Discover if teachers can tell if you used AI to write an essay and why relying on it can be risky for your transcript.
The software isn’t catching you. Your teacher is.
I’ll plant my flag early: yes, they can tell. Not every time. Often enough that betting your transcript on it is bad math.
And it’s almost never the detector. It’s the human who read your two discussion posts in week three, watched you fumble a definition out loud on a Tuesday, then opened 1,400 words of frictionless prose about Foucault.
Teachers grade in batches. Forty essays on a Sunday, coffee going cold, dog barking. By essay nine your voice is a fingerprint. You overuse “basically.” You love a rhetorical question. Your paragraphs sprawl and then apologize for sprawling.
Then one week the sprawl vanishes. That gap is the tell.
The sharper ones also run a cheap trick: a handwritten in-class paragraph in week two, collected and filed. Ten minutes of class time, a baseline for the whole semester.
The stuff that actually blows your cover.
Citations. Always citations. A model will hand you “Halberstam, 2017, p. 43” with total composure, and the article will not exist. Ask Steven Schwartz. He filed a brief in Mata v. Avianca citing six invented cases, got sanctioned $5,000 in June 2023, and became the cautionary tale every professor now tells in week one.
Your grader doesn’t need a tool for that. They need thirty seconds and a library search box.
Then there’s the class-shaped hole. You wrote a competent essay on Beloved that never once touches the argument your professor spent forty minutes on, the one about Sethe’s scar as text. Every student in the room used it. You didn’t. Why not?
Other giveaways worth knowing, because they’re fixable in your own writing too:
- Three-item lists everywhere. Always three. Never two, never five.
- Paragraphs of near-identical length, like bricks.
- Zero typos and zero specifics in the same document.
- Words nobody in your seminar says out loud. “Delve” got so common in 2024 that researchers tracked its spike across PubMed abstracts.
- A conclusion that restates the intro with the furniture rearranged.
The pattern underneath all of it: confident, tidy, and weirdly uninterested in the particular text you were assigned.
The detectors themselves are a mess.
Here’s where the AI hype vs reality gap gets ugly, and where I think a lot of teachers have been sold something.
Turnitin shipped AI detection in April 2023 and ran it across tens of millions of papers. Vanderbilt turned it off that August, publicly, citing false positives and no transparency on how the score is produced. They weren’t alone.
A Stanford group tested seven detectors on TOEFL essays written by real humans. Over half got flagged as AI. Non-native English writers use simpler vocabulary, and simpler vocabulary reads as machine-generated to these tools. That’s not detection. That’s a bias machine with a percentage attached.
OpenAI quietly killed its own classifier in July 2023. Its own numbers: it caught about a quarter of AI text and falsely accused human writers roughly nine percent of the time.
So if you’re accused, the score alone is weak ground, and you should say so calmly. Ask what else supports it. Most honest instructors will admit the number is a flag, not proof.
But don’t get comfortable. The weakness of the software is exactly why teachers moved to oral defenses, in-class drafts, and questions like “walk me through your third paragraph.” Those you cannot bluff.
The receipts you forgot you were generating.
This is the part nobody warns students about, and it’s the most useful thing in this article.
Google Docs keeps version history. File, Version history, See version history. It stores every save, every big paste, timestamped. A 900-word block appearing at 11:47pm in a single event looks like exactly what it is.
There’s a Chrome extension called Draftback that replays a Doc being written, keystroke by keystroke, like a film. Teachers use it. Students also use it — to prove they wrote the thing.
So flip it. If you actually write your essays, your draft history is an alibi, and it’s free. Work in one Doc from the first bad outline onward. Don’t compose in Notes and paste the finished block in.
Word tracks similar metadata, and Grammarly now stamps authorship sessions too. Canvas logs your submission time and every file you swapped. The trail exists whether or not anyone looks at it.
If you get accused and you didn’t do it.
This happens, and it happens most to students who write plainly. Panic is the wrong move. So is a long emotional email at 2am.
Do this instead. Ask, in writing, what the accusation rests on. If the answer is a detector percentage, ask which tool and what its documented false positive rate is.
Then hand over the boring evidence. Version history link. Your notes photo. The library checkout. The three browser tabs you had open, still in your history from that Thursday.
Offer a conversation about the essay. Fifteen minutes, no notes. Anyone who wrote the thing can talk about it for fifteen minutes, and graders know that better than any score does.
One more thing people skip: check your school’s policy page for who handles appeals. Usually it’s a named academic integrity officer, not your instructor, and they’ve seen the false positive problem up close.
Where AI is fine, and where I’d stop.
I’m not anti-AI. I use these things daily. But the question “is AI worth it” has a different answer for every task, and pretending otherwise is how people get caught.
Everyday AI tools earn their keep on the boring edges of writing. Turning your messy bullet list into a working outline. Catching comma splices at 1am. Reformatting fourteen citations into Chicago style. Asking it to attack your thesis so you can patch the weak spot before your grader finds it. That last one has saved my arguments more than once.
AI for small tasks is genuinely good, and you can do nearly all of it on free tiers. Using AI without paying covers most student work — the paid version mainly buys speed and longer context, not better judgment about your Hegel paper.
When not to use AI is simpler than the internet makes it sound. Not for the argument, not for the citations, not for the reading, because a summary of a chapter you never opened collapses the second someone asks a follow-up question.
There’s also the part that’s easy to say and annoying to hear: the struggle of the first bad paragraph is the thing school is actually selling you. Outsource that and you’re paying tuition for a receipt.
What I’d do if I were you.
Write the ugly draft yourself. Keep it in one document. Let AI clean the edges, not build the house.
And if you’re going to cheat anyway — at least check whether the sources exist. Steven Schwartz didn’t.
