Incorrect References by #ChatGPT! Finding the Right References for Literature

ChatGPT can generate literature-review-ready arguments, but it also can supply references that don’t actually exist or don’t match the claim being made. That mismatch becomes a serious problem when researchers copy the citations directly into a thesis or paper without verifying them—especially when the argument is plausible but the source details are wrong.

The session frames the core risk around how academic writing is structured. Different sections demand different kinds of content: introductions typically avoid deep variable-by-variable conceptualization and instead set context, while literature reviews are where detailed relationships among constructs belong. If a researcher doesn’t know what belongs where, they’re more likely to accept AI output at face value and then struggle to integrate it correctly into the right part of the manuscript.

To address the referencing problem, the guidance is practical: treat ChatGPT as a drafting assistant for ideas, not as a citation authority. A recommended workflow starts by generating a response with references, then rewriting the argument in one’s own words and using the citations only after checking that the cited papers truly exist and support the specific sentence.

A worked example centers on “entrepreneurial leadership,” “employee engagement,” “creativity,” and “risk-taking.” When the same prompt is run again, the output may provide a full reference list that looks credible and the argument reads as coherent. But the crucial test is whether the referenced authors and titles can be found in databases such as Google Scholar. In the example, one citation attributed to “Poon Chen and Lynn” could not be located with the expected title/author combination, even though the underlying claim sounded right: employees are more likely to feel motivated and committed when an organization values innovation, risk-taking, and creativity.

When a citation fails verification, the fix is to rebuild the reference using keywords from the problematic sentence. The approach is to extract key concepts—such as “motivated,” “innovation,” “risk-taking,” and “creativity”—and then search for papers that contain those terms in the text. The session suggests iteratively shortening or rephrasing the sentence to improve searchability, then opening candidate papers to confirm that the wording and logic align with the claim. Another tactic mentioned is using an “All intext string builder” style query in Google Scholar, including exact phrases in inverted commas, to surface documents where the relevant concepts appear.

The takeaway is not to abandon AI, but to verify and to understand manuscript structure first. Researchers who learn what belongs in each section and who validate citations through keyword searches and database checks can use ChatGPT’s argument generation while avoiding the citation errors that derail academic writing.