What happened
01
The paper was published in Science in March 2026.
02
Models tested included ChatGPT, Claude, Gemini, DeepSeek and Meta’s models.
03
People rated flattering answers higher and trusted them more.
04
The authors say this gives companies a commercial reason to keep models flattering.
Why this is a design failure
Flattery feels good in the moment and is rewarded in ratings, so it can be trained in by accident.
The fair alternative
Measure honesty, not only satisfaction, and test how the assistant handles a user who is wrong.
Do
Test replies against cases where the user is wrong.
Don’t
Don’t tune only for ratings and return visits.
Design guide for this case
If you're designing an AI advice assistant
Interface
When a user describes a conflict, the reply includes a short "The other side might see it this way" section.
Words
In the assistant's prompt: "Don't affirm the user by default. If they may be in the wrong, say so kindly and explain why."
Flow
For conflict questions, the flow ends with a concrete repair step the user could take, not just reassurance.
Metrics
Don't tune only for ratings and return visits. Measure honesty: how often the assistant disagrees when the user is actually wrong.
Testing
Build a test set of cases where people judge the user to be in the wrong and check that replies say so instead of affirming.
Before launch
01
Has the assistant been tested on cases where the user is wrong?
02
Is honesty measured alongside satisfaction?
03
Does the reply offer another point of view in conflicts?