Verification · Practical guide
How to Check AI Answers and Fake References
Fluent writing is not evidence. Check the underlying claim against a source or a reproducible test, rather than asking the same chatbot whether its answer is correct.
Separate claims from opinions
Highlight names, dates, numbers, quotations and statements that could be wrong. Put each important claim on its own line. Prioritise errors that could change a decision, assessment or technical outcome.
Example: “This wireless setting doubles throughput in every environment” contains a universal performance claim. Identify its assumptions before accepting it: equipment, bandwidth, interference, traffic and measurement method all matter.
Trace each important claim to its evidence
Open the cited source yourself. Read the relevant passage, not just the search snippet or article title. Record whether it supports, contradicts or does not address the claim. Prefer original documentation or research for technical facts.
A real webpage can still be irrelevant. A source about signal strength does not automatically support a claim about application throughput. Mark a claim “unverified” when you cannot find supporting evidence.
Check whether a reference exists and matches
Search the exact title and author. Where a DOI is supplied, open it through doi.org and compare the title, authors, year and publication venue with the cited entry. A resolving DOI is not proof that the accompanying citation is correct.
If no DOI is available, use the publisher, journal, library catalogue or repository. A missing DOI does not make a work fake. If you cannot locate it, describe it as unverified rather than asserting fraud. Read the actual work before claiming that it supports your argument.
Recalculate and test independently
For arithmetic, recompute the result with a calculator and check units. If a draft says 12 packets lost from 200 sent equals 12%, recomputing 12 ÷ 200 × 100 gives 6%.
For code, inspect it before running it in a disposable environment with synthetic data. For changing information, check the publication date and a current official source. Do not use an AI confidence percentage as a substitute for validation.
Keep a simple evidence log
Use four columns: claim, source or test, result, correction. Example: “packet loss is 12% | independent calculation | incorrect | replace with 6%.” Record unresolved assumptions instead of hiding them.
Before sharing the result, re-read the corrected text to ensure the correction does not break another conclusion. If your course or employer requires it, disclose the assistance used and the checks performed.
Practice with a safe example
Ask an AI tool for a short explanation of a familiar topic. Select three factual claims, verify them independently and record one uncertainty. The goal is a defensible result, not proving that every AI response is wrong.
Worked evidence log: distinguish fact, inference and invention
Use this fictional source card: “The workshop has 80 places. Twelve places are booked. Registration closes on Friday at 16:00.” A generated draft says “Only 12% of places are booked, so demand is low and registration will probably be extended.” Separate this into three claims before correcting the wording.
Claim 1 concerns a calculation: 12/80 × 100 = 15%, so 12% is incorrect. Claim 2 concerns demand. The source does not describe how long registration has been open, the expected booking pattern or other evidence of demand, so that interpretation is unsupported. Claim 3 predicts an extension. The source states a deadline but gives no extension policy; the prediction should be removed or explicitly identified as speculation outside the supplied evidence.
A supported revision is “Twelve of 80 places are booked, representing 15% of capacity. The stated registration deadline is Friday at 16:00.” Do not add a calendar date or timezone when the source omits them. If the message is to be used operationally, resolve those details with the authorised source before sending it.
Practice answer and a stopping rule for verification
For a 15-minute exercise, spend three minutes extracting claims, five checking them, four writing the corrected version and three recording remaining uncertainties. Use four evidence-log columns: claim, source or calculation, finding and action. In the workshop example, enter “12% booked | 12/80 × 100 | incorrect | replace with 15%”; “low demand | source card | not established | remove”; and “extension expected | source card | not established | remove”.
A claim can be unverified without being proved false. This distinction matters when checking references: failure to locate a paper in one search does not establish fabrication. Record the attempted search, check the publisher or repository where appropriate, and describe the limitation honestly. Never invent a DOI or a publication detail to make a reference look complete.
Stop when the important claims are supported, corrected or explicitly bounded, and the revised answer still addresses the task. Repeatedly asking a model to approve the same text does not add independent evidence. If a material uncertainty remains, identify what source or test would resolve it instead of polishing it away.
Self-review: can another reader follow your calculation or source trail? Have you distinguished source statements from your interpretations? Does the corrected text preserve important conditions and acknowledge unresolved details? Retain the log with the final version so a later source change can be traced to the affected claim.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
List the factual claims in the text below. Separate claims, assumptions and opinions. For each factual claim, say what evidence would be needed to verify it. Do not invent sources or mark a claim verified without evidence. Text: [paste non-sensitive text].Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.