Guide for Medical Researchers: Did My Student Use AI in a Paper?

The use of generative AI (GenAI) in academic texts is not always easy to detect. Unlike plagiarism detectors, AI text detectors are not always accurate, with frequent false positives and scores varying across tools. Moreover, one has to deal with the distinction between human-written, AI-assisted, or AI-generated work. This might because most students today are using GenA) at four stages:

Stage 1: Planning (brainstorming, literature search)

Stage 2: Writing (outlining, text generation, language checks)

Stage 3: Analysis (inputting confidential, unpublished data)

Stage 4: Dissemination (communicating research outputs)

Using AI at each of these stages has different strengths and weaknesses, as well as risks. Finding out that your student is using AI is not necessarily a cause for alarm; not all AI use signals misconduct. Rather than outright forbidding AI use in your lab, it would be prudent to understand how your students are using AI and guide them to do so responsibly.

Here is a checklist for identifying if and how your student used AI in their writing.

STAGE 1: AI USE AT THE PLANNING STAGE (for brainstorming, searching the literature)

Risks: An overreliance on GenAI at this stage could lead to the inclusion of fake references and false claims in the text. The student might not understand their own hypothesis fully and/or diminish their own ability to critically analyze the literature.

Questions to ask

[ ] Ask the student for the search strategy used. Which databases, search terms, and if applicable, prompts, were used?

[ ] Spot-check the references: choose five to ten. Are they real? Do the authors, year, DOI, and journal match? Does the paper say what the sentence claims? Specifically check for:

– Nonexistent papers

– Inaccurately cited papers: irrelevant to the claim, misquoted, or misrepresented

– Missing landmark papers.

[ ] Check the student’s reading. Ask to see annotated PDFs. A student might have collected papers, but did the student read them?

[ ] Ask “What did the tool suggest, and what did you choose?” Brainstorming with AI is fine, but the student should be able to defend their hypothesis.

STAGE 2: AI USE AT THE WRITING STAGE (outlining, text generation, language checks)

Risks: Using AI for grammar and language checks of one’s own text is widely accepted. Using AI for outlining is low to moderate risk. But generating substantive text is where verification matters most, as a student may not understand the draft themselves. Finally, a journal might flag heavy AI use and reject the manuscript.

[ ] Check that numbers trace back to a table, figure, or source file. Do sample sizes, percentages, p-values, confidence intervals, and units agree across abstract, results, and tables?

[ ] Do the methods describe what was done in your lab/clinic: instruments, catalog numbers, software versions, ethics approval number, consent process? Generic-looking methods could point to AI generation.

[ ] Does the Discussion match with your data, or summarize the field a bit too smoothly and then claim more than you found?

[ ] Compared to the student’s prior writing samples (drafts, lab-notebook entries, email history), do you see sudden changes in the following?

– That student’s vocabulary or sentence complexity

– Their tone or spelling

– Their ability to explain the argument orally

[ ] Search for or fingerprints, which signal copy-pasting of GenAI output:

• Run [Ctrl+F] for GenAI residues such as “Certainly! Here’s a revised…” “As an AI language model,” and “Regenerate response.”
• Run Ctrl+F for series of hashtags (###) or asterisks (**), formatting markers that might appear when chatbot output is pasted into a document.
• Check for prompt instructions or user directions pasted directly into the final text: “Make this sound more professional” or “reduced word count.”

[ ] Check version history: did the text build over days or arrive suddenly in large blocks?

[ ] Before asking for the polished manuscript, always ask for an outline to see the thinking process in action. Then ask for completed sections, at various stages of preparation.

[ ] Ask which tool and version was used for language polishing. You could also ask for a demonstration.

* Treat these indicators as reasons to ask further questions, not as proof of AI use.

STAGE 3: AI USE AT THE ANALYSIS STAGE (entering confidential, unpublished data)

Risk: Data that has gone into a third-party tool cannot be “unsent”! Further, generating fake data or unethically modifying it is strictly forbidden.

Questions to ask

 [ ] Ask your student if any of the following were entered into a public tool and which tool it was:

• identifiable patient data
• protected health information
• genomic data, human-subject information
• unpublished results
• confidential grant text
• confidential peer-review material

[ ] Ask if data were de-identified and how.

[ ] Ask if any data, clinical content, images, code, participant descriptions, or statistical results were fabricated or altered improperly.

STAGE 4: AI USE AT THE DISSEMINATION STAGE (communicating research outputs: posters, slides, plain-language summaries, press releases, social media posts)

Risks: AI can add unnecessary spin or make the findings a tad exaggerated. AI-generated figures might be misrepresentative.

Questions to ask

[ ] Does the plain-language summary or press release say more than the paper? Watch for “breakthrough,” “proves,” and “cures.”

[ ] Do the numbers in the abstract, poster, slides, and posts match the paper?

[ ] Verify that AI-generated or AI-modified images do not misrepresent experimental results. Check such images for anatomical inconsistencies, gibberish text, melting letters, unreadable background logos, or texture blending.

THE MOST RELIABLE TEST: A ROUTINE CONVERSATION

[ ] Ask the student to explain the central research question in their own words.

[ ] Ask why a particular method was chosen and what alternatives were considered.

[ ] Ask them to interpret a key result and explain its clinical or biological significance.

[ ] Ask them to identify the study’s main limitation and how it affects the conclusions.

[ ] Ask them to walk through one citation and explain how it supports the relevant statement.

If the student cannot answer these satisfactorily, GenAI might have been used for more than just writing assistance. They did not engage sufficiently with the science. In fact, you may ask these questions to all the students in your lab routinely, in the tone you’d use for any general query.

ONE LAST CHECK: HOLDING THE MIRROR

[ ] Did you read the student’s entire draft, or only the abstract and the figures?

[ ] Did you use an AI tool to summarize the student’s draft?

As a lab head/PI, your role is to ensure that every manuscript remains accurate, original, and fully attributable to the researchers who take responsibility for it. Once you use this checklist, you will understand the pattern of AI use in your lab. Going a step further, you can take the following measures to establish your own lab policy on acceptable AI use. You can build on the following sample. Remember to update the policy as needed; tools and their capabilities keep changing!

Policy on Acceptable AI Use: Lab XYZ

• Regularly review the institution, department, journal, funder, and laboratory policies governing AI use.
• Check whether the manuscript, protocol, grant, thesis, or report you are working on permits GenAI use in brainstorming or outlining, grammar and style editing, literature searching, coding or statistical assistance generation, or alteration of figures/images.
• Check whether disclosure of AI assistance is required and, if so, where it must appear.
• Maintain raw data and version histories of drafts.
• Decide who signs off on preprints, social posts, and media statements before they are posted publicly.
• Cultivate a healthy, regular reading habit (a reading club?) in the lab to encourage critical thinking and analysis.
• Maintain an “AI Use Log” containing:
• AI Tool and version used
• Date and task
• Input provided, with confirmation that no confidential data were included
• Output used or rejected
• Human edits and verification performed
• Disclosure in the final document