How to Collect an In-Class Writing Sample and Read an Essay Against It

A few hundred words, written in class early in the term with nothing open and nobody helping. This article's working rule is 300 to 500 words in the first two weeks; it is a rule of thumb, not a measured threshold. The prompt has to produce the same kind of writing the course will grade later. GPTZero's teacher guide puts a short in-class piece first, ahead of any scan, and calls it a snapshot of a student's natural writing voice ¹.

Length matters more than it looks. A very short sample leaves too little of a student on the page for a comparison to mean much, and the working rule assumes a class period does not yield much more than 500 words by hand.

Two samples beat one. The reason is the objection that arrives in every misconduct meeting: students get better. A single September page makes ordinary growth look like substitution. A September page and a late-October page show the slope, which is what a December essay has to be read against.

Say out loud what the sample is for on the day it is collected: so that nobody is accused on the strength of a score alone. A class that learns of it first in a disciplinary meeting has reason to read it as surveillance.

The prompt has to pull the same muscles as the assignments that follow, and a personal narrative pulls none of the ones a source-based argument needs. Give the class a short passage from the course reading and ask for a claim about it with two reasons, in whatever form comes naturally, then keep that prompt reusable so the second sample can be read against the first.

The shape that travels: a passage everyone has already read, a question with more than one defensible answer, and an instruction to take a position. No research, no citation format, no choice of topic. Choice of topic is one of the quickest ways to make two samples incomparable.

The sample is not a graded assignment. Marking it for correctness turns it into a performance, and a performance is the wrong baseline. Collect it, read it once, then file it. That one reading is the part that pays off in December.

GPTZero's guide tells teachers to pay attention to patterns ¹. Quality is the wrong thing to read a sample for; habits are what survive into a later essay. A word the student reliably misspells. British and American spellings used interchangeably. Long unbroken sentences, or short flat ones. Vocabulary that is either narrow or unusually adventurous ¹. Those are what a comparison can see.

A course where the writing happens in visible stages collects the same evidence without a separate exercise. The assignments that survive AI are the ones where a proposal, an annotated source list and a rough draft all exist before the final file does.

In the course space the school already runs, under each student's name. Not a personal drive, not an email thread, not a phone camera roll. Under FERPA (34 CFR 99.10, current as of 2025), which covers US schools that receive federal education funds, a student who is 18 or in college, or a parent of a younger student, can ask to inspect education records, and the school must comply within a reasonable period and no later than 45 days ⁴.

A sample filed there is likely to count as an education record; the school's records office decides.

Retention is the easier half of the question. Keep a sample through the grade appeal window and any open integrity case, then delete it or hand it to the school's records process, whichever the school's retention schedule requires. Retention is the school's rule, not the instructor's, and files about a named student that no longer serve a purpose are worth clearing out where that rule allows.

Three places a sample should not end up:

• A personal laptop or a private cloud account that leaves with the instructor at the end of term.
• An email thread, which scatters copies the school cannot gather again on request.
• A third-party detection tool, uploaded as a reference document. Handing a student's unpublished writing to an outside vendor to store is a different decision from running one check, and it belongs to the school rather than to one instructor.

Handwritten pages photograph fine. Name each image by student identifier and date. One image per page, sitting in the same course space the essays land in, is enough to make the comparison reproducible six weeks later.

Read for the habits that persist. Set the two texts side by side and look at four things: recurring spelling and punctuation habits, how long the sentences run and how much that varies, the way sources get introduced, and the words the student reaches for under time pressure. Quality is expected to rise over a term. Habits mostly don't.

Four comparisons that carry information:

• Habitual error. A student who has misused the semicolon since September, then suddenly never does, in a paper whose argument is no sharper than before.
• Sentence rhythm. Supervised, timed writing often runs uneven. A paper whose sentences all land within a few words of each other may have been through something: a grammar tool, a tutor, a model. Or that is how the student writes when there is time. Either way it's one question to ask, not a conclusion.
• Source handling. How a student introduces a quotation is a habit the course teaches. A new convention appearing correctly on the first attempt is worth one question.
• Register. Vocabulary that climbs two levels while the argument stays where it was.

None of the four settles who wrote the essay. What a comparison does is narrow a vague suspicion into something specific enough to ask about, and the next move is a request for the drafting history rather than a scan. The version history of the document shows whether a paper was built or whether it arrived.

Then ask the student. Two questions about one paragraph go further than a percentage: why that source, and what got cut. That conversation is close to the oral defense that Harvard College's academic integrity resources list as a way to evaluate comprehension and let students explain their process, including how AI was used ⁶.

The whole sequence, with software last, is the five checks in order.

Because style moves for reasons that have nothing to do with a chatbot. A writing center appointment. A grammar checker. A term of reading, or a topic the student already knew well. An essay drafted over two weeks instead of 40 minutes. When the Stanford team had ChatGPT simplify the word choices in US eighth-grade essays, the average misclassification rate across the seven detectors went from 5.19 percent to 56.65 percent ².

Detection software makes the same case for keeping the comparison first. Bloomberg Businessweek ran two detectors over 500 college application essays submitted to Texas A&M University in the summer of 2022. That was before ChatGPT was released, which all but guarantees they were written by people ³. Between 1 and 2 percent came back flagged on that one set of essays, some with near total certainty ³.

The bias has a seam, and it is the one a baseline is meant to protect. In the same Stanford study, having ChatGPT enrich the vocabulary of the human TOEFL essays dropped the average false positive rate from 61.22 percent to 11.77 percent ². The flag rate moved with word choice. A class with many non-native English writers has a particular reason to keep the comparison first, since the flag rate in that study was highest on their essays ², which is why detectors and non-native English writers is a policy question before it is a classroom one.

A baseline does not settle a case either; in the one case this article cites, the student's evidence was the drafting history ⁵. The referral at UC Davis began with a professor running an exam through a detector. William Quarterman, the senior who wrote that exam, answered the accusation with the edit history of the document he wrote it in, which showed the paper being built over two or three hours, and was eventually cleared ⁵. The sample is what tells an instructor whether asking for that history is worth the afternoon.

Software goes last, and a teacher who reaches for it after the side-by-side reading has a specific question for it to answer. Human fits that slot: paste up to 1,200 words of the suspect essay at human.olive.is (one paper a day, no account) and it reports how much of the text reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. Its own in-house measurement as of September 2026, on its own held-out set, with no outside audit, 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%. Human is tuned for college-level academic writing; how that rate carries to another class's essays 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. Human's detector scores the text in memory and de-identifies any personal information for additional training. Read beside the baseline, the reading is one more difference to ask the student about, and the drafting history is still the next thing to ask for.