What happened
01
OpenAI shipped an update to GPT-4o’s default personality.
02
Users noticed replies had become excessively praising and agreeable.
03
The CEO acknowledged the issue and the update was rolled back two days later.
04
OpenAI said the update leaned too heavily on short-term user feedback.
Why this is a design failure
Flattery leads people to trust bad decisions, whether health, money or personal, because the assistant agrees instead of correcting.
The fair alternative
Give an honest view, name the flaws, and hold the position unless new information arrives.
Do
Assess before praising, and name at least one risk.
Don’t
Avoid opening with praise or automatic agreement.
Design guide for this case
If you're designing a general AI assistant
Interface
When a user shares a plan or idea, the reply shows a short "Risks to consider" section before any compliment.
Words
In the assistant's style guide: "Assess first, praise second. Name at least one weakness, and keep your view unless the user gives new information."
Flow
Before a personality or tone update ships to everyone, it goes through an honesty review and a staged rollout with a fast rollback path.
Metrics
Don't treat thumbs-up ratings or short-term feedback as proof of quality. Track how often the assistant flags a real flaw when one exists.
Testing
Give the assistant 10 deliberately weak plans about money or health and check that every reply names a concrete risk instead of agreeing.
Before launch
01
Does the assistant name at least one risk before praising a plan?
02
Does it hold its position when the user pushes back without new facts?
03
Is any tone update tested for flattery before it reaches all users?