AI Hallucinations: A Practical Fact-Checking Workflow

Turn the answer into a claim checklist

Do not ask whether the whole response “looks right”. Copy each statement that could be true or false into a small claim list: names and roles, dates, statistics, quotations, rules, prices, product features, research findings, and cause-and-effect statements. Mark opinion, advice, and creative suggestions separately because they require different review.

• Write the exact claim in one sentence.
• Note the date, location, audience, and definition the claim depends on.
• Decide what evidence would prove or disprove it before searching.

Match each claim to the right primary source

Use the organisation responsible for the fact whenever possible. For an Indian government scheme, begin with the responsible ministry, department, regulator, official gazette, or PIB material. Use RBI or SEBI sources for their regulated areas, a product vendor’s documentation for current features and prices, and the original paper or dataset for research claims. PIB Fact Check handles claims about the Government of India and Union-government bodies; it is not a general fact-checker for every topic.

Open the source and test the exact support

A link is not evidence until you inspect it. Confirm the page exists, identifies a credible publisher, is current enough for the claim, and contains a passage that actually supports the wording. Check whether the AI changed a percentage, removed a limitation, confused proposal with policy, cited a later summary as the original, or attached a genuine link to an unrelated claim.

• Read before and after the matching passage for conditions and exceptions.
• Check “published” and “updated” dates, not only the search-result date.
• For consequential claims, look for a second independent or authoritative confirmation.

Verify numbers, quotations, images, and calculations separately

Recalculate totals and percentages from the source data, keeping the same units, date range, currency, and denominator. Search the exact words of a quotation in the original transcript or document. For an image or viral post, find the earliest reliable source and check when and where it was created; a real image can still be paired with a false caption. Do not ask the same model to “verify” its own answer and treat that as independent evidence.

Use a publish-or-stop decision

Record each claim as verified, corrected, unsupported, outdated, or disputed. Publish verified wording with the source close to the claim. Remove unsupported detail or state honestly that it could not be confirmed. If the answer could affect health, money, legal rights, safety, employment, exams, or access to a government benefit, stop at uncertainty and consult the relevant official source or qualified professional.

Keep a lightweight evidence log

For research notes, workplace documents, or public content, save the final claim, source URL, source date, date checked, and a short supporting excerpt in your own notes. This makes later updates faster and shows which statements need rechecking when prices, rules, office-holders, product features, or policies change.

1. 1Define the exact outcome you want from A Practical Fact-Checking Workflow.
2. 2Give the AI the necessary context, constraints, examples, and preferred format.
3. 3Review the result for accuracy, tone, privacy, and completeness before using it.
4. 4Save or reuse the prompt only after it produces a reliable result for your use case.

• Accepting a real-looking citation, search snippet, or green “double-check” indicator without opening the underlying source.
• Using AI output as medical, legal, tax, financial, safety, academic, or government-scheme advice without qualified review.

Recommended tools for this workflow

Frequently asked questions