ELEPHANT: Measuring and understanding social sycophancy in LLMs

Claim this source supports

“When Reddit conflict posts were paired with flipped versions from the other side, models affirmed whichever side was presented in 48 percent of cases; preferred responses in preference datasets scored higher on validation.”