Skip to content

Verify AI's Facts and Citations

When to use this

AI gave you a specific fact, statistic, quote, or source. You're about to actually use it — in a document, a decision, or something you'll repeat to someone else. Not needed for pure brainstorming or a first-pass explanation you already know you'll double-check later.

The prompt

Before I use anything you just told me, help me verify it.

For each specific fact, statistic, or claim in your last answer, tell me separately:
1. The exact claim.
2. Whether you have a specific, real source for it (name it, and link it if you can) — or whether this is your own synthesis or inference rather than something a source directly states.
3. Your honest confidence: well-established, actively debated, or something you're not fully sure about.

If you don't have a real primary source for something, say so plainly instead of presenting it with the same confidence as something you do.

This surfaces what the AI is willing to flag as shaky — it doesn't reliably catch its own confident fabrications. A claim labeled "well-established" or "sourced" still needs the source opened, not less checking.

Before and after

Before: You ask an AI tool for the typical price range for a service you're about to quote a client. It gives you a confident-sounding number with no source attached.

After (excerpt):

Claim: "typical range is $8,000-$15,000." Source: no specific report backs this figure — it's my own synthesis from general patterns, not a cited statistic. Confidence: moderate at best. Treat this as a rough starting point, not something to quote as fact.

Now you know to check the number independently before it goes in front of a client. You won't repeat an estimate as if it were sourced.

Make it yours

  • If the answer is long, ask it to list only the claims it's least confident about — that's where your verification time is best spent.
  • If the AI names a source, actually open it. Don't count a plausible-sounding title as verified — open the link and check it actually says what the AI claimed. This is the one step that touches real ground truth; the rest of this page is just routing your attention toward it faster.
  • Some tools now do part of this for you. Gemini's "Double-check response" feature cross-references its own answer against Google Search results. It highlights where they agree or disagree, so you can see at a glance which parts to trust more.

Watch out for

Grounding an AI's answer in real search or retrieval helps a lot, but it changes the odds. It doesn't eliminate the problem. One independent study looked at deep-research tools built specifically for citation-backed research. Even those tools still generate a real share of hallucinated citations: "We find that 3-13% of citation URLs are hallucinated. They have no record in the Wayback Machine and likely never existed" (Rao, Wong & Callison-Burch, 2026).

Grounding is also only as good as what it's grounded on. A Stanford study of AI legal-research tools found retrieval-based tools cut hallucination sharply compared to general-purpose models. Even so, one tool "hallucinated more than 34% of the time" (Stanford RegLab/HAI, 2024). Researchers stopped short of calling any tool tested "hallucination-free." Treat "the AI searched for this" as meaningfully better than "the AI just generated this from memory" — never as a guarantee.

Sources

  • Rao, Wong, Callison-Burch (University of Pennsylvania) — Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents, 2026. We find that 3-13% of citation URLs are hallucinated. They have no record in the Wayback Machine and likely never existed. View source · verified 2026-07-05 · primary
  • Stanford RegLab / Stanford HAI — AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries, 2024. the Lexis+ AI and Ask Practical Law AI systems produced incorrect information more than 17% of the time. Westlaw's AI-Assisted Research hallucinated more than 34% of the time. They hallucinated between 58% and 82% of the time on legal queries. View source · verified 2026-07-05 · primary