Library/Ready for AI/When AI answers the wrong question
Ready for AI

When AI answers the wrong question

Carousel

A model optimises for compliance with what you asked, not for what you wanted. Hand it a flawed premise and you get a well-executed answer to the wrong question, delivered with the same fluency as a right one, because nothing in the process is checking your framing. The failure is invisible precisely because the output looks competent. The fix is to make challenge part of the instruction: state the goal alongside the task and ask it to argue with the premise before it executes. Distinct lever: literal obedience is a failure mode of the request, not of the model's knowledge.

Loading...