A Google Docs Version History Shows Three Patterns, and None Names an Author

A Docs history takes one of three shapes across a stack of student papers. One or two entries hold the finished text. Or the typing runs straight through in one sitting, with nothing deleted. Or the document grows across days, and paragraphs keep getting cut and rewritten. To open the panel, click Last edit at the top right of the document, or File, then Version history, then See version history ¹.

The single arrival is the easiest to see. One entry, or two, holds text that was not there a minute earlier and needs no further work. The writing happened somewhere the panel could not watch. That is all it says. A phone's notes app, a Word file on a laptop, a legal pad transcribed at midnight and a chatbot window all produce the same shape in the same panel. An entry records an arrival, and an arrival has no author attached to it.

The unbroken sitting fills the document steadily between 9 and 11 on a Sunday night, with few deletions and no reordering. That is what a confident writer under a deadline looks like. It's also what retyping looks like, and the panel cannot separate them.

Only the third shape shows the drafting itself.

The drafting curve spreads entries across days. Paragraphs move. A thesis changes in the middle. Whole sections disappear and come back a week later. A 2026 framework in Frontiers in Artificial Intelligence asks courses to require drafts, revision logs and reflective checkpoints, on the ground that such documentation is less easily outsourced to generative AI alone ². A drafting curve is the nearest thing a Docs panel has to that record. It is also the shape a careful student produces by accident, which is why it works better as a record than as a requirement.

Each shape points to a different next move, and not one of them points to a misconduct referral on its own. After a single arrival, the question is where the text was composed. An unbroken sitting raises the same question with less urgency behind it. With a drafting curve, the next move is a note in the margin, because the record shows engagement rather than authorship ².

That phrasing is the framework's own. Its authors write that "process documentation should be interpreted as evidence of engagement, not proof of independent authorship, since process artifacts are not immune to fabrication" ². The 2026 paper reports no empirical outcomes or validation data, and says so in its abstract ².

Grammarly Authorship records the channel a passage arrived through: typed, pasted from a browser, or pasted from a source it cannot see, such as a private browsing window ⁶. A channel is not an author, and Grammarly itself says the record may not capture every case accurately ⁶. Its categories are typed and pasted ⁶, and nothing on its page says a generated paragraph typed by hand would be recorded as anything other than typed.

Students read the same panel from the other side of the desk, and what a teacher can actually see when text is pasted in is the version of this question they ask. The answers differ by role. The mechanics do not.

It misleads once by accusing a student who drafted somewhere else, and once by clearing a student who retyped generated text. The first failure is the expensive one, because it puts a referral in front of someone who did the work. The second is quiet, and costs only a wrong conclusion held with confidence.

Students draft in a phone's notes app, in Word, or on paper, and text pasted into Docs in one go produces the single-arrival shape. So does writing in a shared family account and copying the paragraphs across. So does a student who lost a document once and now keeps the real draft somewhere with a backup. The panel Google describes ¹ gives no way to separate any of that from a chatbot window opened at 11:40 on a Sunday night.

The 2023 UC Davis case ran the other way. William Quarterman was referred for academic dishonesty after his professor ran his exam through AI detection software, including GPTZero ⁴. He answered with the edit history of the Google document, which gave time-stamped evidence of the paper being built over two or three hours. The university dropped the case after reviewing his evidence, the history among it ^{3 4}.

The reverse needs no skill. A student reading a generated draft on one screen and typing it into Docs on another builds a history that looks like hours of work. What such a record can honestly be offered as, from the student's side, is set out in what a version history proves.

Three product behaviors change what the panel shows. One of them happens without the student doing anything. Browsing earlier versions of a file requires permission to edit that file, so a document shared as view-only carries no visible history ¹. Unnamed versions may occasionally be merged ¹. The owner can also permanently delete every version that was never named, so a wiped history still shows any named versions ¹.

The merging matters more than it sounds. Google's help page describes creating a named version as the way to make sure versions are not merged, which means the unnamed entries a teacher scrolls past may already have been folded together by the product ¹. A thin list can be partly an artifact of that folding. Deletion is the other end of it, and Google marks the step as permanent and impossible to undo ¹.

Word is different. Microsoft states that version history in Microsoft 365 works only for files stored in OneDrive or SharePoint ⁵, so a paper kept on a local drive and emailed as an attachment arrives with no history to open.

Check the sharing setting before reading anything into a short list.

Classroom itself adds nothing, because Google Classroom runs no AI detection of its own.

Ask the student about the paper. Which source was dropped between the second and third versions. Which paragraph was rewritten last, and what the counterargument was before it lost. A writer who built the argument usually has something to say about each; a student who does not is a reason to keep asking, not proof. The panel is at its most useful as the thing that prompts that conversation, and at its least useful as the thing that ends one.

Ten minutes with the student is Quarterman's own suggestion ³. For course design, Harvard College lists oral presentations as a way to evaluate comprehension and let students explain their process, including AI application ⁷, and the Frontiers framework's oral defense runs 8 to 12 minutes per student ². Neither is written as a response to a sparse history; both show that a short spoken check is an accepted form. A conversation at a desk is the first place a sparse history should go, before anything is written down.

A history is a reason to ask, not a reason to file.

The text itself is the second record. Whether a teacher may paste student work into an outside tool is the school's decision. Up to 1,200 words of a paper can be pasted into Human, one paper a day, with no account; it reports how much of that text reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. Human's detector scores the text in memory and de-identifies any personal information for additional training. A longer paper is read only in the part pasted. Below 50 words it declines to score at all. Its own held-out measurement, as of September 2026, in-house rather than independently audited, is 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. That rate was measured on Human's own held-out set of college-level academic writing, and assuming it holds for a particular class's papers is an assumption, not a measurement. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The order does not change once a reading arrives: the history opens the conversation, and the score joins it.