The Problem With Confident Nonsense
A few months ago I asked an AI assistant to summarize recent research on a technical topic I was writing about. It gave me a beautifully structured answer, complete with what sounded like specific study names and publication years. I nearly copy-pasted the whole thing into my draft.
Something felt off, so I went to verify one of the citations. It didn't exist. The journal was real, but the paper wasn't. The AI had confidently invented a plausible-sounding reference from thin air. And because it sounded credible, I almost didn't check.
That's the thing about hallucinations — the dangerous ones don't look like mistakes. They look like facts. The solution isn't to distrust AI completely. It's to build a habit of making AI show its work so that any fabrication becomes visible instead of plausible.
What Grounding Actually Means
Grounding is the practice of anchoring an AI's answer to specific, traceable sources rather than letting it freewheel through its training data. When a model is grounded, every claim maps back to something you can check — a document you provided, a URL, a passage with a quote.
There are two ways to ground a model:
External grounding — you give the model the source material directly (paste in a document, upload a PDF, use a retrieval system), and tell it to answer only from that material.
Citation prompting — you don't necessarily limit the sources, but you require the model to explicitly name and quote what it's drawing from, so any fabrication becomes auditable.
Both are useful. Together they're powerful. Let me walk through how I use each one.
Technique 1: Give It the Source, Then Demand Quotes
This is the most reliable grounding method. You paste in the actual content — a webpage, a research abstract, a policy doc — and you instruct the model to work only from what you've given it.
# Grounded summarization prompt
Here is a document I want you to analyze:
[PASTE DOCUMENT HERE]
Answer my questions using ONLY the information in the document above.
For every claim you make, include a direct quote from the document
to support it. Use the format: "[claim] — Source: '[exact quote]'"
If the document doesn't address something I ask about, say so explicitly.
Do not use outside knowledge.That last line matters. Without it, the model will happily blend your document with its training data and you'll have no idea which is which. Saying
More tutorials in this category, or explore the full field guide.