Old Assignments as a Writing Baseline: Checklist, and What It Cannot Prove

No, and the tool built to do exactly this comparison says so in writing. A style baseline is a set of your own earlier, graded assignments, offered to show that the flagged writing looks like the writing you were already handing in. It is a genuine category of evidence: the University of York's academic misconduct policy instructs its panels to consider "the previous assignments submitted by the student for a comparison". But Turnitin sells a product that does this professionally, and Turnitin's own FAQ says that product "will not be able to indicate if it was AI written". A resemblance argument answers the question "does this look like you". The allegation is about how one document came to exist, which is a different question.

So the honest position is that a style baseline is corroboration, not an answer. It sits underneath the process evidence rather than in front of it.

The rest of this covers who is already running this comparison and on what, which pieces of your own work are worth offering, and three ways it can turn against you: a comparison may already have been made without you seeing it, anything you hand over can be put through the detector again, and if your writing has improved, the baseline argues against you. If you take one thing from this page, make it the table under "What this evidence can and cannot carry".

Three published documents, read on 28 August 2026. Two of them tell staff to compare your new work against your old work. The third does something different, and the difference is worth knowing. (On the Western Washington and South Carolina pages the lines quoted below sit inside collapsed panels, so expand them if you go to check.)

There is a documented case of the comparison being run and going wrong. The Office of the Independent Adjudicator reviews unresolved student complaints against higher education providers in England and Wales, and it publishes anonymised summaries of what it decided. Six of them came out in July 2025 under the heading "AI and academic misconduct". In the first, a student with autism denied using AI and supplied their planning and notes. The summary records that "a disciplinary panel considered the student's written evidence and also compared the essay to some other examples of the student's work", and found misconduct.

The ombudsman upheld the student's complaint. One of its reasons:

"The student had not been given a fair opportunity to comment on how this essay compared to their other essays because some of the evidence the panel relied on had not been shared with them."

The student later told the ombudsman that when the provider reconsidered the case, it found there had been no misconduct at all.

That gives you a question to ask before you volunteer anything: has a comparison with my other work already been made, and can I see it? York's policy states the principle from the other side, that the panel "will not use any material to make its judgement unless the student has had sight of it in advance and the opportunity to respond".

The product mentioned at the top of this page is called Authorship, and how it differs from the AI detector is covered in Turnitin Authorship vs AI Detection: Two Different Systems, Explained. What matters here is what it compares against, and the guide written for investigators says it is your back catalogue:

"Authorship for Investigators also provides new evidence about a student's writing over time that can be used to confirm suspicions of a possible misconduct violation. This evidence is provided in an Authorship Report that is generated by either manually uploading files to compare against each other, or by entering a Turnitin paper ID and selecting from a list of the students' past work which files you would like to compare against."

The same page adds a limit in its own words: "It is important to note that Authorship for Investigators does not identify contract cheating. It takes human judgment to determine whether contract cheating has occurred based on the balance of probabilities."

Turnitin's public guide page for that report shows a reader one note: the guidance "is only visible when accessed from within the report itself". So if a comparison was run on your past submissions, the person best placed to explain what it shows is whoever opened the report, which is one more reason to ask to see it.

The AI detection FAQ, as it read on 28 August 2026, never uses the words baseline, style, portfolio or voice. So the detector's documentation makes no claim about your writing style in either direction: nobody can quote it against your style, and you cannot quote it in support of yours.

Neither the York policy nor the Western Washington page names a number. York says "previous assignments", plural, with no count. Western Washington says "a baseline writing sample", singular. So pick on quality rather than to hit a quota.

• Returned and graded, with the marker's comments still on the file. A returned assignment carries a date, a marker and a grade that came from somebody else. A clean copy you retyped carries none of that.
• Match the genre before you match the module. A lab report and a reflective essay are different instruments and they make you write differently. If you can only have one of the two, take the matching type of writing.
• Anything written under invigilation. A timed exam script (if your department will release it), an in-class piece, a supervised test. Authorship is not in question for those, which is the one thing coursework samples cannot give you no matter how many you produce.
• The pieces that have your habits visible in them. Not your best work, your most typical. The point is consistency, and a prize-winning essay from second year proves nothing about how you write on an ordinary Tuesday.
• Anything already assessed inside a formal process. If a previous flag on your work was investigated and dropped, that outcome is a documented fact and is stronger than a resemblance argument, for reasons in the section on your writing getting better, below.

What not to reach for: a piece you cannot locate the graded original of, anything from a course where you worked in a group, and anything you would not want re-read closely. One place a dated run of your own prose piles up without any effort is a weekly online discussion board. Why that format builds one by accident is in Discussion Board Post Flagged as AI.

Western Washington's page names four things: writing style, quality, quantity, and mechanics, with mechanics broken out as spelling, punctuation and structure. Use it precisely because it is somebody else's list. If you build your comparison on those, you are talking in the terms the review is already using.

Three of those four are judgement calls. Mechanics is the closest to countable, and Word will give you two numbers, though as the formula below shows, neither of them looks at spelling or punctuation.

In Word for Microsoft 365, Home > Editor > Document stats reports the Flesch Reading Ease score and the Flesch-Kincaid Grade Level. Microsoft's support page prints both formulas. Reading Ease is:

"206.835 – (1.015 x ASL) – (84.6 x ASW)"

where ASL is average sentence length in words and ASW is average syllables per word. The Flesch-Kincaid formula uses the same two inputs.

So before you put the numbers in an email: both scores are functions of exactly two variables, how long your sentences are and how many syllables your words have. They know nothing about your vocabulary choices, your argument structure, your citation habits or your paragraph shapes. A matching Flesch-Kincaid grade level across four documents is weak corroboration. A mismatching one is not evidence of anything either, because a research-heavy chapter runs to longer sentences and bigger words than a reflective piece, and that difference is about the genre, not about authorship.

Two Word behaviours to know if your documents are not monolingual, both from the same Microsoft page. If a file contains text in several languages, Word "displays readability statistics for text in the last language that was checked". And for some European languages inside an English document, Word "displays only information about counts and averages, not readability".

A style baseline argues that the flagged work resembles your earlier work. Now say that claim out loud with the dates attached. You are two years into a degree, you have had feedback every term, you have read a few hundred papers in your field, and you are claiming that your writing has not moved.

For a lot of people that claim is simply false, and the person reading it can see the grade trend. Worse, it is a claim you may not want to win. Arguing that years of teaching left no mark on your prose is an odd thing to put in writing to your department. (Whether to hold your style steady on purpose from here on is a separate question, taken up in Should You Change How You Write to Avoid Being Flagged?)

There is a second use of old work that avoids all of this, and the ombudsman file shows a student making it. In that first case summary, the student's argument was not "this looks like my other essays". It was that the detection software "had previously flagged another essay they had written for suspected use of AI but in that case it had been agreed that the student had not used AI", which the student said "suggested that the detection software was biased against their writing style".

That is a different argument built from the same material. One says my work resembles my work. The other says this instrument has produced a result about my writing before and that result was set aside. The second one does not require you to claim you have stopped developing, and it points at the instrument rather than at you. It is also narrow: one tool, one document, one occasion, and it does not generalise into a claim about accuracy rates. Our page on why the same text scores differently on every detector covers the mechanism, and Is 20% AI too high? covers why no number works as a threshold.

If you are choosing between the two, the second is usually the stronger one, and it is available only if it actually happened to you.

The baseline you volunteer is a set of documents you are inviting somebody to read closely, and running them again is something Turnitin documents, not a hypothetical. That is not a reason to hold them back. It is a reason to decide before you send them, not after.

There is a related trap in the York policy that cuts the other way for anyone who kept nothing. The same clause that authorises the comparison also says: "Lack of preparatory work may be considered evidence of false authorship." Preparatory work means notes and drafts. A stack of finished old essays does not fill that hole. The material that does is covered in how to keep version history in Word, Google Docs and Overleaf.

1. Ask first whether a comparison already exists, and ask to see whatever it produced. This is not a confrontational request. York's policy, quoted above, says the panel will not rely on material you have not seen, and it was one of the reasons the ombudsman upheld the complaint above.
2. Send the returned files, not retyped copies. The marker's comments, the submission date and the grade are the part that came from outside you. A clean Word document with the same words in it has lost the evidence and kept the text.
3. Name the dimensions yourself. Not "please compare these". Something closer to: sentence length, how I use subheadings, the way I introduce quotations, and my habit of putting the citation at the end of the sentence. Point at three or four things somebody can check in two minutes.
4. Say what has changed since. If your writing has moved, say so before anyone else notices. "The structure is tighter now than in first year, and here is the feedback that changed it" reads as accurate. Silence on the point reads as something else once it is spotted.
5. Put it second, not first. The question in the room is how this document was made. Process evidence answers that; the baseline corroborates it. How to prove your work is original lays out the process layers, and how to appeal a false AI flag covers what universities have accepted.
6. Keep a copy of what you sent and when. If the case goes to a second stage, the record of what was already in front of the first reader matters.

Nothing above is something a rewriting tool can do for you. Assembling a baseline is your work. There is one way a rewriter can work against the argument in this article. The more of the flagged paper that changes, the harder it is to set beside the samples you are about to hand over, so if you revise, keep it to the passages in question.

HumanPen can keep a revision of an English document to those passages: you mark them yourself, or import a detection report and let its highlights set the range (the report route is set out in Rewriting only the paragraphs a Turnitin report flagged), and you confirm the paragraphs before anything runs. It is designed to preserve terminology, citations, tables and layout, and what comes back is still an editable file.

Eligible passages can be re-run at no charge. We do not make claims about what a score will be afterwards.

How many old assignments should I bring? No number appears in the two policy documents quoted above. York says "previous assignments" without a count and Western Washington says "a baseline writing sample" in the singular. Pick on quality: graded, returned, same genre, and typical of you rather than your best.

Can my university run my old essays through the AI detector? Turnitin's FAQ says previously submitted assignments can be checked for AI writing if they are re-submitted and the account has AI writing enabled. So old work is not automatically out of scope, and that is worth knowing before you volunteer a folder of it.

Is there a tool that compares my writing to my writing? Turnitin has one, called Authorship, sold separately from AI writing detection. It compares files against each other or against a student's past submissions, and Turnitin's FAQ says it "will not be able to indicate if it was AI written". Turnitin says its guidance for reading that report is shown from inside the report itself.

My old work also comes back with a high AI score. What now? Then the resemblance argument is not the one to lead with, and the flags on the older work may be a fact about the instrument's behaviour on your prose rather than about any of the documents. You are closer to the second argument in the section on your writing getting better, and that is a conversation about the tool. Non-Native Speaker Writing Flagged as AI collects what the documentation says about the groups where false positives are acknowledged.

Is a style baseline worth doing at all? As a supporting layer, yes, and it is cheap. As the centrepiece of a response, no. It answers whether the writing looks like you, and the question being asked is how this particular document was produced.