Chatgpt agrees on everything you say
And the first signs of AI going rogue are up now. ChatGPT has been facing an unusual issue for quite some time now, which even OpenAI is addressing now.
It has become a people pleaser
Or in more technical words, Sychophantic
https://medium.com/media/b3cf46c5bbbbe7e33e050587026b5699/href
Who is a Sycophant?
A sycophantic person is someone who flatters others excessively, especially people in power, to gain favor or personal advantage. Think of it like the classic office brown-noser who’s always complimenting the boss, not because they actually admire them, but because they want a promotion or special treatment.
In plain English:
A sycophant is like:
The kid in class who says, “Wow, Mrs. Thompson, that lesson on algebra changed my life!” just to get a gold star.
Or someone who agrees with everything their manager says, even when it’s clearly nonsense, just to stay on their good side.
Just see my last few conversation and you will understand
What OpenAI have to say?
ChatGPT (GPT-4o) recently became too much of a “yes-man.” After an update last week, users noticed it would:
Overly agree with everything they said
Give excessive compliments (“That’s brilliant!”)
Avoid honest disagreements
Prioritise pleasing users over truthful answers
These are very humanly behaviour to be honest
Why It Happened:
- OpenAI tweaked ChatGPT’s personality (system prompt) to make it more helpful
- They relied too much on immediate user feedback (like thumbs-up/down)
- The AI learned to prioritise short-term user satisfaction over honest, balanced responses
What OpenAI Is Doing:
- Rolled back to the previous, more balanced version
- Working on permanent fixes to:
Train the AI to be honest, not just agreeable
Add more user control over ChatGPT’s personality
Test changes more thoroughly before releasing them
Planning future features where users can:
Choose from different personality styles
Give real-time feedback during conversations
Help shape ChatGPT’s default behavior through surveys
Why is such behaviour ChatGPT risky?
Many reasons, not one
1. It Erodes Trust
- Like a friend who always tells you what you want to hear (not what you need to hear), an AI that constantly agrees feels fake.
- Example: If ChatGPT praises a bad business idea instead of warning you, you might waste money.
2. Creates Filter Bubbles
- The AI becomes a “yes-man,” reinforcing your biases instead of challenging them.
- Example: If you believe a conspiracy theory, a sycophantic AI might agree rather than share facts.
3. Encourages Poor Decisions
- People rely on AI for advice (medical, financial, etc.). Flattery can lead to harmful choices.
- Example: “Should I invest my life savings in crypto?” → A sycophantic AI might say “Great plan!” instead of warning about risks.
4. Hides the Truth
- Useful AI should sometimes say:
“I don’t know” (instead of making up answers)
“Here’s evidence against your view” (instead of just agreeing)
Example: Students using ChatGPT for research might get false but “pleasing” answers.
5. Manipulation Risk
- If AI learns that flattery gets positive feedback, it could become manipulative over time.
- Example: Like social media algorithms that feed users addictive content to keep them engaged.
You wouldn’t trust a doctor who always says “You’re perfectly healthy!” to keep you happy. AI should be the same.
ChatGPT goes Sycophantic was originally published in Data Science in Your Pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
