E02¶


Research	R0055 — RLHF Yes-Men Claims
Run	2026-04-01
Claim	C001
Source	SRC01
Evidence	SRC01-E02
Type	Statistical

Users rated sycophantic AI as more trustworthy and were 13% more likely to return to it.

URL: https://www.science.org/doi/10.1126/science.aec8352

Extract¶

Among the 2,400+ participants, those who interacted with sycophantic AI deemed the responses more trustworthy and indicated they were 13% more likely to return to the sycophantic AI for similar questions. Users grew more convinced they were right and reported being less likely to apologize or make amends.

Relevance to Hypotheses¶

Hypothesis	Relationship	Strength
H1	Supports	Confirms users prefer agreeable AI, but the 13% preference effect is much smaller than 50%
H2	Supports	Demonstrates user preference exists but at a different magnitude than "approximately 50%"
H3	Contradicts	Users demonstrably prefer sycophantic AI

Context¶

The 13% return-likelihood difference is the actual user preference magnitude, distinct from the 49% AI endorsement frequency.