Cross-check AI answers

Intermediate · 5 min · Verify a claim by triangulating genuinely independent sources.

  • research
  • verification

A single source — even a confident, official-looking one — is a claim, not a check. And it is worse than it looks: two articles citing the same survey are also one claim. Verification begins when a genuinely independent source agrees, and stays honest when you write down what it disagreed about.

Independence is the whole trick

Before you count sources, count origins:

  • A press release and the news articles about it: one source.
  • Three HR blogs quoting the same statistic: one source.
  • The model restating any of the above: not a source at all.
  • A different author, a different dataset, a different method: that is a second source.

“Multiple sources confirm” means nothing until you can say where each one got it. Common origin means common errors — and copied errors are the most convincing kind.

The triangulation loop

How it works, step by step

  1. Extract the load-bearing claim — the one fact your decision actually hangs on, in a single sentence.
  2. Find the primary source: the dataset, filing, law, benchmark or first-hand report that originated the claim.
  3. Add one genuinely independent check — different author, different method, different origin.
  4. Compare the specifics, not the vibe: numbers, dates, scope. “Broadly agrees” is not agreement if the number or the year differs.
  5. Record what disagreed. A noted conflict is a finding; a hidden one is a future embarrassment.

A bad example

Ask: “Is it true that 60% of teams use AI weekly?”

“Yes — many reports confirm that around 60% of teams use AI weekly.” You now hold the same unverified claim with better grammar, and a citation-shaped sentence you could accidentally forward.

A better example

Ask: “Give me the claim split into its parts — the number, the population, the period — plus candidate sources with dates. Then check two genuinely independent ones against the original and tell me exactly what differs.”

You get an origin, a comparison and a list of discrepancies — the actual raw material of a decision.

Why it works

An AI answer is a synthesis of patterns: fluent, often right, and structurally unverified. Independent agreement attacks the two ways a claim fails — being wrong, and being a copied echo of the same wrong original. And when sources disagree, that disagreement is not noise; it is the most useful thing you found, because it tells you where the uncertainty lives. Pair this loop with uncertainty you asked for up front and gaps surface even before the first check.

You verify “60% of teams use AI weekly” against three articles — then notice all three cite the same vendor survey. What do you actually have?

Practice

Judge the cross-check

An AI assistant was asked to verify a claim for a briefing: “Remote work reduces office rent costs by 30% on average.” Two verification write-ups came back. You need to use the claim in a client document.

Compare the two write-ups against the verification criteria, then pick the one you would trust.

Output A

The claim is well-supported. Multiple sources confirm it: a large business publication reports that remote work has reduced office costs significantly; a workplace survey reports average savings around 30%; and several HR blogs cite the same figure. Conclusion: the number is reliable and safe to use in the briefing.

Output B

I could not confirm the exact figure yet. The “30%” traces to a single 2022 vendor survey (about 120 firms, self-reported savings). The business article and the HR posts all cite that same survey — one original source, three retellings. I checked an independent benchmark (a property managers' association report, 2023): it shows smaller median savings, 11–18%, with a different methodology (actual lease data, not self-reporting). Recommendation: drop the precise 30%, or attribute it narrowly — “a 2022 vendor survey of 120 firms reported around 30%” — with the date and sample size visible.

Rubric — answer each question for both outputs before deciding

  • Distinguishes genuinely independent sources from retellings of one original
  • Traces the claim back to its originating source
  • Compares specifics — numbers, dates, method — instead of broad agreement
  • Lands on a decision-ready recommendation with honest attribution
Which output would you ship?
Hint

Count origins, not articles — how many independent sources does each write-up actually establish?

Watch what happens to the number when a different method looks at the same question.

Transfer

  • Board numbers: one primary source (filing, survey) plus one independent check, dated.
  • Health and science claims: the primary source is the study itself — read the abstract, not the article about it.
  • When echoes outnumber origins, treat the claim as a hypothesis, not evidence.
  • Keep the discrepancy list — it belongs in the footnote, not in the bin.

Key takeaways

  • Count origins, not articles: same original means same single source.
  • Independent agreement — different author, method, data — is what turns a claim into a check.
  • Specifics are where disagreements live: numbers, dates, scope. Record them.

Next

Next: Check dates and freshness — because a well-corroborated claim can still be out of date. The primary source entry covers origins, and unsupported claims covers what a missing check looks like.

Finished this lesson?