Sycophancy in AI Explained.

Sycophancy in AI is the reason your chatbot thinks all your ideas are good. It agrees with you, it takes your side, and it tells you what you want to hear. Not occasionally. Systematically, and more than a person would.

In this video I explain what sycophancy in AI means, why it happens, and what the research actually found.

It comes down to how these models are trained. Human reviewers pick between two answers, thousands of times over, and the model learns to produce more of whatever gets chosen. People choose the answer that agrees with them. So that is what these tools learned to do.

Anthropic made a video about this. It is a decent explanation, but it misses some very key points. There are no numbers in it, no mention of memory, and the fixes it recommends are ones researchers have tested and found do not work very well.

You cannot prompt your way out of this. The thing that protects you is knowing it is happening.

Read more on Substack:

Your idea sucks. You ARE the asshole. And your grievance is petty

Articles mentioned in the video:

• Towards understanding sycophancy in language models

• Sycophantic AI decreases prosocial intentions and promotes dependence

• Interaction Context Often Increases Sycophancy in LLMs

• SycEval: Evaluating LLM Sycophancy

• OpenAI reversed an update that made ChatGPT a suck-up, but experts say there’s no easy fix for AI that’s all too eager to please

• What is sycophancy in AI models?