One line that cuts down made-up facts: "mark [UNSURE]"
AI makes up facts with confidence. You can't stop it completely, but you can make it show where it is unsure, and that is where you check. Add this line to any task that involves numbers, names or laws.
Hallucination (when the AI gives you names, numbers or "studies" that don't exist) is part of how it works, not a bug. But you can make it honest about how confident it is.
Add this to the end of any prompt that involves facts:
For every fact, number, name or reference in your answer: if you are not 100% sure, mark it with [UNSURE] and tell me what I should check. It is better to mark too many than to be confident but wrong.
You will see a pattern: the AI is confident about general concepts, and unsure about specific numbers, dates, and anything that changed after it was trained. The marked spots are where you open Google or an official source.
Try it now: ask something factual (prices, laws, statistics) with this line. Count the [UNSURE] tags that come up.
Terms: Hallucination and Grounding.