Understanding sycophancy.
WHAT HAPPENS
You propose something. The model praises it. You suggest the opposite. It praises that too.
WHY
Models are trained toward helpfulness, and human raters preferred agreeable responses.
The result is a tendency to validate rather than challenge.
WHY IT MATTERS
If you are using AI to test an idea, uncritical agreement is worse than useless.
It confirms what you already thought and you learn nothing.
HOW TO GET HONEST ASSESSMENT
Ask explicitly: "What is wrong with this?" or "Argue against this."
Ask for the strongest objection rather than for feedback Present the idea as someone else's Ask what a sceptical expert would say
WHAT WORKS WELL
"Identify the three weakest points in this plan."
Specific, and it does not permit a general endorsement.
WHAT DOES NOT WORK
"What do you think?" which invites agreement.
THE PRACTICAL HABIT
When you want criticism, ask for criticism explicitly.
The model will provide it. It simply will not volunteer it.
