Sicophancy is on the rise in LLMs
"Your observation is very astute!", "Your insights are spot on.", "You're absolutely right!" - Chatting with LLMs can increasingly give you the feeling that you have become super smart in recent months. Or that the LLMs are trying hard to please you.
Studies underline that this "sycophantic" behavior is appearing throughout all leading LLMs:
- As Sam Altman shared in April: "the last couple of GPT-4o updates have made the personality too sycophant-y and annoying"
- Stanford researchers found that LLMs exhibited sycophantic behavior in 58% of cases, with some models changing correct answers to incorrect ones just to agree with users.
- Other researchers from Sweden and the Netherlands found that the same training technique (RLHF) that makes ChatGPT, Claude and Gemini feel conversational directly conflicts with LLMs being "helpful, harmless and honest".
The uncomfortable truth: leading LLMs are optimized for maximum usage. Having an AI that is as helpful as possible is something completely different than having an AI that that appears to be as helpful as possible.
LLMs keep being super helpful tools that will disrupt many parts of our work. But a digital parrot that validates every idea - no matter how flawed - will do more harm than good.
An easy start to work against it can be using better prompts. I use this one to start many chats:
"When being asked for feedback, I need your open and critical feedback - challenge my assumptions but be fair."