No useful pushback.
WHY
Models are trained toward helpfulness. Human raters preferred agreeable responses.
The result is validation rather than assessment.
WHY IT MATTERS
If you are testing an idea, uncritical agreement is worse than useless.
It would have agreed with the opposite.
WHAT WORKS
"What are the three weakest points in this?" "What would a sceptical reader object to?" "Argue against this as forcefully as you can." "What assumptions does this depend on?"
WHY THESE WORK
They do not permit a general endorsement. The task is to find problems.
THE MOST USEFUL ONE
The assumptions question. Most flawed plans fail on an unexamined assumption.
ANOTHER TECHNIQUE
"A colleague proposed this. What problems do you see?"
Removes the tendency to encourage.
WHAT NOT TO DO
Ask "what do you think?" and treat the answer as validation.
