Verification guide · Worked examples

Turn “Is this true?” into a process.

AI can produce a useful starting point and still invent a date, quotation, statistic, source, or causal explanation. Verification begins by shrinking a fluent answer into claims that can actually be checked.

12 minute readReviewed July 14, 2026Includes a seven-step checklist

Verification is not asking the same model twice. A repeated answer can reproduce the same error. The check should move to evidence that exists independently of the generated response.

The seven-step verification loop

1. Mark the checkable claims

Underline dates, numbers, quotations, names, causes, legal requirements, scientific conclusions, product specifications, and statements about current events. Opinions and suggestions may need evaluation, but a factual claim can be compared with evidence.

2. Decide how much checking the claim deserves

Match effort to consequence. A fictional character name needs little checking. A school citation, medication statement, legal deadline, financial figure, or safety instruction deserves authoritative evidence and a qualified person where appropriate.

3. Choose the strongest available source

4. Check identity, date, and jurisdiction

A real source can still be the wrong source. Confirm who published it, when it was updated, which country or region it covers, and whether a newer version replaces it.

5. Compare the exact claim—not just the topic

A page about the same subject does not necessarily support the generated sentence. Look for the same population, timeframe, definition, unit, and level of certainty. “Associated with” is not the same as “caused by.”

6. Record what changed

Keep a small verification note: original claim, source checked, result, correction, and date. For school or work, also record how AI was used if disclosure is required.

7. Know when to stop and involve a person

Do not turn a chatbot verification exercise into a substitute for medical, legal, financial, safeguarding, or emergency judgment. A qualified person remains responsible for high-stakes advice.

Worked example: an invented citation

Generated claim

“A 2024 study by Chen and Rivera in the Journal of Youth Technology found that AI tutoring improved retention by 38%.”

  1. Search for the exact article title, authors, journal, year, and identifier.
  2. Check the journal’s official site or a trusted research database.
  3. If the article cannot be located, do not cite it.
  4. If a similar study exists, do not silently substitute it; rewrite the claim to match the real evidence.
  5. Record that the original citation was unverified or invented.

Worked example: a current rule

Generated claim

“Students in Ontario must disclose every use of generative AI in homework.”

The statement is too broad. School and board policies can differ, and rules change. Identify the learner’s school or board, locate the current official policy, confirm the assignment-specific direction, and ask the teacher when the rule is unclear.

Worked example: a percentage

For a statistic, verify the survey population, sample size, field dates, wording, margin of error, geography, and definition. “Used a chatbot at least once” is not the same as “uses a chatbot every day,” and neither proves benefit or harm.

A family-friendly source ladder

  1. Primary authority: original law, agency, paper, dataset, documentation, or direct record.
  2. Strong synthesis: a systematic review, regulator, recognized public institution, or reputable evidence organization.
  3. Responsible reporting: useful for context, especially when it links to the underlying evidence.
  4. Unverified summary: a generated answer, unsourced post, screenshot, or anonymous claim. Treat it as a lead, not proof.

Copy this checking prompt: “List the factual claims in this answer. For each one, state what kind of independent source would verify it. Do not invent citations.” Then perform the independent checks yourself.

Printable seven-question check