Case Report Flagged as AI: What You Can Rewrite When the Facts Are Fixed

Very little of a case report is yours to reword, and the parts that are tend to sit in two places: the introduction and the discussion. The case description belongs to the record, the section order belongs to the reporting guideline, and the declarations belong to the journal. Find the editable surface first, because it is smaller than the page count suggests.

This is a different problem from a flagged methodology chapter, where you can at least retreat into background prose. A case report has almost no background prose. It is a compressed account of one patient, written to a checklist, with a consent document attached to it.

Two things follow from that, and they pull in opposite directions. The document is short, so the score arrives blunt. And the document is dense with fixed content, so there is not much room to respond to the score. What is below is about working inside that.

The CARE guidelines were developed to improve the accuracy and transparency of case reports, and their checklist is the document most journals point at. Reading the published checklist as a list of owners rather than a list of tasks changes what a rewrite looks like.

The checklist item for therapeutic intervention says the report should cover `Administration of therapeutic intervention (dosage, strength, duration)` and `Changes in therapeutic interventions with explanations`. There is no version of rewriting that sentence which is safe if the numbers move. There is also not much room to move the wording around them without making the sequence ambiguous.

The Patient Perspective item is the one people forget. The checklist says the patient should share their perspective on the treatment they received. Rewriting that paragraph for style means editing somebody else's account of their own care.

If your flag lands mostly on the introduction and the discussion, you have a normal revision problem. If it lands on the case presentation, the answer is usually not a rewrite at all.

Journal instructions add a second layer of fixed text on top of the guideline, and it is more literal than most authors notice until they read the instructions in one sitting.

Journal of Medical Case Reports, to take one that publishes its requirements openly, asks for an abstract of no more than 350 words containing three named sections: Background, Case presentation, Conclusions. It asks for three to ten keywords. It requires that `Patient ethnicity must be included in the Abstract and in the Main Body under the Case Presentation section`. And it requires a Declarations block containing a fixed set of subheadings, with this instruction attached:

If any of the sections are not relevant to your manuscript, please include the heading and write 'Not applicable' for that section.

So a measurable part of the manuscript is a string the journal handed you and asked you to reproduce. Turnitin limits the analysis to qualifying text, described on its own page as `only prose sentences, meaning that we only analyze blocks of text that are written in standard grammatical sentences`, followed by `This percentage is not necessarily the percentage of the entire submission`. A block of headings followed by the words Not applicable is not prose sentences. The abstract and the case narrative are.

Which is the awkward shape of this genre. Take out the timeline table, the declarations block and the reference list, and what is left in the measured text is the abstract, the background, the case narrative and the discussion. The abstract is capped at 350 words and its middle section restates the case. The case narrative is not yours to alter on content. That leaves the background and the discussion carrying most of what is being scored. Turnitin does not publish a breakdown by section, so treat that last step as arithmetic about composition rather than something the vendor has confirmed.

Turnitin also says that in short documents `the prediction will be mostly "all or nothing" because we're predicting on a single segment without the opportunity to overlap`, and the sentence after it is the one that matters here: `This means that some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated.` A case report checked on its own, before the reference list and the tables come out of the count, is a candidate for that behaviour. Short documents and the all-or-nothing problem works through the mechanism.

When a passage reads as generic, people reach for one of two things. Add back concrete specifics. Or alter details so the passage reads as more distinctive. In a case report both are governed by rules that exist for the patient, not for you.

The ICMJE Recommendations on protection of research participants are explicit about the first one:

Identifying information, including names, initials, or hospital numbers, should not be published in written descriptions, photographs, or pedigrees unless the information is essential for scientific purposes and the patient (or parent or guardian) gives written informed consent for publication.

The paragraph after it opens with two more sentences on the same point: `Nonessential identifying details should be omitted. Informed consent should be obtained if there is any doubt that anonymity can be maintained.`

The second move runs into the last sentence of that same paragraph, which authors rarely quote:

If identifying characteristics are deidentified, authors should provide assurance, and editors should so note, that such changes do not distort scientific meaning.

Read that as a constraint on rewriting, because that is what it is. Any alteration you make to the described characteristics of a patient is something you may be asked to assure an editor about. Changing an age band, softening a comorbidity, generalising an occupation, all of it is inside that rule. The test the editor applies is not whether the sentence reads better. It is whether the case still means the same thing.

Journals apply their own version of the same logic to images. JMCR says images should be cropped only to show the key feature, that photos should not show a name, medical record number or date of birth, and that it does not consider images with patient faces or facial features.

The usual advice for prose that sounds machine-made is to make it more specific and more particular. In a case report, specificity and particularity are the two things the consent framework is designed to limit.

Keep these unchanged while the prose around them moves. Each one is a place where a fluent edit is a factual edit.

• Drug name, dose, strength, route, frequency and duration, including any change in them and the stated reason for the change.
• Laboratory values, imaging findings and their units, and the time at which each was obtained.
• The order of events in the timeline, and any explicit intervals between them.
• Diagnoses considered and rejected. Dropping a differential from a sentence for flow deletes part of the diagnostic reasoning.
• Adverse and unanticipated events, and the wording that separates a suspected association from an established one.
• Intervention adherence and tolerability, including how they were assessed.
• Patient ethnicity, where the journal requires it in named sections.

One rewriting failure in this genre deserves its own line, because it is invisible to a proofreader and obvious to a reviewer. The CARE checklist asks for `Clinician- and patient-assessed outcomes if available`. Those are two different sources of evidence. Turning "the patient reported reduced pain on day four" into "pain was reduced by day four" removes who assessed it, and a case report where clinician assessment and patient report are silently merged is a weaker report than the one you submitted.

The same trap sits in verbs elsewhere. Resolved is not improved. Declined is not was withdrawn. Tolerated is not completed. If a comparison of your two files shows a changed verb inside the case narrative, it needs checking against the notes rather than against your ear.

If you want the general version of this discipline for a study rather than a case, what you can rephrase and what must stay exact covers Methods and Results at length.

Rewriting a case report creates three pieces of downstream work that are easy to miss.

The populated checklist goes stale. JMCR requires the CARE checklist as an additional file and states that `Submissions received without these elements will be returned to the authors as incomplete`. A populated checklist points at where each item lives in your manuscript. Move a sentence between sections during a revision and the checklist is now describing a document you no longer have.

The AI disclosure has to be accurate and it has to be in the right place. The ICMJE Recommendations say journals should require authors to disclose at submission whether AI-assisted technologies were used, and that authors who use them `should describe, in both the cover letter and the submitted work in the appropriate section if applicable, how they used it`. The sentence at the end of that section is the one to read twice: `Nondisclosure of AI use may require corrective action and may be construed as misconduct in some circumstances.` JMCR's own instruction is more specific about location, saying use of an LLM should be documented in the Methods section, or in a suitable alternative part where there is no Methods section. Policies differ between publishers on where the statement goes, and publisher AI disclosure policies compared sets them side by side.

The consent statement still has to be true of the version you are submitting. Consent for publication in JMCR is mandatory for all case reports and must be written and obtained from the patient, a parent or legal guardian for children under 18, or next of kin where the patient has died. If a revision changed what is described about the patient, the question of whether the consented material and the submitted material are still the same material is a real one, not a formality.

That leaves the case where the flag sits on the sections that genuinely are yours, and someone has decided the prose needs work.

The expensive part is not the writing. It is that every touched paragraph is a paragraph you now have to read against the notes, the timeline and the populated checklist. A revision plan that covers the whole manuscript creates a verification job the size of the whole manuscript, in a document where a verification miss is a factual error about a patient.

Cutting that job down is what HumanPen does. Give it the AI Writing Report from Turnitin or iThenticate and the marked passages define the scope, so the case narrative stays out of the run unless you put it there yourself. The site describes the boundary this way: a paragraph is the smallest unit the engine rewrites, a partial selection is expanded to the full paragraph, and the expanded scope is shown to you to confirm before anything runs. It also says only the words actually rewritten are billed, and that where a result is eligible you can continue lowering AI for free. Files go in and come back in the same editable format, which matters when the manuscript is a DOCX carrying a numbered timeline and a reference list.

No part of that is a forecast. Turnitin dates its own detection changes in its release notes, which is the clearest evidence that a report run later is not the same measurement. The only claim being made here is about scope: which words are opened, and which stay exactly as the notes have them.

Can I paraphrase the case presentation to lower the score? You can change sentence shape, but the content is a record of what happened to a person, and the ICMJE asks authors to assure editors that any de-identifying change does not distort scientific meaning. Treat the case presentation as the last section to touch, not the first.

My case report is 1,800 words and came back with a very high percentage. Is that reliable? Turnitin says short submissions predict mostly all or nothing because there is a single segment and no overlap, and that mixed content can therefore be flagged as entirely AI-generated. It also says an AI Writing Report needs at least 300 words of prose. Take out the timeline table, the declarations and the references and a case report is shorter than its word count suggests.

The whole discussion is highlighted and the timeline table is clean. Does that mean the table passed? No. Turnitin says the model does not detect short-form and non-sentence structures, so a table of short entries is largely not measured. It also says long-form prose text in tables is now processed, so a table full of full sentences is measured. A clean table is usually an absence of measurement, not a result.

Do I have to declare that I used a rewriting tool? Check the journal, and read its wording rather than a summary of it. The ICMJE position is that authors should disclose AI-assisted technologies at submission and describe how they were used, and that nondisclosure may be construed as misconduct in some circumstances. That is a stronger sentence than most authors expect, and it is the reason to sort this out before submission rather than after.

Does the CARE checklist require a specific number of items? The published checklist is organised under headings from Title through to Informed Consent, and journals that endorse CARE generally ask you to submit it populated. The practical point is not the count. It is that most of the headings name content whose source is a record, a guideline or a patient, and the ones that leave you room to argue are the introduction and the discussion.

Is a case report treated differently from a research paper by the detector? The mechanism is the same. What differs is the composition of the document. Do journals check papers for AI writing covers what actually runs at submission, which is a separate question from what a supervisor might run on a draft.