What happens to citations, tables and equations in an AI humanizer?

Turnitin excludes bibliographies when it processes the AI writing report, and the report only analyses prose sentences, so rewriting a reference list changes nothing about the number. It can still cost you the references themselves. Markers move when sentences are merged or split, reference-manager fields flatten into dead text, and one recorded review shows a tool that rewrote the reference list and produced a citation that was not in the input.

Everything below is about that second sentence, because it is the one nobody prices in.

There are two shapes of rewriting tool and they fail differently.

The first takes a text box. You paste, it rewrites, you paste back. The tool never sees your document, so every question about tables and fields is really a question about what you did before and after pasting. Whatever structure was in the file, you removed it by hand and you are about to put it back by hand.

The second takes the file. It opens the document, finds the runs of text inside it, rewrites some of them, and writes the document back out. That kind of tool can protect structure, and it can also break structure at a scale you would never manage manually, because it touches every paragraph in one pass.

Neither shape is safe by default. What decides the outcome is scope: how much of the document is in the rewrite, and whether you got to see that boundary before it ran.

Turnitin's own description of the detection mechanism is worth reading once, because it explains why nothing about your citations is protected on the detector's side. Submissions are cut into overlapping segments; every segment is scored between 0 and 1; a qualifying sentence picks up the values of whatever segments cover it; those get pooled and aggregated into the number on the report (AI writing detection FAQs). There is no step in that sequence that recognises a citation.

The same blindness runs the other way in a rewriting tool. To a model rewriting prose, `(Smith, 2019)` is a run of characters sitting inside a sentence, with nothing marking it as load-bearing. Merge two sentences and the marker is left standing behind a claim its source never made. Split one, and half the sentence keeps the marker while the other half becomes an unsupported statement that reads perfectly well.

That failure is invisible to a proofread, because nothing is ungrammatical and nothing is missing. The sentence supports itself. It just no longer says what the source said. We covered how to audit that link claim by claim in revising a literature review without breaking the evidence chain.

The second failure is mechanical rather than semantic. If your citations came out of Zotero, Mendeley or EndNote, they are not text. They are fields, and a plain-text round trip turns them into characters that look identical and no longer update. Word fields, tables of contents and cross-references goes through what that costs in a long document.

This is not hypothetical, and I would rather point at someone else's recording than my own claim.

Tadhg Blommerde lectures at Northumbria University, holds instructor-side Turnitin access, and states in the descriptions of his reviews that they are unsponsored and carry no affiliate links.

His method in a February 2025 review is what makes the finding checkable. The test essay is not his own writing. He asks ChatGPT for it — "1,500 words in UK English", with "intext citations and a reference list in the APA style" — so a known input exists on screen before anything else touches it. That file then goes through Walter Writes, and he reports that it "even humanized the reference list". He also stops on a citation in the output with the line that matters: it wasn't in the ChatGPT version.

That is why the second finding is worth more than the first. A mangled reference is an error you would eventually notice. A citation nobody wrote is a fabricated source sitting in a document you are about to put your name on, and the only reason anyone can say it was fabricated is that the input was produced in the video first and could be compared against the output line by line.

The only open dataset I could find on this comes with a disclosure attached, and the disclosure has to travel with the numbers. The benchmark, HumanizerBench, is run by WriteHuman — one of the twelve tools on the leaderboard it publishes, and the tool sitting at number one in the cycle I am about to quote. That is not something I worked out; they say it themselves on the about page and again in the footer of every page on the site. Independent, then, is not a word that can be attached to it, whatever anyone quoting it calls it.

What it is, unusually for this category, is a benchmark that hands over the files. Every input and output of the August 2026 cycle sits in a public repository under CC BY 4.0. That is what makes it usable in spite of who runs it: you can disagree with anyone's summary of it, this one included, by opening the same file and counting.

So here is the part worth counting, and the rule is as plain as it sounds — a Markdown heading is a line that starts with a hash, and after the rewrite it is either still there or it is not. Across the twelve tools in that cycle, four returned outputs with no heading left in them at all, and three returned every heading their inputs had. "These tools all destroy your formatting" is a comfortable thing to believe, and the published files do not support it. The spread inside this category is the whole range.

Then there is the limit of the measurement, which matters more than any of the percentages. What is being counted is a hash character at the start of a line of plain text. Every tool in that dataset is paste-in, paste-out; no `.docx` changes hands anywhere in it. Styles, numbering definitions, table objects, footnote parts, cross-reference bookmarks and TOC fields live in a different layer, and I could not find a public dataset that goes near that layer at all.

So we counted it ourselves, on our own output. This is us measuring us — three before-and-after pairs from our own runs, read with a script that opens `word/document.xml` rather than trusting anything the product reports about itself. Take the largest of the three: 405 non-empty paragraphs, 46 paragraphs carrying a heading style, one table of 7 rows and 14 cells. Afterwards, all three counts were identical.

Two qualifications, and the second one is the important one.

A heading surviving as a heading does not mean its wording survived. Of those 46 headings, 18 came back with different words under the balanced setting, still attached to the same style. Being reworded and being flattened into body text are separate events, and the Markdown counts in the previous section are picking up the second one.

The bigger limit is what the corpus did not contain. No table of contents field, no cross-references, no footnotes: zero `fldChar`, zero anchors, no `footnotes.xml` anywhere in the package. Those are the parts most likely to break quietly, since none of them are stored next to the words they belong to; a footnote marker and the footnote itself sit in different files inside the package, connected only by a number. So the honest scope of that measurement is paragraphs, heading styles and one table, on three documents we picked.

You cannot reproduce our numbers. The files are not ours to publish, and a number from a vendor's own corpus is worth roughly what you would expect. What you can reproduce is the procedure, on the document you actually care about, in an afternoon: copy the `.docx`, rename the copy to `.zip`, open `word/document.xml`, count the non-empty `<w:p>` elements, count how many of them carry a paragraph style whose name begins with `Heading`, and search the package for `fldChar` and for `footnotes.xml`. Do that before and after any tool you are considering, ours included, and the number you get is about your thesis rather than about our three files.

Anyone who tells you their tool preserves cross-references, us included, should be asked what they counted.

Turnitin's AI writing detection model page carries release notes, and the entry for 9 August 2023 says the model is now able to process long-form prose text in tables. The same entry adds that existing submissions containing tables have to be resubmitted before they are reprocessed.

Two things follow. Prose sitting in a table cell can carry AI highlighting, so "it's in a table, it doesn't count" is out of date by three years. And that puts tables inside a rewriting tool's scope as well: if the cell counts towards the score, a file-level tool has a reason to go in there, cell by cell.

I have not measured what different tools do inside a cell, so treat the rest of this paragraph as where to look rather than as findings. A cell that held two paragraphs is worth reopening. So is a row that grew taller because its text got longer. And in a results table, the cell to check first is the one holding a number with a unit written after it, since that is the content in a table that most resembles a sentence.

I have no measurement to offer on equations, and I would rather say that than fill the gap with a confident paragraph. Our corpus had none.

What I can give you is the boundary that matters, and it is not really about equations. It is about anything whose value is the value. A quoted sentence must come back with the same words inside the quotation marks or it is no longer a quotation. A sample size, a p-value, a confidence interval, a unit, a date and a proper noun all have to survive character for character. Prose can move. These cannot.

What you can rephrase in Methods and Results, and what must stay exact sorts a paper into those two columns properly.

1. Open the file in Word, press Ctrl+A, then F9. Fields that survived will refresh. Fields that were flattened will sit there doing nothing, which is how you find them.
2. Count your reference entries before and after. A number that changed is the loudest possible signal, and it takes ten seconds.
3. Search the output for every in-text marker and check the sentence it now sits in still makes that claim. This is the slow one. It is also the one that catches fabricated support.
4. Read every table cell that contains a number. Not the prose around it. The cell.
5. Diff the quoted material. Anything between quotation marks should be byte-identical to the source.

The full version of this, with the file-versioning setup that makes it repeatable, is in how to review a humanized Word document before submission.

The premise behind HumanPen is that "what got touched" should be an answer you gave, not one you discover in Word afterwards.

The engine will not rewrite anything smaller than a paragraph. Stop a selection halfway through one and the whole paragraph is what gets rewritten, and you see that boundary drawn before anything runs. You can set the boundary yourself, or import a Turnitin or iThenticate report and let its flagged passages set it, in which case only the flagged text is rewritten. Credits are counted on the words actually rewritten, so a job scoped to four paragraphs is priced as four paragraphs rather than as a thesis. Eligible results can continue lowering AI for free.

Our own FAQ ends that section with "Review complex documents after download", and I would not remove that line. A rewrite of a complex document is something you check. What scope buys you is a much shorter list of things to check.

Does a humanizer change my reference list? It depends entirely on whether the reference list was in scope. Nothing about a bibliography helps your AI number, because Turnitin's release notes say bibliographies are excluded when the AI writing report is processed, so there is no reason for it to be in scope in the first place. A recorded review of one tool shows a reference list being rewritten anyway.

Will my in-text citations still match my bibliography afterwards? The markers usually survive as characters. What breaks is the connection between the marker and the claim, when sentences are merged or split around it. Check the sentence, not the bracket.

Do tables get rewritten? They can be, and Turnitin does analyse long-form prose inside tables, so a tool has a reason to go in. Cells holding numbers, units and short labels are the ones to inspect first.

Why did my table of contents break? Almost always because the document went through a plain-text stage. A TOC is a field, not text; flatten the document to a string and the field is gone even though the words are still on the page.