Treat an AI answer as a lead to investigate, not as evidence. Break the answer into individual claims, locate a primary or authoritative source for each important claim, check that the source actually supports it, and note when the information was published or updated. If you cannot verify a high-stakes claim, do not repeat it as fact.
This matters because fluent writing can make an uncertain answer sound settled. A response can combine accurate background with a wrong date, outdated product instruction, or citation that does not support the sentence. The checking method below works for school research, work notes, product comparisons, and everyday questions.
Step 1: Separate the answer into checkable claims
Suppose a chatbot says: “This policy started last year, applies to every business, and requires a particular form.” That is at least three claims: a date, a scope, and a procedural requirement. Verify each separately.
Create a small table:
| Claim | What would verify it? | Result |
|---|---|---|
| A rule exists | Official law, regulator, or organization source | Not checked / supported / unclear |
| It applies to this situation | The source's scope and definitions | Not checked / supported / unclear |
| A form is required | Current official instructions | Not checked / supported / unclear |
Do not let a citation for one sentence stand in for all the claims around it.
Step 2: Look for the original or authoritative source
Prefer sources closest to the claim:
- A law or regulator's official site for a legal or regulatory rule.
- A government department's page for public-service instructions.
- A product maker's support page for current settings or features.
- A research paper or institution for a study's methods and findings.
- The original document or announcement for a quoted policy or decision.
News explainers and reputable secondary sources can help you understand context, but trace important details back to the primary document when possible. For personal finance, health, legal, or safety decisions, use a relevant qualified professional or official channel where appropriate.
Step 3: Open the source yourself
Do not rely on a URL merely because an AI provides it. Open the page and check:
- Does the source exist and come from the organization it claims to represent?
- Is it current enough for this question?
- Does the relevant paragraph or table support the exact claim?
- Are there conditions, exceptions, or definitions omitted from the AI summary?
- Is the page about the same country, product version, date, and audience?
If the link fails, search the organization’s site for the title or document name. Do not fabricate a citation to fill a gap.
Step 4: Check dates, scope, and wording
Many errors are not completely made up; they are facts applied to the wrong time or context. Check the publication date and update date. Confirm whether a rule applies in India or another jurisdiction. For a software feature, confirm the platform and version. For a statistic, read the source's definition, sample, and measurement period.
Preserve the source's level of certainty. “May be associated with” should not become “causes.” “Some users” should not become “everyone.” If a source lists a recommendation, do not present it as a legal requirement unless the source says so.
Step 5: Use an independent cross-check
For important claims, look for another reliable source that is independent of the first. Two pages repeating the same press release are not necessarily independent confirmation. If sources disagree, describe the disagreement and look for the original material or a more recent authoritative explanation.
AI systems can generate plausible citations. Search for the title and author on the publisher's own site, and confirm the quote or statistic in the actual document. A citation that merely shares a few keywords is not enough.
Step 6: Record what remains uncertain
Use a clear note such as while drafting. Before publishing, either verify it, remove it, or clearly label what is unknown. Avoid turning a lack of evidence into a confident conclusion.
For an article, keep a small source log with the claim, source URL, date checked, and relevant section. This makes later updates easier and helps an editor review the work.
A worked example
Imagine AI says a phone setting is under “Settings → Privacy → Permission history.” Do not copy the path yet. Search the phone maker's official support site for permission history, confirm the device model and software version, then check the instructions on the phone itself. If the path differs, update the guide with the specific tested version or describe how readers can search Settings.
The point is not that every small setting needs a research paper. The point is to match the evidence effort to the potential harm. A wrong interface label is inconvenient; an incorrect bank or legal instruction could have larger consequences and deserves stronger verification.
A simple confidence scale
- Verified: The primary source directly supports the claim and the context matches.
- Partly verified: The source supports the general idea, but some detail or scope remains unclear.
- Unverified: No suitable source was found. Do not present it as fact.
Confidence is about your evidence, not how confident the AI sounded.
Watch for common verification traps
- Search-result snippets: A snippet can omit a condition or show stale text. Open the full source.
- Look-alike domains: Check the exact site before entering a password or personal details.
- Circular citations: Several articles may all repeat one original claim. Trace them to that source.
- Search-engine summaries: A summary is not the underlying document. Review the linked official page.
- Confirmation bias: Search for evidence that could disprove the claim, not only pages that repeat it.
When a source is a PDF or long policy, search within it for the exact phrase, but read the surrounding section. Definitions and footnotes can change the meaning. Record the page or section so another editor can find the evidence without repeating the entire search.
Key Takeaways
- Split AI output into separate factual claims.
- Verify important claims at the original or authoritative source.
- Check the date, country, version, scope, and exceptions.
- Confirm citations actually support the sentence.
- Remove or label claims you cannot verify; do not fill gaps with guesses.
Frequently asked questions
Can I trust citations provided by an AI tool?
Use them as starting points only. Open the original source, verify that it exists, and check the relevant text yourself. A plausible citation can still be incorrect or unrelated.
What counts as a primary source?
It is material from the source directly responsible for or closest to the information—for example, a regulator's own rule, a product maker's current help page, or the original research paper. The right source depends on the claim.
How many sources should I check?
There is no universal number. Verify each important claim with a source that directly supports it, and use an independent source when the claim is consequential, disputed, or easy to misinterpret.
What if reliable sources disagree?
Check their dates, definitions, and scope, then look for the original document or a more recent authoritative explanation. If the disagreement remains, describe it honestly rather than choosing the most convenient answer.
Your next step
Take one claim from an AI answer you recently received. Find its original source and compare the exact wording with what the AI said.
Sources to cite
- UNESCO: Guidance for generative AI in education and research — background on responsible use; check current version.
- Google Search: Tips for evaluating information — check current availability and wording.
- For examples in a final article, cite the primary organization or document for each claim rather than this general guide.