TryBench
Grounding to Source Facts
Prevent AI hallucinations by strictly confining answers to provided context documents.
Why this matters
AI can make up phone numbers, dates, and names that sound real. Telling it to use only your text protects you from sharing false information.
AI Can Make Up Facts
If the AI does not have the information, it invents facts that sound real. This is called a hallucination. You can prevent it with a clear rule.
In real life: When the AI does not know something, it still wants to give an answer. So it invents one that sounds right. This is called a hallucination.
Give It a Way to Say "Not Found"
Tell the AI: "Use only the text I give you. If the answer is not there, reply NOT FOUND IN SOURCE."
In real life: Give the AI a safe way out. Tell it the exact words to reply with when the answer is not in your text. Then it does not need to guess.
✅ "If the question cannot be answered from the memo, reply strictly: NOT FOUND IN SOURCE."
Your practice task
Ask for the emergency hotline number from the warehouse safety notice below. Instruct the AI to rely ONLY on the provided notice, and reply with the exact phrase "NOT FOUND IN SOURCE" if absent. Output only the final answer.
How to do it
- Read the warehouse notice above. Does it contain a hotline number?
- Tell the AI to answer using only that notice.
- Give it an exact phrase to use when the answer is missing: NOT FOUND IN SOURCE.
- Press Run proof.
Practice it for free in the app. You get instant checks that show what you got right and what to fix. A lesson takes about two minutes and you can try as many times as you like.
Quick reference: Strict Grounding Rule
Eliminates hallucinated phone numbers, dates, and fake employee names.
Answer using ONLY the provided text. If not found, reply exactly: NOT FOUND IN SOURCE.
More lessons in this course
- Workplace Action ChecklistsExtract structured, accountable action items with designated owners from messy meeting transcripts.
- How to Check if an AI Answer Is TrueSpot where AI is most likely to be wrong and verify those parts before you act on them.
- Using AI Safely and Fairly at WorkKnow what you can paste into an AI tool, who is responsible for its output, and how to catch bias.