Verify an AI answer before turning it into a fact.
A structured answer can combine accurate facts, poorly connected references and overbroad conclusions. Check each claim that matters to the decision, even when a link or precise number is provided.
List names, figures, dates, quotations and recommendations separately. Rank points by the consequence of being wrong. A decorative detail should not consume the time needed for a decisive fact.
2
Test the reference
Open the original document through an independent route where necessary. Check its existence, title, author and whether the passage supports the claim.
3
Check context
Compare population, territory, unit, observation period and publication date. A report published this year may describe an earlier year.
4
Reproduce the calculation
Recalculate from the cited data and record the formula and rounding. Do not compare percentages with different denominators.
5
Rewrite the conclusion
Separate sourced confirmation, reasoned interpretation and unverified points. Remove references that do not support the sentence.
A citation must support the exact sentence
An existing article does not validate the summary attributed to it. Locate the passage, unit and caveats. If a link leads to a general page, find the precise publication; if it remains inaccessible, mark the check incomplete.
A second answer from the same system is not an independent source. Seek the primary document, then separate corroboration where a fact or attribution remains disputed.
Check freshness without inventing an update
A recent consultation does not update the source data. Record both dates and the publication’s coverage. For a role, price or rule that may change, check the responsible publisher directly.
If versions disagree, record the difference, date and context of each. Disagreement may reflect a correction, population difference or error; do not silently choose the convenient figure.
Prepare a shareable validation file
The log can contain claim, primary URL, relevant passage, period, finding and next action. Avoid sending secrets, complete personal records or client data to a tool whose use has not been authorised.
The NIST generative AI profile provides a risk-management framework. This IIS grid is an editorial checking method; it certifies no system and produces no universal reliability rate.
Check the scholarly object behind the citation
A DOI identifies an object without being a quality label. Compare title, authors, publisher and type: preprints, reviews and original studies need not support the same claim. Check publisher corrections, including Crossmark where present.
Check the conclusion’s unit
Separate unit, population, period and measured phenomenon. A share of messages is not a share of people; an average does not describe every case. Separate source observation from generated inference. Keep unsupported inference hypothetical even when the cited publication exists.
Apply the method
Test the evidence chain behind a generated answer
NIST’s generative-AI risk profile describes confidently presented false content, including misleading or fabricated citations. Open the original document and check the passage supporting each material claim, rather than merely confirming that the document exists.
Start with the claim whose error would matter most. Record the claim, primary document, supporting passage, population, period and status. A real study may support only part of a sentence or concern a different context. Another generated answer is not an independent source. Mark each claim as confirmed, context-limited or unverified before reusing it.
Priority
Check decision-critical claims first.
Scope
Connect each fact to the passage supporting it.
Status
Keep limits and unverified claims visible.
Does a genuine study validate the whole answer?
No. Check each material claim and its actual connection to the document.
Try the reasoning
Three AI answers cite the same article. Are those three sources?
Reasoning: They are three formulations with one documentary origin. Verify the claims in that origin.
An AI answer turns a narrow indicator into a population-wide claim
Fictional teaching case, not a client investigation.
Does the citation establish that “40% of all companies experienced an intrusion in 2025”? The report and numbers below are invented for this exercise.
Observation, interpretation and limit
Observation
What it establishes
What remains unproven
The answer gives a report title, with no page or passage.
A reference is asserted.
A plausible title does not establish that the document exists or supports the claim.
The exercise table reports 40% of 200 voluntary respondents experiencing a login alert in 2024.
The figure concerns a particular sample, event and period.
An alert is not an intrusion; a voluntary sample is not all companies; 2024 is not 2025.
Two articles reproduce the same table.
One observation has been repeated twice.
The repetitions are not two independent measurements.
Conclusion to record
The general claim is unsupported. A statement consistent with the exercise is: “Among the 200 voluntary respondents, 40% reported a login alert in 2024.” Do not reuse this fictional figure as a real statistic.
What would change the conclusion?
For a real claim, retrieve the primary document, exact passage, indicator definition, period and sampling method. If the document cannot be found, remove the figure.
Check your reasoning
A working link is insufficient: compare population, period and measured event word by word. Correct the scope of the sentence.
For each point, record what the material actually supports. “Supported” applies to the specific observation, not to the entire case. A contradiction needs explanation; missing material remains an open question.
Define the question
Record the source
Test the observation
State the limits
Entries stay in this page until you leave or reload it. Nothing is saved or submitted by this tool. Use neutral references instead of personal data or credentials.
How to write a useful conclusion
Editorial examples, not findings about your case
Material
Useful wording
Next check
One dated source
The consulted source stated X on the recorded date.
Check its origin and whether it still applies.
Multiple copies of one claim
Several pages repeat X; independent corroboration is missing.
Find the original and a source with independent evidence.
Conflicting sources
A states X while B states Y; the difference is unresolved.
Compare dates, scope, identifiers and definitions.
No matching result
No result was found in the recorded scope.
Describe queries, coverage and access limitations.
Common pitfalls
Four shortcuts that weaken the result.
Trusting confidence
Good prose does not replace a source.
Counting links
Check relevance rather than quantity.
Only asking the AI to confirm
Return to original documents.
Hiding doubt
Keep limitations in the delivered document.
Practical questions
Frequently asked questions.
Is a sourced answer reliable?
Links make checking possible. Their presence alone guarantees neither accuracy nor support for every sentence.
Must every word be checked?
Prioritise decisive claims, then precise facts, quotations and calculations. State the scope actually verified.
What should be delivered?
A corrected summary and a checking table linking each important claim to its status, source and limitation.
Public references
These primary references support the checks described here. The sequence and working sheet are an IIS editorial method; a cited organisation does not endorse this guide. Read each reference within its stated date, jurisdiction and scope.
Crossref — Crossmark. Reference for the technical fields and interpretation boundaries described in this guide.
DOI Foundation. Reference for the technical fields and interpretation boundaries described in this guide.
Editorial scope
Published by Internet Intelligence Service on 3 October 2026. Last content update: 29 October 2026. This educational guide describes a lawful, defensive method. It is not legal advice, an emergency service or authority instruction.