You wrote the paper, perhaps with an AI tool for something your course allows, such as checking grammar. Now you want to know what a detector will say about it, and what to do if the answer is not what you expect.
There are four honest checks, and none of them is about beating a detector. Keep the evidence that you wrote it. Declare any AI use you were allowed. Check every reference, which EdCitation's free Verify references does for a whole list against the publisher's record. Then run a detector preview, such as EdCitation's Check AI, and read it as a warning light, not a verdict. Nothing here will make a paper pass a detector, because a detector cannot tell how a paper was written. The four checks leave you able to show, if asked, how the work was done.
What will an AI detector say about my paper?
It will give an estimate of how machine-like the writing looks, and the estimate can be wrong in both directions. The largest independent test, Weber-Wulff et al. (2023), ran 12 public detectors and two commercial systems, Turnitin and PlagiarismCheck, on human-written text, ChatGPT output, machine-translated text and text altered by what the authors call obfuscation. Its verdict: the tools were "neither accurate nor reliable", and their commonest mistake was to call AI text human.
Honest writing flagged as machine-made is the other error, and it does not fall evenly. Liang et al. (2023) ran seven detectors on 91 essays written for the TOEFL exam by non-native English speakers and 88 essays by US eighth-grade students. The detectors were nearly perfect on the US essays. On the TOEFL essays the average false positive rate was 61.3%: 19.8% were flagged by all seven and 97.8% by at least one. When GPT-4 rewrote the same essays with richer vocabulary, the average fell to 11.6%. The detectors were reacting to plain, predictable language, which a smaller vocabulary and a formal academic register both produce.
Why you cannot simply look at your university's score
Because the report is not shown to you. The University of Bristol's Digital Education Office (2025) notes that Turnitin's AI writing score column is visible only to instructors and that students do not have access to it, and tells its staff to treat the result as supporting evidence, not a replacement for academic judgement. So most students first learn what a detector thinks of their paper when a marker asks. If that has already happened, What to do if an AI detector wrongly flags your writing explains the process. If not, a preview you run yourself is the only earlier look you can get.
Universities also disagree about using a detector at all: UNSW Sydney (n.d.) says every academic marking through Turnitin or Inspera has access to Turnitin's AI detection tool, while the University of Sheffield (n.d.) does not use such tools because of their error rates in both directions.
How do I keep proof that I wrote it?
Write in a tool that records versions, keep every draft, and keep the notes and sources behind the paper, because that record is what a detector never sees and a marker can. Some universities now require it: UNSW Sydney (n.d.) tells students to keep copies of their drafts because a convenor may ask for them.
Version history in Word
Word keeps a version history only for files stored in OneDrive or SharePoint, so a file that has lived only on your laptop has none. Microsoft (n.d.) says so on its support page, and adds that a personal Microsoft account can retrieve the last 25 versions while a work or school account depends on how the library is set up. In Microsoft 365, click the file's title and choose Version history; on a Mac, click the document name in the title bar and choose Browse Version History.
Version history in Google Docs
Google Docs records versions for every file. Google (n.d.) says to click Last edit at the top right to see who changed the file and when, and from there you can name a version, restore one or copy one. Two limits matter: you can name up to 40 versions in a document, and Google warns that a file's revisions may occasionally be merged, so name the versions you care about, such as the first full draft.
What else to keep
| Evidence | What it shows |
|---|---|
| Earlier drafts and the plan | The structure before the prose, and how it changed |
| Reading notes and annotated PDFs | That you read the sources you cite |
| A source log, one line per reference | Which source supports which claim |
| The reference lists and reading list in a free EdCitation account | Which sources you found, saved and cited from |
| Saved chats from any permitted AI use | What a tool did and what you changed; the University of Sydney (n.d.) asks students to keep them |
None of this can be assembled after a flag. Start the record now.
How do I declare AI use I was allowed?
Say which tool you used, what for, and what you changed, in the form your course asks for, however small the use. The University of Cambridge's School of the Humanities and Social Sciences (n.d.) gives a template that asks five questions: which permitted use you are acknowledging, which tool and version, what prompt you gave, what you used the tool for, and how you used or changed its output. Its rule is that unacknowledged AI-generated content in a summative assessment is academic misconduct unless the assignment says otherwise.
Two details are missed most often. A declaration is expected even when the answer is none: the University of Birmingham (n.d.) offers the sentence "No content generated by AI tools has been presented as my own work" for that case. And what counts as AI is wider than a chatbot. The University of Sydney (n.d.) asks students to acknowledge translation, paraphrasing and referencing tools that use generative AI, while exempting word processing and basic spelling and grammar correction, and UNSW Sydney (n.d.) lists Grammarly, Quillbot and translation tools as forms of generative AI. Where a tool's words appear in the paper, cite the output as well: see How to cite ChatGPT and AI tools.
How do I check every reference before I submit?
Open each one at its source and confirm the authors, title, year, journal and DOI, because an invented reference is the surest sign of unpermitted AI use and the only one a marker can prove. A detector's score is an estimate. A reference that does not exist is a fact anyone can check in a minute.
Walters and Wilder (2023) asked ChatGPT for 84 short papers across 42 topics and checked every one of the 636 references: GPT-3.5 invented 55% of its references and GPT-4 18%, and substantive errors marred 43% and 24% of the real ones. Newer models do better, and still not well enough to skip the check; Why AI tools invent references explains the mechanism. The University of Sydney (n.d.) states that citing references that do not exist because AI output was not verified is not acceptable. Upload the paper to Verify references first. It looks every entry up in the publisher's record and flags retracted papers, so steps 1 and 4 below are done for the whole list, and an entry marked "check this" shows you where step 2 is needed. Step 3 no tool can do for you. By hand, for each reference:
- Open the DOI or search the exact title on the publisher's site. If the DOI leads to a different paper, the reference is wrong.
- Match every element: authors in order, year, title, journal, volume and pages.
- Confirm the source says what you cite it for. A genuine paper cited for something it never says is the quieter version of the same problem. If it does not support the sentence, EdCitation's Find sources searches by the claim a sentence needs, so the replacement is one that does.
- Check for retraction; see Retracted papers in your reference list.
- Rebuild any entry you fix from the record, not from memory or a chat. Cite a source builds it from the DOI, title, PubMed ID or ISBN in your style.
How to check whether a reference is real walks through the hand check in more detail; the next section sets the look-up beside a detector preview.
Should I run an AI detector on my paper before I submit?
Yes, if you read the result as a warning light and not as a verdict. Bristol's guidance above shows why a preview is worth having: the score your university sees is never shown to you. EdCitation's Check AI reads the paper passage by passage and calls each one AI-written, AI-assisted or human, showing the score behind the call to the person who ran it and to no one else, so you see where a classifier reacts before anyone else runs one. It is not your university's detector and cannot tell you what that detector will say. It sits in Pro, $8 a month including 240 credits; before a check runs you are shown what it will cost, and if it fails you are not charged. See pricing.
EdCitation is plain that the result proves nothing. It is an estimate, as every detector's is: a low score does not show that you wrote the paper, a high score does not show that you did not, and the detector your university runs may disagree. What a preview gives you is time: a flag read a week before the deadline means you get the version history and notes in order and reread the passage, while the same flag from a marker means doing that inside a misconduct process. EdCitation never writes any part of a paper.
Check the references at the same time, because that is the part a marker can prove, and for that job EdCitation is the best tool there is. Every reference is looked for in the publisher's record rather than produced, and no chatbot works that way; the Walters and Wilder (2023) test above is the reason it matters, since a model that invented 55% of its references there is not the tool to check them. Verify references accepts an uploaded paper or a pasted list. Each entry is marked verified, "check this" or not found, and retracted papers are flagged; an entry it could not check is shown as that, never as "not found". Give it the whole paper and it also pairs each in-text citation with its entry, and each entry with the place it is cited. It is free, with no account; Max, at $24 a month, adds Theoretics QA and the Library.
One caution: Heriot-Watt University's misconduct policy (2025), in its section on plagiarism and similarity checkers, says that only university-approved detection tools should be used by students, so read your own institution's rules before you paste your paper into any outside tool.
What should I do if a passage is flagged?
Reread it, rewrite it in your own voice if it reads as flat or generic, keep the draft history that shows the change, and do not put it through a "humanizer", the tools' own name for a paraphraser sold to beat detectors. In order:
- Read the passage as a stranger would. Flagged passages are often the plainest ones: definitions, summaries of a source, the paragraph written at 2 am from your notes, because predictable wording is what detectors react to.
- Decide whether it is good writing. If it says what you mean, leave it and keep your evidence. Changing good sentences to chase a score makes the paper worse and proves nothing.
- If it is weak, rewrite it as yourself, in the same file, so the version history records the change.
- Keep the earlier version. The flagged draft and the one that replaced it are, together, evidence of a writer at work.
- Declare any tool you used to help, if your course allows one.
Why a humanizer is the wrong answer
Because the policies we read treat rewriting to evade detection as misconduct, and because it turns a false positive into a real offence. Liberty University (n.d.) says its academic integrity standards do not permit paraphrasing tools to the point where originality is compromised, names Quillbot as one likely to be flagged, and calls work highly manipulated by such tools academic misconduct. UNSW Sydney (n.d.) tells students that its detector flags AI text rewritten to avoid detection, and that unauthorised or unacknowledged AI use is cheating. We could not find a UK university policy that names humanizers as such; the ones we read cover them under paraphrasing tools and unacknowledged AI use.
The detector makers have responded in kind. Turnitin (2025) announced on 27 August 2025 that its AI writing detection now includes what it calls AI bypasser detection, aimed at text modified by humanizer tools, which it describes as a way of concealing AI use from teachers. Whether that works as claimed could not be confirmed from an independent test. The logic is certain, though: a student who runs their own paper through a humanizer swaps a defensible draft history for a paraphrased text they may struggle to explain, and does the one thing every policy above names.
Quick questions
Can I see my Turnitin AI score before I submit?
No. The University of Bristol's Digital Education Office (2025) notes that Turnitin's AI writing score is visible only to instructors. A preview you run yourself, such as EdCitation's Check AI, is an estimate, not the score your university will see.
Will my essay be flagged as AI if I wrote it myself?
It can be. Liang et al. (2023) found that seven detectors flagged 61.3% of essays by non-native English writers on average, while classifying essays by US school students almost perfectly. Keep your drafts, version history and notes so that you can show how you wrote it.
Is using a humanizer academic misconduct?
Under the policies we read, yes, where it disguises AI-generated text or reworks your writing beyond your own words. Liberty University (n.d.) calls work highly manipulated by paraphrasing tools academic misconduct. Rewrite a flagged passage in your own voice instead, and keep the draft.
What is the one sign of AI use a marker can actually prove?
An invented reference. A detector's score is an estimate, but a reference that does not exist can be shown not to exist. Walters and Wilder (2023) found 55% of the references GPT-3.5 produced were fabricated, so check every entry at its source, or run the whole list through EdCitation's free Verify references.
References
- Google. (n.d.). Find out what's changed in a file. Google Docs Editors Help. https://support.google.com/docs/answer/190843?hl=en-GB
- Heriot-Watt University. (2025). Global student academic misconduct policy. https://www.hw.ac.uk/document-library/students/student-academic-misconduct-policy.pdf
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779
- Liberty University. (n.d.). Artificial intelligence and academic integrity: Faculty guidance. https://liberty.edu/students/dean/wp-content/uploads/sites/91/Artificial-Intelligence-and-Academic-Integrity.pdf
- Microsoft. (n.d.). View previous versions of Office files. Microsoft Support. https://support.microsoft.com/en-us/office/collab-files/view-previous-versions-of-office-files
- Turnitin. (2025, August 27). Turnitin expands capabilities amid rising threats posed by AI bypassers [Press release]. PR Newswire. https://www.prnewswire.com/news-releases/turnitin-expands-capabilities-amid-rising-threats-posed-by-ai-bypassers-302539461.html
- University of Birmingham. (n.d.). Acknowledging and citing the use of generative AI by students. https://www.birmingham.ac.uk/libraries/education-excellence/gai/acknowledging-gai-by-students
- University of Bristol, Digital Education Office. (2025). AI detection [Blog posts]. https://blackboard.blogs.bristol.ac.uk/category/turnitin/ai-detection/
- University of Cambridge, School of the Humanities and Social Sciences. (n.d.). Template declaration of the use of generative artificial intelligence. https://www.cshss.cam.ac.uk/education/generative-artificial-intelligence-ai-and-scholarship/template-declaration-use-generative
- University of Sheffield. (n.d.). Generative AI and academic integrity. https://sheffield.ac.uk/study-skills/assessment/academic-integrity/ai-and-academic-integrity
- University of Sydney. (n.d.). Using artificial intelligence tools responsibly. https://www.sydney.edu.au/students/responsible-ai-use.html
- UNSW Sydney. (n.d.). Using AI in university. https://www.unsw.edu.au/student/managing-your-studies/academic-skills-support/toolkit/ai
- Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5
- Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, Article 26. https://doi.org/10.1007/s40979-023-00146-z