Six AI-generated research details to verify
For writers editing a research-based article, the task is to check whether each small factual detail matches the source and the claim it is being used to support. Names, dates, numbers, quotations, source titles and scope words can look incidental, yet a tiny change may make a citation hard to trace or make evidence appear to say more than it does. Treat each category differently: some details need exact transcription, while others need a check of context and boundaries.
check identity, spelling and attribution
Names are easy to “correct” into a familiar form. A middle initial may disappear, a surname may be misspelled, two authors may be merged, or an organization may be credited with a report it only hosted. Similar names can also point to different people. These are distinct errors: spelling affects identification; attribution affects who did or said the thing.
Compare the name against the source page itself, especially its byline, author note or organization’s official page. Do not rely on a search-result snippet or a bibliography copied into another article. In a research citation, check author order and initials as well as spelling: they help readers locate the work. If the source names a group author, retain that group name rather than guessing at individual authors.
**Counterexample:** A source lists “M. Lee” as one of four authors. Expanding that to “Michael Lee” because a related staff page has a Michael Lee is an unsupported identification, even if the person might be the same.
A real example of why a plausible-looking reference deserves a direct check comes from Walters and Wilder’s 2023 study of ChatGPT-generated bibliographies. The authors examined citations produced in literature reviews across 42 topics and reported both fabricated references and errors in real citations. The study concerns the tested systems and prompts; it does not establish an error rate for every AI tool or writing task. ([Scientific Reports study](https://www.nature.com/articles/s41598-023-41032-5))
distinguish the event, publication and update
A date can be accurate and still answer the wrong question. Research pages may show when a study was conducted, first published online, assigned to a journal issue, updated, or accessed. Reports may also cover a period that differs from their publication date. When prose says “the study appeared in 2021,” using the date of a later webpage update would be misleading, even if that update date is displayed prominently.
Match the date to the claim’s verb and noun. “Published” calls for a publication date; “surveyed in” calls for the fieldwork period; “updated” calls for a revision date. If a source gives only a year, do not invent a month or day. When two legitimate dates matter, label both rather than silently choosing one.
**Counterexample:** A report page says “Updated 14 March 2024” and describes a survey conducted in 2022. Calling it “a 2024 survey” confuses the page’s update with the data collection period.
verify the unit, denominator and comparison
Numbers often survive paraphrase while their meaning changes. Check the unit (people, papers, citations, or responses), denominator, time period, population and whether the figure is a count, percentage, estimate or rounded value. Then check the comparison: “rose by 10%” and “rose by 10 percentage points” are not interchangeable. A share of citations that contain an error is not the same as a share of articles with at least one error.
Return to the table, chart, methods or source passage where the figure originates. Recalculate only when the inputs and operation are clear; otherwise quote the source’s reported figure and preserve its qualifiers. Don’t convert a narrow sample result into a general frequency claim.
For a simple illustration, 12 of 20 gallery visitors and 20 of 40 gallery visitors are different counts, yet the first is 60% and the second is 50%. An AI summary that copies only “20 visitors” or only a percentage drops the comparison needed to understand the claim. These figures are invented for illustration; verify real numbers against the original table and its notes.
Quotations: verify the words and the speaker
A fluent sentence in quotation marks may be an exact quotation, a loose paraphrase, or text the model invented. Find the original passage and check the words, speaker, date and surrounding context. A quote attributed to a group may have been spoken by one participant; a memorable phrase may be the article author’s summary rather than a participant’s words.
If the original source does not contain the exact wording, remove the quotation marks. Paraphrase what the source actually says and cite it, or omit the line when you cannot confirm it. Do not silently tidy a quotation into smoother prose while presenting it as exact.
Source titles and identifiers: keep the reference findable
A plausible article title is not proof that the article exists. Search the exact title and author in the publisher’s page or a bibliographic index, then compare the DOI, date and journal or book title. A changed word in a title may lead to a different work or no work at all. Crossref’s metadata search can help locate a registered publication, but a matching record still does not show that the work supports the sentence beside the citation.
Walters and Wilder examined 636 references in 84 short reviews generated by GPT-3.5 and GPT-4 in 2023. They reported fabricated references and errors within references to real works. The result applies to those tested versions and prompts; it is evidence for checking references, not a current error rate for every AI writing tool.
Scope words: check who, where, when and how strongly
Small words can quietly expand a claim. “Some,” “all,” “usually,” “caused,” and “proved” say different things. If a source reports one workshop on one date, a paragraph about all workshops is broader than the evidence. If two measures moved together, calling one the cause of the other needs more support than their shared timing.
Mark every phrase that widens the original finding: a larger audience, longer period, new setting, stronger certainty or causal verb. Restore the source’s boundaries in the sentence. If the source says “participants in this trial,” retain that group. If it says “may,” do not change it to “will.” This check asks what the evidence covers, rather than whether the sentence merely sounds confident.
A six-detail scan before publication
Run one focused pass over names, dates, numbers, quotations, source titles and scope words. For each detail, ask whether the original source supports this exact form and whether a reader could find it. Correct, narrow or delete a detail that fails the check. Then read the paragraph once more to confirm the repaired details still support its overall meaning.
This scan is deliberately narrower than a full argument or source-quality review. It catches small factual changes that can make an otherwise careful research article unreliable, while leaving broader decisions about framing and reasoning for a separate editorial pass.