When AI helps you research the useful question is whether the source actually supports the sentence
By Arbab Naseebullah Kasi, Chief Executive Officer - CEO, Feel Worldwide Foundation Inc.
A paragraph can look thoroughly researched before anyone has checked its evidence. There is an author, a date, a journal title, perhaps even a link. The formatting is neat. The claim sounds reasonable. It would be convenient to move on.
But a citation is a direction to look, not proof that someone has already looked. When an AI tool supplies references, the work still includes finding the source, reading the relevant material, and deciding whether it supports the sentence beside it.
That distinction matters far beyond academic papers. A small nonprofit may use research in a funding proposal. A trainer may prepare a handout. A business owner may compare services or explain a new policy to customers. In each case, an impressive-looking reference can travel from a draft into a decision before its connection to the claim has been tested.
The goal is not to make people afraid of using helpful tools. It is to keep the convenience of drafting separate from the responsibility of saying something is established.
Why a plausible reference deserves a real check
A 2023 study by William H. Walters and Esther Isabelle Wilder examined 636 references in 84 short literature reviews generated on 42 topics. In that particular sample, 55 percent of references produced by GPT-3.5 and 18 percent produced by GPT-4 were fabricated. Even references to real works sometimes contained substantive errors. Authors' research record at CUNY
Those figures describe historical models, prompts, and outputs studied in 2023. They are not current error rates for every AI assistant, and they should not be presented that way. Their lasting relevance is more modest: a reference can look like an ordinary scholarly citation while failing to identify a real work.
NIST's July 2024 Generative AI Profile also identifies confidently presented false content as a risk. Its discussion specifically includes fabricated citations that can encourage misplaced trust, and its suggested risk-management actions include reviewing and verifying sources in generated outputs. NIST Generative AI Profile
Neither source tells us that every generated reference is wrong. Both give us a reason not to treat polished citation formatting as a completed check.
Existence is only the first question
Start by locating the work independently. Search the title in a library catalogue, a scholarly database, or the publisher's website. If the citation includes a digital object identifier, commonly called a DOI, follow it and compare the resulting record with the reference you were given.
Does the title match? Do the authors match? Is it the same publication rather than another work with a similar name? A link that opens successfully may still lead to the wrong item.
The University of Alberta Libraries advises readers to check whether generated sources are real and, when they exist, read them to verify the information attributed to them. Its guidance also warns that invented citations may borrow the names of real researchers or journals. University of Alberta citation guidance
This is a useful place to separate three situations. A source may be verified as real. It may be inaccessible to you. Or it may remain unlocated after a reasonable search. These are different findings.
An inaccessible article is not automatically a fake article. Equally, a plausible title that you cannot locate should not quietly become a verified reference. Leave it out of the finished claim until you have sufficient evidence, or seek help from a librarian or someone with appropriate access.
Read for the claim you are making
Once you find the source, the next question is not merely whether it mentions the topic. It is whether it supports your wording.
Imagine a fictional training organization preparing an article about online courses. An AI-assisted draft says, “Online training consistently improves employment outcomes.” The attached source turns out to be a real survey in which participants reported that they liked a particular course.
The citation exists. The sentence still does not follow from it. Satisfaction is not employment, one course is not all online training, and a participant survey does not automatically establish what caused a later outcome.
A responsible revision might describe the satisfaction finding accurately, if it is relevant and sufficiently documented. Alternatively, the writer might remove the sentence and look for research that actually examines employment outcomes. What should not happen is keeping the original claim because the reference looks respectable.
The same check applies to smaller words. “May help” is different from “will improve.” “Associated with” is different from “caused.” “In this sample” is different from “for everyone.” These distinctions can disappear during editing even when the source itself is careful.
Treat the evidence and the sentence as a pair. If the sentence becomes stronger, broader, or more certain, its supporting evidence needs another look.
Numbers need their surrounding sentences
A percentage without context can be easy to quote and hard to interpret. Before carrying a number into your own work, identify who or what was counted, when the observation was made, and what the denominator includes.
The historical citation study above illustrates the point. Its percentages refer to references produced by particular models within the researchers' sample. They do not describe the percentage of all AI answers that are false. Changing the unit from references to answers would change the meaning, even if the numerical value remained identical.
Also look for exclusions, response rates, comparison groups, and the difference between a reported estimate and a measured result. You do not need to reproduce every methodological detail in a short article. You do need enough context to avoid giving readers a stronger conclusion than the evidence allows.
For quotations, inspect the actual wording and its surroundings. A quotation should not acquire words that were never there, and removing a qualifying phrase can misrepresent an otherwise accurate excerpt. When you cannot inspect the quoted passage, do not present it as personally verified.
Keep a small record of the evidence connection
For a document with several factual claims, a simple working record can make review easier. Beside each consequential claim, keep the source link and a short note identifying the relevant section, table, or page. Add the limit that matters most to the sentence.
For example, a working note might say: “Participant satisfaction survey, not employment results. Use only for reported satisfaction.” Another might say: “Historical model evaluation. Do not describe as present-day performance.” These notes are editorial aids, not extra material that must appear in the published piece.
This suggested practice is not a guarantee of accuracy. It makes the reasoning inspectable. A colleague can see why a source is attached rather than having to reconstruct the connection from a bibliography alone.
Keep the record proportionate. A short community announcement may contain only a date and a venue to verify. A report making recommendations may need a more substantial review. The effort should follow the consequences of being wrong, not the decorative length of the reference list.
Do not make confidence do the work of evidence
Asking an AI tool to reconsider an answer can help expose a problem or suggest a better search. It does not, by itself, establish that the revised answer is correct. A confident reassurance is still a statement to evaluate.
A more useful request is to ask the tool to identify which claims need checking, distinguish quotations from paraphrases, or help generate search terms. Then verify important claims against material you can actually inspect. If the tool cannot access a source, that limitation should remain visible rather than being covered by a fluent summary.
The same standard applies to a second tool. Agreement between two generated answers is not necessarily independent corroboration. For your finished document, look for the evidence beneath the agreement.
Be careful with private documents as well. Verification should not become a reason to upload confidential participant records, internal proposals, or personal information to an unapproved service. Use appropriate access and your organization's rules. Where possible, check the public factual claim without sharing unrelated private material.
Make correction an ordinary part of finishing
Finding a weak citation before publication is useful work. It is not an embarrassment to hide. Remove the unsupported claim, narrow it, or replace the evidence with something that genuinely fits.
If a document has already been shared, consider what the error changes. A broken link may need a straightforward repair. A false statistic supporting a recommendation may require a visible correction and direct notice to the people relying on it. The response should match the consequence, not merely the ease of changing the file.
Teams can make this easier by assigning a reviewer to the evidence as well as the prose. “Looks good” is ambiguous. “I checked these three factual claims against the linked sources” tells the next person what review actually happened.
AI-assisted writing can produce a draft quickly. A dependable document is finished only after the important connections between its claims and its evidence have been examined. That work is not a nuisance added after writing. It is part of what makes the writing worth trusting.
Which sentence in your next report would you most want someone to check before a reader acts on it?

