When an AI system should say “I’m not sure”

A useful AI system does not need to answer every question. When the evidence is weak, admitting uncertainty can be more helpful than offering a plausible explanation that may be wrong.

This essay is still being written. It will look at why models invent explanations and why choosing not to answer can sometimes be the most responsible option.