Why is Turnitin flagging my references and citations? What the exclusions actually cover

Turnitin's AI writing report excludes bibliographies and analyses only prose written as standard grammatical sentences, so your reference list is neither raising your AI percentage nor available to dilute it. Colour sitting on a bibliography almost always belongs to the Similarity Report instead, where a matched reference entry is what a shared citation style is supposed to produce.

Which leaves three narrower questions that the usual answer skips. Why do some AI reports have colour on reference entries anyway. Whether an annotated bibliography is the same case as a plain reference list. And who, if the colour is on the similarity side, actually holds the control that would change it.

The thread running through all three is that these exclusions are not facts about your document. They are steps in how a report gets made, applied at the moment it is made. Change the moment and you change the report, without touching a word of the file.

Turnitin's AI writing detection model page carries dated release notes. The entry for 9 August 2023 reads:

"We have fixed a bug that was occasionally highlighting AI writing within references listed in a bibliography. Bibliographies are now excluded when processing the AI writing report."

The sentence directly after it is the one that gets dropped every time this note is quoted: "Resubmit to reprocess existing submissions that contain highlighted reference sections." So the exclusion is applied while a report is being produced. A report that already existed on 9 August 2023 kept whatever highlighting it had, and stayed that way until somebody submitted the paper again.

The same note handled tables in the same shape. "We are now able to process long-form prose text in tables", followed by "Resubmit to reprocess existing submissions that contain tables." Two changes, both conditional on a resubmission that nobody performs automatically.

For most people holding a report generated this year, none of that is the explanation for anything. What it is good for is the habit it should install: before you try to account for what a report shows, find out when it was produced, because the rules it was produced under are the rules of that date. Two reports on the same file, generated a year apart, are not two readings of one measurement. We took the general version of that apart in comparing two Turnitin reports after a recheck.

It also explains why so much of the advice circulating about flagged bibliographies sounds confident and dated. A lot of it was written against screenshots from 2023.

Two questions arrive constantly and they have one answer. Did my bibliography push my AI percentage up? And would a longer bibliography bring it down?

Turnitin defines what the model looks at as qualifying text, and its detection FAQs say it "only analyze blocks of text that are written in standard grammatical sentences and do not include other types of writing such as lists, bullet points (short non-sentence structures), or other non-sentence structures", with the consequence stated in the next sentence: "This percentage is not necessarily the percentage of the entire submission." A reference entry is not a standard grammatical sentence, and bibliographies are removed before processing anyway. So the reference list is not in the numerator and it is not in the denominator either.

That kills two plans at once. Deleting your references before submission does not change what the report analyses, and it hands in an incomplete paper to fix a problem you did not have. Padding the bibliography does not dilute anything, because dilution needs the added words to be in the denominator and these are not.

If you want to see what the denominator actually is in a document full of tables and lists, how to read a Turnitin AI writing report works a numerical example through.

The word bibliography is doing a lot of work in all of this, and Turnitin's own pages contain two separate claims about reference-shaped material that people tend to merge into one.

One Turnitin help article says: "The model does not reliably detect AI-generated text in the form of non-prose, such as poetry, scripts, or code, nor does it detect short-form/unconventional writing such as bullet points, tables, or annotated bibliographies."

Read that next to the release note. The release note describes a processing rule: bibliographies are taken out before the model runs. The help article describes a reliability limit: for certain writing types, the model's output is not something Turnitin stands behind. Those are not the same guarantee, and the second one is much weaker.

The distinction bites hardest on annotated bibliographies, because the annotations are prose sentences that you wrote, sitting under entries that you did not. If your annotated bibliography comes back with colour on it, that is not the same event as colour on a plain reference list, and the release note is not the document that covers it.

Turnitin's published description of the calculation names a chain of operations. Sentences are extracted, grouped into overlapping segments, each segment gets a value between 0 and 1, each qualifying sentence inherits the score of the segments covering it, and those pooled sentence scores are aggregated into the document number. A citation is not an input to any step in that chain.

The overlap has a consequence people rarely follow through. A sentence near the edge of one segment sits inside two, so it carries more than one inherited score. The highlight on your citation-carrying sentence is partly borrowed from the sentences on either side of it, which is why the fix, if there is a fix, is never located at the bracket. We traced what does and does not follow from that in does rewriting one flagged sentence lower the score.

It also means several tempting edits have nothing to act on. Renumbering, switching a parenthetical marker for a superscript, adding a signal phrase in front of the claim, moving the attribution to the end of the sentence: none of these change the property being scored, because none of them change the prose.

For the version of this question that starts from "which report am I even holding", why did Turnitin flag my references walks the branches.

Now the other case. Somebody has forwarded you a percentage, the colour is sitting on your bibliography, and you have been asked to bring the number down. The thing that would move it is not in your file.

Similarity reports are produced with exclusion filters, and bibliography is one of them, quotations another. Whoever set the report up chose which were active, and that choice does not travel with a percentage pasted into an email. Two reports on one unchanged file can differ by a lot for that reason alone. The two reports side by side covers the filters and where to find their state; why your quoted and cited text still matches covers what happens to attributed passages when nobody switched them on.

Which is also the argument against the obvious move. A reference entry matches other people's reference entries because the style guide told all of you to produce the same string. An entry rewritten until it stops matching is an entry that stops following the style, and if a tool does it across the whole list at once you will not notice until a marker does. What happens to citations and tables inside a rewriting tool has a recorded case of exactly that.

So the useful reply to the email is a question rather than a revision: which exclusions were active on that report, and where is the colour outside the bibliography. The second half is the part that ever needs an answer from you.

None of this needs an account, an appeal or an expert. It needs the actual PDF rather than a screenshot of a number, and about five minutes.

1. Find the generation date first. Every exclusion and model version described anywhere is the set that was in force when that file was produced, not the set described on the help site today.
2. Look at the AI indicator before you look at anything coloured. Turnitin attributes no score and no highlights in the 1% to 19% band and shows an asterisk instead, so if you are looking at an asterisk, nothing highlighted in front of you came out of the AI report.
3. Follow one coloured passage and see what it offers you. A match that leads to a named source with its own percentage is the Similarity Report. AI highlighting does not appear there and does not lead anywhere.
4. Ask which exclusion filters were active. Bibliography and quotations are separate switches, neither state is visible on a forwarded number, and the person who ran the report knows.
5. If it is the AI report, read the highlighted prose and stop counting lines. The percentage is computed over qualifying text and the highlights sit in a document that also contains everything else, so the two will not reconcile by eye.

Written down, that is a list of questions about the report rather than about your writing, which is roughly the correct ratio for this particular problem.

HumanPen is a document rewriter, and the design decision relevant here is about scope rather than output. You either select the passages yourself, or you upload the document together with its AI writing report and let the flagged passages define the scope. Only the passages matched from the report are rewritten; the rest of the document is preserved verbatim.

For this article's subject that has a specific consequence. A reference list has no reason to be inside the scope of a rewrite in the first place, because it is outside the AI calculation on both sides. A tool that reaches it anyway is spending your risk on something that could not have helped your number.

Billing counts only the words actually rewritten, so a report that flags a fifth of the document costs roughly a fifth of a full pass. Eligible results can continue lowering AI for free.

What we will not tell you is what your next report will say. Our own page says HumanPen aims to preserve meaning, structure, terminology, citations, layout and styles while rewriting only the necessary language, and then says to review complex documents after download. Both halves of that are meant.

My AI report has colour on the reference entries themselves. Is that even possible? It was, before Turnitin's release note dated 9 August 2023 announced that a bug highlighting AI writing inside bibliographies had been fixed and that bibliographies were then excluded when processing the report. The note adds that existing submissions have to be resubmitted before they are reprocessed, so an older report kept its highlighting. If your report is recent, check whether what you are holding is the Similarity Report instead.

Will a longer reference list dilute my AI percentage? No. The percentage is calculated over qualifying prose, and Turnitin says the result is "not necessarily the percentage of the entire submission". Words that are not in the denominator cannot dilute anything.

My annotated bibliography is highlighted. Is that covered by the same exclusion? Not by the same statement. The release note is about bibliographies being excluded from processing. One Turnitin help article separately lists annotated bibliographies among the writing types the model does not reliably detect AI writing in, which is a claim about reliability rather than a rule about what gets removed. The annotations are also prose you wrote, which a plain reference list is not.

Does switching citation style change either number? Not on the AI side, because nothing in the published description of that calculation reads citations at all. On the similarity side, how a report classifies an attributed match depends on a recognition step with documented limits, and why your quoted and cited text still matches goes through them.

What should I ask for before replying to an email about this? The full report as a PDF rather than a screenshot, the date it was generated, which report it is, and which exclusion filters were active. Those four answers settle most of this before anyone has to discuss your writing.