IEEE Format and AI Detection: What the Style Manual Takes Out of Your Sentences

Two things are going on and they are worth keeping apart. IEEE's editorial style manual specifies your citation grammar, your abstract length, your heading labels, your caption openings and a list of specific words, which flattens the parts of the paper a detector actually reads. And a two-column PDF is not a text file, so the text a machine gets out of it may not be in the order you wrote it. The first is about what your sentences are allowed to look like. The second is about whether the thing being analysed is even your paper in the right sequence. You can check the second one yourself in about thirty seconds.

Nothing on the vendor's side is about IEEE, or about formats at all. We read the three Turnitin pages that define the AI Writing Report (Using the AI Writing Report, the AI writing detection capabilities FAQs, and File requirements for an AI Writing Report) on 27 August 2026, 38,174 characters of rendered article text between them, and searched: `IEEE` 0 hits, `bracket` 0, `numbered` 0, `column` 0, `layout` 0, `caption` 0, `figure` 0, `heading` 0, `abstract` 0, `citation` 0. On the same pass `prose` returned 12 hits, `qualifying` 12 and `PDF` 7, so the instrument was working.

The other direction does not come back empty, and what is in it deserves counting exactly. IEEE's editorial style manual runs to 141,011 characters of extracted text, and `artificial intelligence` occurs twice in it. The first is a disclosure rule, quoted in full further down. The second is a line in the abbreviations appendix: `GenAI generative artificial intelligence`. One rule and one glossary line. Not two rules. Immediately under the rule IEEE prints its own sample acknowledgment wording, which is where `AI-generation` and `<AI system used>` come from, so those are the next things you land on if you go searching the file yourself. One warning if you do: the text layer in that PDF drops a lot of the spaces between words, so `artificialintelligence` finds both hits and the version with the space in it may find neither. `Turnitin` returns 0 hits, `plagiarism` 0, `AI detection` 0, and the two hits for `detector` are both entries in that same appendix, one of them `maximum-likelihood sequence detector`. So IEEE has a position on you using AI and no position on anyone detecting it. Those get confused constantly and they are separate problems.

Start with the rule almost nobody quotes, because it is buried in the References section of the IEEE Editorial Style Manual for Authors:

`Do not say "in reference [1] …"; rather, the text should be written to read simply, "in [1]…" The author's name should not be included in a text reference with a number (i.e., "In Smith [1]") and should be changed to "in [1]" except in such cases where the author's name is integral to the understanding of the sentence (e.g., "Smith [1] reduced calculated time …").`

Two sentences later the same section removes the other identifier:

`Reference dates should not be used as reference identifiers and should be deleted in text except in rare cases where the date is somehow relevant to the article's subject.`

Now compare the same paragraph in two conventions. In an author-date discipline, a related-work paragraph is full of proper nouns and years: Zhang and Okoye (2021) did this, Kumar (2023) found that, and each sentence opens differently because it opens with a different name. In IEEE, the manual has taken the names out and the years out and left you with numbers in brackets. What is left to vary is the verb.

There is one more rule that produces the long bracket runs specifically:

`The numbering of references is employed by citing one reference per number. Every reference in a Transactions reference list should be a separate number entry. Use of one reference number to designate a group of references is not permitted.`

That is why a survey sentence ends `[12], [13], [14], [15]` rather than with one bracket covering the group. The conference template restates it, and adds a rule about where the punctuation goes:

`The template will number citations consecutively within brackets [1]. The sentence punctuation follows the bracket [2]. Refer simply to the reference number, as in [3]—do not use "Ref. [3]" or "reference [3]" except at the beginning of a sentence: "Reference [3] was the first ..."`

None of this is a complaint about IEEE. Numbered citation is compact, and compactness is the whole point in a page-limited venue. It is a statement about which parts of your related-work section were your decision.

The style manual's list of article parts puts a number on it directly: `Abstract, must be one paragraph and between 150 to 250 words.` The Abstract section then adds the constraints:

`In order for an Abstract to be effective when displayed on IEEE Xplore as well as through indexing services such as Compendex, INSPEC, Medline, ProQuest, and Web of Science, it must be an accurate, standalone reflection of the contents of the article. They shall not contain numbered mathematical equations, numbered reference citations, nor footnotes.`

IEEE's conference-side guidance says the same in different words, asking for a single paragraph up to 250 words that is `self-contained with no abbreviations, footnotes, references, or mathematical equations`.

Read that as a description of a text object rather than as advice. Your abstract is a fixed-length single paragraph, required to reflect the article accurately and to contain nothing else, with every device that would break up prose explicitly barred from it. No equations. No brackets. No footnote markers. No headings. It is pure continuous prose, and it is a condensation of a document you already wrote.

Turnitin's own list of what its false positives tend to involve includes `text that has been paraphrased without developing new ideas`, and the sentence after it advises the reader to take that into account when looking at the percentage. Those two documents are not talking about each other, and the vendor gives no frequency for that property, so nobody can tell you how much of anything it accounts for. But an IEEE abstract is that shape because a manual made it that shape.

Turnitin defines what it reads as prose sentences inside paragraphs, and says the percentage `is not necessarily the percentage of the entire submission`. Applying that definition to IEEE furniture is our mapping, not a vendor ruling.

The last row is worth being precise about, because it is the row people get backwards. The only exclusion Turnitin publishes for the AI writing report is the bibliography, and it comes from an August 2023 release note saying a bug that highlighted AI writing inside references was fixed and that `Bibliographies are now excluded when processing the AI writing report.` The same note says a submission has to be resubmitted before that applies to it.

What does not follow from that is any exclusion for the `[12]` sitting in the middle of your sentence. Searching those three pages, `citation` returns 0 and `bracket` returns 0. A sentence with a bracket in it is still a sentence. The broader version of this confusion, including which exclusions belong to the Similarity Report rather than to this one, is in Why is Turnitin flagging my references and citations? What the exclusions actually cover and What Content Does Turnitin AI Skip? Non-Prose, Code, and More.

One arithmetic consequence. Strip the headings, captions, equations, index terms, first footnote and reference list from a six-page conference paper and the remaining prose is a lot less than the page count implies. The vendor's file requirements say a submission needs at least 300 words of prose in a long-form writing format before a report is produced at all.

This is the part that costs nothing to check and that almost nobody checks.

A PDF stores glyphs at coordinates. It does not store your reading order. Getting text back out is a reconstruction, and on a two-column IEEE layout the reconstruction has to guess where one column stops.

To see what that looks like on a real file rather than in the abstract, we took a five-page ICASSP 2026 paper from arXiv, `arXiv:2608.25794v1`, and ran `pdfplumber` over it on 27 August 2026. The whole document came back as 26,572 characters and 2,162 whitespace-separated tokens. Lines five through twelve of that output alternate between the abstract in the left column and the introduction in the right one, one line each, so that a sentence from the abstract is immediately followed by an unrelated sentence from the introduction. Word spacing does not survive either: the title comes out as `COOPERATIVEMULTI-AGENTREINFORCEMENTLEARNINGFORADAPTIVE`. The figure on page two contributes tokens like `(cid:16) (cid:17)`, 29 of those across the file. In the reference list, entry `[3]` and entry `[15]` occupy the same lines.

Two honest limits before anyone runs with this. That is one extractor on one file, and Turnitin does not publish which extractor it uses or how it handles multi-column layouts, so nothing here says a Turnitin report on that paper would look any particular way. And the citation brackets themselves came through cleanly: 51 bracketed tokens, including the ranges written as `[3,4]` and `[17,18]`.

What it does say is that the assumption underneath every conversation about a score is worth verifying. Open your submitted PDF, select all, copy, paste into a plain text editor, and read the first screen. If the abstract and the introduction are interleaved there, that is the shape of the thing anyone downstream is working from.

Turnitin's own file requirements page has a line that touches this from the other side, in its best-practice note rather than in a rule:

`When using Turnitin's AI writing detection for assignments, it's best practice to have all students submit their work in the same file format (e.g., all .docx or all .pdf). This helps maintain uniformity in how submissions are processed.`

So the vendor considers file format enough of a variable to recommend consistency across a class. Its accepted types are `.docx, .pdf, .txt, .rtf`. If your institution's assignment accepts a DOCX, sending that instead means nobody downstream has to work out where a column ended, because a word processor file already stores the text in reading order. If you are compiling from LaTeX, that choice is not always available, which is worth knowing rather than discovering afterwards. Why Zotero and EndNote exports sometimes miss the required style covers the neighbouring problem of what your reference manager emits into that file.

Being fair to the manual: it is a house style, and it says so. Its own editing philosophy section states that IEEE does `a mechanical edit` and that `we do not try to change an author's style of writing`. The rules quoted above are about citation grammar, labels and lengths, not about how you argue.

So the paper splits, and the split is usable.

There is one more thing IEEE prescribes that belongs here, because it is the question people actually mean when they ask about AI and a paper. The style manual carries a disclosure rule, and it is specific about where the disclosure goes:

`The use of content generated by artificial intelligence (AI) in an article (including but not limited to text, figures, images, and code) shall be disclosed in the acknowledgments section of any article submitted to an IEEE publication. The AI system used shall be identified, and specific sections of the article that use AI-generated content shall be identified and accompanied by a brief explanation regarding the level at which the AI system was used to generate the content. The use of AI systems for editing and grammar enhancement is common practice and, as such, is generally outside the intent of the above policy. In this case, disclosure as noted above is recommended.`

Read the last two sentences slowly, because the verb changes. Text, figures, images and code `shall be disclosed`, and the disclosure has to name the system and the sections. Editing and grammar enhancement is carved out as common practice, with disclosure `recommended` rather than required. Whether a given revision falls on one side or the other of that line is a judgement you have to make and, if it is close, ask about. It has nothing to do with what any report says about your paper.

Prescribed, and it should stay prescribed. Bracket format and citation order. `Fig.` at the start of a sentence. Roman-numeral section labels. `in Section II-A`, with the manual explicitly telling you not to write `Subsection`. The abstract length. Reference entries in IEEE format. Also, on the conference side, the template's guidance text, which the templates page warns about in one sentence: `Failure to remove template text from your paper may result in your paper not being published.` That is the one item on this list with a stated consequence attached, and it is worth a search before submission rather than after.

Yours, and usually the thinnest part of the draft. Why this problem and not the adjacent one. What the prior approach could not do, stated concretely enough that a reader could disagree. What your result does not cover. What you would measure next and why. IEEE's conference guidance asks the introduction to start with the state of research, progress to your specific topic, and `End with a description of the exact question or hypothesis that your paper will address`. That is a three-move shape, not a script, and the moves have no prescribed wording at all.

There is a related point that has nothing to do with any score. A related-work paragraph that is one sentence per reference, each saying that somebody proposed something, is doing less work than the same space spent on why those approaches divide the way they do. That is a better paper to read and a better paper to review, and it is also the part where sentence shape has somewhere to go.

The slow part of revising an IEEE paper is not writing. One touched paragraph can carry four bracket citations, a cross-reference in the form `[1, eq. (8)]`, two `Fig.` callouts and a `Section II-A` pointer, and every one of those has to be checked against the current numbering before the paragraph can be trusted. Renumber one reference and the checking spreads. Count the cost in paragraphs opened, not in words changed.

So scope is the thing to settle first. Ours is HumanPen, and what follows is a first-party description of where the boundary sits, not a recommendation. Two ways to draw it: add the passages yourself, or hand over the AI report you already have from Turnitin or iThenticate and let its flagged passages do it. The site states the rule as a paragraph being the smallest unit the engine rewrites. Mark half of one and the boundary widens out to the paragraph edge. What you then approve is that widened version, and nothing past it gets opened. Billing sits on the same boundary and counts only the words actually rewritten. Terminology, citations, structure, layout and styles are on the list of things the rewrite aims to preserve, the file comes back in the format it went in as, and the site's own instruction is to review complex documents after download. A passage that qualifies and is still flagged can be re-run at no charge.

For a paper full of numbered cross-references, "review after download" means a specific check rather than a read-through: every bracket, every `Fig.` and `Table` callout, every `Section` pointer, against the numbering in the file you are actually submitting. The document-level version of what breaks when a file makes a round trip, fields and cross-references included, is in Word fields, tables of contents and cross-references.

No part of this predicts a number. Turnitin says its model `may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student`, and the sentence after that one asks for human judgement on top.

Do the `[1]` citations in my sentences get counted? Nothing published says they are excluded. The only exclusion the vendor states for this report is the bibliography, from an August 2023 release note, and in the read described above `citation` and `bracket` return zero hits across the three pages that define the report. A sentence containing a bracket is a prose sentence.

Is my reference list scored? The same release note says bibliographies are now excluded when the AI writing report is processed, and adds that an existing submission has to be resubmitted before that applies to it. So a report generated before that change is not the same measurement as one generated after.

Should I submit the DOCX or the PDF? That is usually decided by whoever is collecting the file. What you can control is knowing what your PDF copies out as before you hand it over. Turnitin's own best-practice note is that a class should submit in one consistent format because it helps maintain uniformity in how submissions are processed, and its accepted types are `.docx, .pdf, .txt, .rtf`.

My related-work section is one continuous highlight and I wrote every sentence. That combination is what the vendor's own false-positive paragraph describes, and its closing sentence advises the reader to take the nature of the text into account when reading the percentage. It does not settle authorship in either direction. What your instructor is told to do when your AI percentage is high collects that guidance.

Is a conference paper treated differently from a journal article? The mechanism is the same, but the composition is not. A page-limited conference paper has proportionally more figures, tables, equations and captions, so proportionally less qualifying prose, which is where Conference Papers and Turnitin AI Detection: What to Know and Short Documents and the All-or-Nothing Problem in Turnitin come in.

Do I have to declare that I used a rewriting tool? That is a disclosure question, and it is separate from anything a detector reports. IEEE answers it directly: content generated by AI `shall be disclosed in the acknowledgments section`, with the system named and the affected sections identified, while AI used for editing and grammar enhancement is carved out as common practice with disclosure `recommended` rather than required. Nothing here tells you which side of that line your own revision falls on, and we are not in a position to tell you either. Read your target venue's own wording rather than a summary of it. Publisher AI disclosure policies, compared puts the major ones side by side.

Do my equations affect the score? Turnitin says it does not reliably detect AI-generated text in non-prose, and displayed equations are not prose sentences. The paragraph introducing an equation is. How Turnitin Handles Math Equations and Non-Prose Content goes through that in full.