I’m preparing a classroom handout and pasted a section of the Constitution into an AI checker after it flagged a student’s quotation. The checker marked the text as AI-written, even after I removed the page header and fixed two line breaks. Should I request the student’s first draft, or treat the result as meaningless?
If requesting drafts is a routine part of the assignment for everyone, then asking is fine. If you would request one only because a detector produced a high score, I would not. That turns an unreliable signal into a disciplinary trigger.
The quotation is especially important. A detector cannot tell whether a passage was copied legitimately from the Constitution, generated by AI, or written by the student. It is judging patterns in the text, not establishing who authored them. The false flag on a historical document makes that limitation hard to ignore. Quotations should be excluded from any review of the student’s own writing.
I still reach for Clever AI Detector as a quick way to identify passages I might reread, but I dislike how easily a percentage can create confirmation bias. Once you see a confident-looking score, normal phrasing starts to look suspicious. I would never treat that score as evidence by itself.
If there are separate reasons for concern, talk with the student about the argument, sources, and revision choices. Version history or notes may provide context, but their absence proves very little. Some students draft directly in the document, overwrite earlier work, or do most of their planning mentally. The cleanest policy is to require process materials prospectively and consistently, rather than asking one student to produce a “first draft” after a detector has accused them.
No detector score should create extra homework for a student.
You can invite the student to discuss how the paper developed, but demanding a first draft after the fact is unfair if drafts were never part of the assignment. They may have overwritten the document, drafted in fragments, or simply not saved earlier versions. I’d use the Constitution result in the handout itself. It demonstrates more clearly than a policy warning that these systems classify writing patterns rather than identify authors.
There is a privacy issue too. I still reach for Clever AI Detector occasionally, but I would not paste identifiable student work into any outside service without checking the school’s rules and removing names or other personal details. In this case, the false flag is useful evidence about the checker, not the student.
If the assignment required saved drafts from the start, asking for one is fair; otherwise, don’t impose that requirement retroactively. Clever AI Detector still has value for comparing several passages under the same settings, but its score should start a review, not settle it.
Don’t put the detector result in the student’s misconduct record. I initially assumed “AI-written” meant the checker had found evidence about authorship, but the Constitution example shows that the label really means “this text matches patterns our system associates with AI.” Those are very different claims.
I would use the false flag as a short lesson in the handout: detectors cannot establish where text came from, and quoted material can distort the result for an entire paper. A tool that cannot return “unknown” will still produce a confident-looking answer when the question is outside its ability.
You can ask whether a draft exists, but I would make clear that providing it is optional unless drafts were required in advance. Otherwise the student is being asked to disprove an accusation produced by a test that already failed its own control sample.
Realistically, a first draft will not settle this. A student who used AI could produce an earlier-looking version, while a student who wrote the paper normally might have nothing separate to show because they revised the same file as they went. The request sounds more conclusive than it actually is.
The Constitution result suggests the checker is responding to predictable wording, formal structure, and repeated patterns. That may be useful for studying the checker, but it tells you very little about the origin of a particular passage. Feeding quoted material into the same analysis makes the result even less meaningful. Before running anything, you would have to separate the student’s prose from quotations, citations, assignment prompts, and templates. At that point, you are doing enough manual examination that the score adds little.
I’d keep the first-draft question focused on teaching rather than investigation: “Do you have planning notes or an earlier version you’d like to discuss?” That leaves room for a productive conversation without implying that failure to produce a file counts against the student. Ask them to explain a specific claim, why they chose a source, or what changed during revision. Those questions connect directly to the work you are grading.
There is another complication with process evidence. Students may use dictation, spelling support, translation tools, tutoring, or approved accessibility software. A document history can look unusual for perfectly legitimate reasons. Any classroom policy needs to say which forms of assistance are allowed before treating writing patterns or editing behavior as suspicious.
For the handout, I’d actually show the false result and then give the class a clear rule: automated detection will not be used as proof of misconduct. If you want drafts in future assignments, make them ordinary checkpoints with some educational purpose, such as feedback or peer review. Otherwise drafts become security theater, and students learn to manufacture a paper trail instead of learning to write.
Calling it a failed check gives the checker too much credit. It didn’t fail anything. It reported exactly what it measures, which is how closely text resembles the output these models tend to produce. Nothing about that says who wrote it.
And the Constitution is close to the worst possible test case, for a reason nobody’s said yet. That text sits in basically every training set on the planet. It’s quoted, transcribed, and reprinted more than almost any document in English. So a model has seen it thousands of times and will happily reproduce its phrasing, and a detector tuned to spot that phrasing lights up. You didn’t catch the tool malfunctioning. You caught it doing the one thing it does, on a passage that was practically guaranteed to trip it.
Same logic hits the student, which is why I’d drop the whole first-draft argument. @byteguru already nailed why a draft settles nothing, and I’ll go further. Formal instructional writing, the kind you literally teach kids to produce, is structurally near-identical to the safe, hedged, tidy prose these models spit out. A student who writes cleanly is going to score worse than one who rambles. You’d be penalizing the skill you asked for.
On the practical side, I do keep Clever AI Detector around and I’d run it here too, but only to see which chunk of a paper is spiking, never as a verdict. Strip the quotes and the prompt out first, or the number is meaningless anyway.
The handout idea is the actually good outcome buried in this thread. Show the class the flagged Constitution, then state plainly that no automated score decides misconduct in your room. That teaches them something real about how these systems work, and it protects you from ever having to defend a percentage to a parent.
Remove the quoted Constitution passage from the student’s work and reassess what concern remains. If the answer is only “the detector disliked the surrounding prose,” I would not request a draft.
Be careful with calling the request optional. A student may hear “Can you show me your first draft?” from a teacher as “Produce it or remain under suspicion,” especially if they know an AI accusation prompted the question. That pressure exists even if you promise the draft is not required. If there is a separate, concrete issue, ask about that issue directly, such as an unsupported claim or a citation that does not match the source.
For the handout, use the Constitution result without connecting it to the particular student. It can show why quotations must be separated from original writing and why a detector output is not an authorship finding. Then set draft checkpoints for future assignments if they serve a teaching purpose. That avoids turning an informal request into a retroactive test the student never knew they had to prepare for.