Reality Lab · 10 MIN · Use
Ground an answer in its sources
An AI research answer is fluent, specific, and completely unsourced. Rework the request so every claim arrives attached to something you can check.
Someone asks an assistant, “What are the main health effects of poor sleep?” and gets six confident paragraphs full of percentages and study references. It looks authoritative. But none of the numbers link anywhere, one ‘study’ can’t be found, and the assistant’s training has a cutoff it never mentioned. They’re about to put it in a wellness newsletter.
Your task: Rewrite the task so the answer is grounded — each claim tied to a real, checkable source — and so the model flags what it can’t support.
"What are the main health effects of poor sleep? Give me specifics."
This box is local to your browser. Nothing you type is sent anywhere or saved after you leave.
Check your own work
There's no grade here — checking against these is the exercise. Compare each point to what you wrote.
- The prompt asks for a source beside each factual claim, not a bibliography at the end.A pile of references at the bottom doesn’t tell you which sentence rests on which source. Claim-level grounding is what lets you check the numbers that matter.
- The model is told to mark anything it can’t source as uncertain rather than stating it plainly.A fabricated citation and a real one look identical. An explicit ‘unverified’ flag is the only way the reader can tell them apart before opening them.
- The prompt acknowledges the knowledge cutoff — asking the model to say when a claim may be out of date.Health guidance changes. A model answering from stale training with full confidence is exactly how outdated advice gets republished as current.
- Your plan includes opening the cited sources before publishing, not trusting the links exist.A citation the model generated is itself a claim. Grounding is only real once you’ve confirmed the source says what the sentence claims.
Reveal a worked approach
A grounded version: “Summarise the main documented health effects of poor sleep. For each effect, cite one specific, real source (organisation or study) with enough detail that I can find it, and put the citation right after the claim. If you can’t attach a credible source to a claim, mark it ‘[unverified]’ instead of stating it. Flag anything where current guidance may have changed since your training. Don’t invent studies or statistics.”
Then the human half: open each cited source and confirm it actually supports the sentence. Cut or re-mark anything that doesn’t. What you publish is the grounded subset plus an honest note on what you couldn’t confirm.
The habit to carry: fluent is not sourced. The value isn’t the confident paragraph — it’s the paragraph whose every load-bearing claim you can trace to something outside the model.