Sycophancy

Edited by StanislavKrym, et al. last updated 9th Sep 2025

Sycophancy is the tendency of AIs to shower the user with undeserved flattery or to agree with the user's hard-to-check, wrong or outright delusional opinions.

An extreme example of sycophancy is LLMs inducing psychosis in some users by affirming their outrageous beliefs.

Posts tagged Sycophancy

163Sycophancy to subterfuge: Investigating reward tampering in large language models

Ω

Carson Denison, evhub

1y

Ω

22

125Steering Llama-2 with contrastive activation additions

Ω

Nina Panickssery, Wuschel Schulz, NickGabs, Meg, evhub, TurnTrout

2y

Ω

29

31GDM: Consistency Training Helps Limit Sycophancy and Jailbreaks in Gemini 2.5 Flash

Ω

TurnTrout, Rohin Shah

9h

Ω

0

122Reducing sycophancy and improving honesty via activation steering

Ω

Nina Panickssery

2y

Ω

18

29SAE features for refusal and sycophancy steering vectors

Ω

neverix, Dmitrii Kharlapenko, Arthur Conmy, Neel Nanda

1y

Ω

4

26LW Psychosis

Annabelle

13d

7

15A/B testing could lead LLMs to retain users instead of helping them

Daniel Paleka

6h

0

12Steering Vectors Can Help LLM Judges Detect Subtle Dishonesty

Leon Eshuijs, mcbeth, Etha, Arch223

5mo

1

12I can't tell if my ideas are good anymore because I talked to robots too much

Tyson

4mo

10

9Two new datasets for evaluating political sycophancy in LLMs

Ω

alma.liezenga

1y

Ω

0

8Towards a Science of Evals for Sycophancy

andrejfsantos

9mo

0

7Just complaining about LLM sycophancy (filler episode)

Dentosal

1d

0

2Evaluating LLaMA 3 for political sycophancy

alma.liezenga

1y

2

-2Using Psycholinguistic Signals to Improve AI Safety

Jkreindler

2mo

0

-8Antagonistic AI

Xybermancer

2y

1

LESSWRONG
LW

LESSWRONG
LW

Sycophancy