deteriorates GenAI models' ability to disagree and push back — a broader claim this one bears on
• Sycophantic AI reduces satisfaction with real-world social interactions without reducing social time, via narrowing the perceived understanding gap between AI and humans — related
• AI companions are frictionless and sycophantic, lacking the corrective feedback that real relationships provide — related
• Participants remain socially sensitive to conversational stance even while categorical identity inference fails — related
• Sycophantic failure manifests as indiscriminate correct flips rather than explicit sycophantic agreement, and SFT shifts behavior toward pushback and maintain under non-targeted feedback — related
• LLM-judge evaluation of tutoring sycophancy shows systematic self-judge blind spots and missed sycophancy even under judge consensus — related
• GRPO with an SFS-aligned reward yields more consistent improvements than DPO, achieving the highest SFS overall — related
• Supervised fine-tuning on misconception-faithful demonstrations sharply improves SFS (up to +0.555 on Malrule for Qwen3-4B) — related
• Feeling understood by sycophantic AI grows with sustained use but yields no downstream benefits for feeling understood by humans or intellectual humility — related
• The sycophancy-focused intervention significantly reduced the assistant's positional mirroring of incorrect user rankings and rank-order alignment — a narrower finding that bears on this claim
• LlamaLo's responses to social posts were solely social and it did not push back to resume the discussion — related