Writing in your own voice when everyone suspects AI comes down to three habits: write with the words you already use, argue from detail only you have, and keep a dated record of how the paper grew. This is EdCitation's view: the answer to suspicion is a record, not a new style. Changing how you write to please a detector costs you your voice and proves nothing, since no wording can guarantee what any detector will say.
EdCitation publishes this guide. It draws on four universities' writing guides, six universities' integrity procedures, the ombuds for student complaints in England and Wales, and two detector studies, each read on 24 September 2026. Why a dash proves nothing is argued in the em dash is not proof of AI, and second-language writers have a guide of their own. This one is about the writing and the record.
One part of the record can be settled tonight: whether every source exists. EdCitation's free Verify references looks each one up in the publisher's record.
What does "your own voice" mean in academic writing?
Your own voice is the part of a paper that belongs to you: the claim, the analysis and the judgement you make about your sources, in the words you actually use. It can be formal without being borrowed.
Your voice among your sources' voices
The University of Hull's guide to critical writing names four voices in academic prose (University of Hull, 2026). Your own voice makes a point or evaluates other people's work; the other three report sources, by summary with a citation, by summary naming the author, or by quotation. Good writing blends them, and referencing lets a reader tell your ideas from everyone else's. The sciences, Hull notes, lean on summary and the writer's own voice and rarely quote. A paper made only of summary has no voice to recognise.
Formal, in words you already know
Vanderbilt University's Writing Studio (2008) says academic voice does not mean stiff or robotic prose, and that "you can write a formal paper with the words you already know." It suggests reading a draft aloud to catch sentences knotted by the effort to sound formal, and warns that readers can usually tell when a writer uses a word they do not know.
Where AI tools stand
The University of Central Florida's writing centre (2025) leaves AI tools to the writer and the course, with one line: generative AI "should never replace your ideas, language, or voice." Your course's AI policy is the rule.
Why does your voice matter if a paper is questioned?
Because markers compare the paper with your earlier writing and with what you can explain. University guidance for markers names what they look at, which tells you what to keep.
| What a marker may look at | Named by | What you can keep |
|---|---|---|
| A change of style or voice against earlier work | University of Bath (2024); University of Liverpool (2024); University at Buffalo (n.d.) | Earlier marked work, and the feedback that changed your writing |
| Whether you can explain the paper and how you made it | Buffalo; Royal Holloway (n.d.) | Your plan, notes and reading |
| False references | Liverpool; Bath | A list checked in Verify references |
| Drafts, search history, file metadata | Royal Holloway; Office of the Independent Adjudicator for Higher Education (2025) | Version history and dated drafts |
Bath's list adds a caution worth knowing: students are expected to improve as they act on feedback, so a change of style has to be significant before it counts. If your writing changed because you followed a tutor's comments, that feedback is evidence.
Where universities disagree
On detectors, they point opposite ways. Buffalo suggests instructors use Turnitin's AI detection, which it calls highly reliable, though a case needs more than that report. Bath forbids its staff to put students' work into AI detectors. Sydney's student page says Turnitin's tool may be used, never as the only evidence (University of Sydney, 2026), while its teaching blog says its educators have never had access to detection tools (Bridgeman et al., 2025). Read your own university's current page.
How do you keep a record of your writing as you go?
Write where version history is kept, name a version at each stage, keep dated notes, and touch nothing after you submit. Royal Holloway's guidance for students facing an AI allegation lists drafts, screenshots of search history and research notes as evidence of authorship, and Loughborough University (n.d.) tells students to save drafts to show how their own ideas developed.
What the software keeps
History depends on where the file lives. Microsoft (n.d.) says Word's version history works only in OneDrive or SharePoint, which leaves a document kept on a laptop without one. A Google Docs history can be browsed only by someone with permission to edit the file (Google, n.d.), so a view-only link shows a reader none of it. Where to click is in how to check your paper for AI before submitting.
The procedure, from the first day
- Start one file on day one, in Google Docs or a Word file kept in OneDrive, with the plan at the top.
- Name a version at each stage: plan, first draft, the draft after feedback, the submitted text. A personal Microsoft account keeps only the last 25 versions of a Word file, so save a dated copy at each stage too.
- Build references as you read in Cite a source. With a free EdCitation account, everything you cite stays in a reference list, next to a reading list.
- Keep notes where they carry dates: annotated PDFs, dated notebook pages, saved searches.
- Log any tool your course allows, with the passage before and after, and declare it; how to declare AI use gives wording.
- Keep the brief and each piece of feedback you acted on.
- Before you submit, run the list through Verify references, and export the named versions as a PDF or screenshots: Royal Holloway cannot open HEIC photos or links to documents.
- After you submit, leave everything alone. Never tidy, rebuild or back-date a draft: Royal Holloway disregards evidence it suspects is not authentic, and may impose a heavier penalty.
Does a record prove you wrote it?
Not on its own. Sydney's teaching blog says students should not have to prove their innocence, that a lack of drafts cannot be read as guilt, and that common AI tools can easily generate drafts and revision histories (Bridgeman et al., 2025). Royal Holloway calls sending evidence of authorship before the meeting very important. Keep the record either way: where the burden sits with the university it costs nothing, and where a panel asks, it is what they want.
How do you put detail only you have into a paper?
By arguing from the particular: the passage you read closely, the data you collected, the point where two sources disagree, what you changed after feedback. That detail makes a paper worth reading, and you can talk about it when asked. It is not a way past a detector.
What the writing centres allow
The University of North Carolina at Chapel Hill's Writing Center (n.d.) treats "I" and personal experience as separate questions. Experience has a place in a response to a work of art, in showing how a theory applies to life, or as evidence for an abstract principle, provided it serves the argument. It fits less well in the sciences, history and anthropology. When the brief is silent, ask.
Detail you can source
Most of the detail that makes a paper yours is not personal. It is the exact figure from the exact study:
| A sentence anyone could write | The same point, with detail you chose |
|---|---|
| Detectors are often unfair to some students. | Seven detectors wrongly flagged 61.3% of 91 human-written TOEFL essays on average (Liang et al., 2023). |
| Studies show detectors are unreliable. | Six of the 14 tools Weber-Wulff et al. (2023) tested labelled human-written texts as AI, more often after machine translation. |
The right-hand column is harder to write and easier to defend. Royal Holloway's panels ask which sources influenced you most, and Bath suggests asking what a student read that shaped their approach.
What do the studies on detector false positives show?
In two well-known tests of 2023, detectors labelled some human writing as AI-written, most often when the vocabulary was narrower or the text machine-translated. Both are small, both tested 2023's tools, and neither can say what a detector will make of your paper.
| Study | Human writing tested | Finding | Limits |
|---|---|---|---|
| Liang et al. (2023), Patterns | 91 TOEFL essays; 88 essays by US eighth-graders; seven detectors | US essays almost all correct; TOEFL essays wrongly flagged 61.3% of the time on average, 19.8% by all seven detectors, 97.8% by at least one | Small samples; groups differ in more than first language |
| Weber-Wulff et al. (2023), International Journal for Educational Integrity | 9 English texts never posted online; 9 machine-translated from seven languages; 14 tools, Turnitin among them | Human texts 96% correct, translated ones about 20% less; six tools gave false positives; none reached 80% accuracy overall | 54 test cases; tools as of March to May 2023 |
Weber-Wulff et al. (2023) concluded that the systems they tested should not be used in academic settings, and that written assessment should focus on how a student's skills develop rather than on the final product.
The warning inside the 2023 study
When ChatGPT enriched the TOEFL essays' vocabulary, the average rate of false positives dropped to 11.6%, and making the US essays' wording simpler pushed misclassification up. The authors did not offer that as advice. They warned that detectors might push non-native writers to use AI more, to "sound more native", and strongly cautioned against detectors in education.
In our view that settles the question of style. Inflating your vocabulary to look human repeats what the study did with ChatGPT; Vanderbilt's rule, the words you know, is better. The vendors' own figures are in what to do if an AI detector wrongly flags your writing.
What should you do if your writing is questioned?
If you are accused of AI use, read the whole case, send your record before the meeting in a form the panel can open, and prepare to discuss the whole paper, not only the flagged passages. Your own university's regulations govern your case.
Before the meeting
- Read the case file. At Royal Holloway it holds the marker's statement and evidence, and may include your other work for comparison.
- Ask what is compared. The ombuds says universities should name the earlier work they are comparing, so you can explain how your approach differed, and explain what they think file metadata shows, with a chance to respond.
- Send evidence on time. Royal Holloway wants it by 5pm the day before, as documents, not links.
- Bring someone. Liverpool lets a student be accompanied by another member of the university; Royal Holloway allows a students' union adviser, a fellow student or staff.
In the meeting
Royal Holloway's panel asks how you found and used your sources, how you wrote and referenced, and which sources influenced you most, and asks you to expand on your ideas well beyond the marker's points. Bath says a viva should test whether a student understands what they wrote and how they reached their conclusions, not whether they recall it. Afterwards, Royal Holloway gives three days for any further evidence and seven to check the meeting notes, and decides on the balance of probabilities.
How can EdCitation help you keep your writing yours?
EdCitation checks what can be looked up, and it never writes any part of a paper. That matters here: a tool that rewrote your sentences would hand your voice to someone else.
Verify references is the best check there is for the one point in an authorship case that a record can settle, whether your sources exist: each entry is matched against the publisher's record and none is ever generated, which is more than any chatbot can promise. On 24 September 2026 we gave it Liang et al. (2023) as a writer might cite it from memory, with the year, the journal and the title wrong and three of the five authors left out. The verdict was "check this", and the reason: "A work with this title is recorded in Patterns (2023), not in the source your entry names. Check your source." The three verdicts are verified, "check this" and not found; a retracted paper is flagged, and a reference the records could not answer for is marked "could not check", never "not found". Upload the whole paper and the citations in your text are checked against the list too. Free, with no account.
Cite a source then rebuilt that entry from the record. From the DOI it returned the APA 7 reference from Crossref, Patterns, volume 4, issue 7, Article 100779, with no retraction notice, and five other styles.
Find sources takes the claim a sentence has to carry and searches about 300 million published works. "AI detectors misclassify essays by non-native English writers" brought back 239 works, with the Liang study at the top.
Check your paper turns a brief into a checklist, free. From our sample brief for a placement essay it drew nine rules, each shown with its sentence, among them 1,800 to 2,200 words, six or more sources from the last ten years, APA 7 and the required headings. The lines on handing in an outline and draft, and on using AI only for grammar, went under "What it asks you to do, which no rule can test". That is fair: they concern your process, which only your record shows. Judging the paper against the checklist is Theoretics QA, part of the $24 Max plan.
Pro, at $8 a month, adds Check AI. It marks each passage of your text AI-written, AI-assisted or human, each with its score, in a report only you see. Pro includes 240 credits a month; you see what a check will use before starting it, and one that fails uses nothing. Its reading is one detector's: the one your university uses may disagree, and a score proves nothing either way about who wrote a passage. Pricing has the plans.
Quick questions
What does writing in your own voice mean in an essay?
The claims, analysis and judgements are yours, in words you actually use. Vanderbilt's Writing Studio says a formal paper can be written with the words you already know.
Does version history prove I wrote my essay?
Not on its own. Royal Holloway asks for drafts and notes, but Sydney's teaching blog warns that AI tools can generate drafts and revision histories, so a record made as you worked counts most.
Should I use more advanced words so my writing does not look like AI?
No. Detectors in one 2023 test flagged essays less after ChatGPT enriched them, but the authors warned against pushing writers that way, and no choice of words can promise a detector's verdict.
Can I use personal experience in academic writing?
Sometimes. UNC's Writing Center says it can serve an argument about a work of art, a theory applied to life or an abstract principle, and fits less well in the sciences and history.
What can EdCitation check before I hand in?
Whether each reference exists, free and with no account, in Verify references. With Pro, Check AI gives you a private, passage-by-passage reading from one detector, which is a reading and not proof.
References
- Bridgeman, A., Liu, D., Brockley, C., & Bassett, E. (2025, October 30). False flags and broken trust: Can we tell if AI has been used? Teaching@Sydney, The University of Sydney. https://educational-innovation.sydney.edu.au/teaching@sydney/false-flags-and-broken-trust-can-we-tell-if-ai-has-been-used/
- Google. (n.d.). Find what's changed in a file. Google Docs Editors Help. https://support.google.com/docs/answer/190843
- 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
- Loughborough University. (n.d.). Use of generative AI tools. Student Handbook. https://www.lboro.ac.uk/students/handbook/assessments/coursework-feedback/generative-ai/
- 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
- Office of the Independent Adjudicator for Higher Education. (2025). Casework note: Complaints relating to AI and academic misconduct. https://www.oiahe.org.uk/resources-and-publications/learning-from-our-casework/ai-and-academic-misconduct/casework-note-complaints-relating-to-ai-and-academic-misconduct/
- Royal Holloway, University of London. (n.d.). Student guidance to artificial intelligence (AI)-related offences [PDF]. https://intranet.royalholloway.ac.uk/students/assets/docs/pdf/academic-investigations/2526/student-guide-ai-related-allegations.pdf
- University at Buffalo. (n.d.). Artificial intelligence guidance. Office of Academic Integrity. https://www.buffalo.edu/academic-integrity/instructors/protect/ai-guidance.html
- University of Bath. (2024). GenAI and academic misconduct: Guidance for colleagues [PDF]. https://teachinghub.bath.ac.uk/wp-content/uploads/2024/02/Guidance-on-GenAI-Suspected-Misuse.pdf
- University of Central Florida. (2025, December 12). Generative AI in the University Writing Center. https://cah.ucf.edu/uwc/generative-ai/
- University of Hull. (2026, September 4). Your voice. Critical writing. https://libguides.hull.ac.uk/criticalwriting/voice
- University of Liverpool. (2024). Academic integrity process for the misuse of generative AI [PDF]. https://www.liverpool.ac.uk/media/livacuk/centre-for-innovation-in-education/digital-education/generative-ai-teach-learn-assess/academic-integrity-process-for-the-misuse-of-gai.pdf
- The University of North Carolina at Chapel Hill Writing Center. (n.d.). Should I use "I"? https://writingcenter.unc.edu/tips-and-tools/should-i-use-i/
- University of Sydney. (2026, June 2). Artificial intelligence. Academic integrity. https://www.sydney.edu.au/students/academic-integrity/artificial-intelligence.html
- Vanderbilt University Writing Studio. (2008). Academic voice [Handout; web version 2021]. https://www.vanderbilt.edu/writing/resources/handouts/academic-voice/
- 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