Did My Student Use AI? A Teacher’s Detection Guide

You’re grading a stack of student essays late at night, and one paragraph stops you cold. The vocabulary is suddenly two CEFR levels above what this student produces in class. The transitions are flawless. There are no article errors, no preposition slips, no L1 interference patterns you’ve spent the whole semester correcting. Your gut says ChatGPT. But how do you actually know — and what do you do next?

This guide walks ESL teachers through the practical detection methods that work in real classrooms, the red flags that look damning but aren’t, and the conversations to have with students once you suspect AI use. No detector software required.

Why AI Detection Is Different for ESL Teachers

L1 English teachers detecting AI work usually compare a polished output to a baseline of polished classroom writing. The signal is subtle. ESL teachers have a different advantage and a different problem: our students’ authentic writing has a strong, individual fingerprint — recurring grammar errors, L1 interference patterns, predictable vocabulary ceiling. When that fingerprint disappears in a single assignment, the change is loud.

The problem is the inverse. Strong B2–C1 students often produce writing that mimics AI patterns naturally — clean grammar, hedged conclusions, balanced structure. If you accuse a high-achiever based only on “it sounds like ChatGPT,” you’ll be wrong roughly half the time. Detection has to combine linguistic evidence with what you already know about the student.

Linguistic Red Flags in AI-Generated Student Writing

AI text has stylistic tells. None of these are proof on their own, but two or three appearing together in writing from a student who normally doesn’t produce them should raise your eyebrows.

1. The disappearance of characteristic errors

Your Mandarin L1 student who consistently drops articles suddenly uses “the” and “a” perfectly. Your Japanese L1 student who always confuses “go to home” writes “I went home” without prompting. Your Spanish L1 student stops translating “have” for age. The vanishing of a stable error pattern in a single assignment is one of the strongest signals available to an ESL teacher.

3. Hedge phrases and balanced conclusions

AI loves the safe middle. Look for clusters of phrases like “it is important to note,” “on the other hand,” “in conclusion,” “plays a crucial role,” “a multifaceted issue,” and conclusions that refuse to take a position. A student opinion essay that ends with “both sides have valid points” without picking one is a classic AI fingerprint — most learners write more decisively, even ungrammatically.

3. Vocabulary spikes far above the student’s productive range

An A2 student suddenly using “furthermore,” “nevertheless,” “intricate,” “facilitate,” or “endeavor.” The gap between receptive vocabulary (what they can recognize) and productive vocabulary (what they can use accurately) is usually wide. AI writing collapses that gap unrealistically. If a word would never appear in their spoken English or earlier drafts, ask where it came from.

4. Uniform sentence rhythm

Human writers, especially learners, produce inconsistent sentence lengths — a long run-on followed by a fragment, then a medium clause. AI tends to produce paragraphs where every sentence is between 15 and 25 words, with consistent comma placement. Read the essay aloud. If the rhythm is metronomic, that’s a signal.

5. Em dashes and parallel structure

Em dashes (—) appear in AI output far more than in student writing, where commas, parentheses, or no punctuation at all are more common. Likewise, three-part parallel structures (“not only X but also Y, and ultimately Z”) arrive at a frequency that human learners rarely sustain. One em dash means nothing. Six em dashes in a single 300-word essay from a B1 student means something.

Behavioral Signals Outside the Page

The most reliable detection happens before you ever read the essay. Behavioral signals are harder to fake than text patterns.

• Submission time anomalies: The student who normally submits at 11:55 PM after agonizing for a week submits at 8:00 PM on day one, with no drafts in the version history.
• Topic mismatch: Their essay covers angles you never discussed in class, in a structure that doesn’t match the template you provided.
• Inability to discuss their own work: Ask them to summarize their thesis in spoken English. A student who genuinely wrote the essay can paraphrase their argument. A student who pasted from AI often cannot.
• Citations they can’t explain: If sources appear, ask which one was most useful. AI-generated citations are often fabricated or were never actually read.
• Sudden absence of questions: Students working on real essays usually ask clarifying questions. A perfectly silent student handing in perfect work is unusual.

Why AI Detector Tools Aren’t the Answer

You may be tempted to paste suspect essays into a detector like GPTZero, Turnitin’s AI module, or Originality.ai. Resist the urge to make decisions based on these scores alone. Independent research has consistently shown two failure modes that hit ESL teachers especially hard.

First, false positives on non-native writing. Stanford researchers found that AI detectors flag essays written by non-native English speakers as AI-generated more than half the time, while flagging native-speaker essays correctly the vast majority of the time. The detectors mistake the simpler vocabulary and more predictable syntax of L2 writers for AI output. If you rely on these tools, you will accuse your hardworking A2 students far more often than your cheating C1 students.

Second, false negatives on lightly edited AI. A student who runs ChatGPT output through a paraphraser, or simply changes five words per paragraph, can drop detector confidence below 20%. The tools cannot reliably distinguish heavily edited AI from original work.

Treat detector scores as one weak signal among many — never as evidence on their own.

The Conversation: What to Do When You Suspect AI Use

Once you have a real suspicion built on linguistic and behavioral evidence, do not lead with an accusation. The fastest way to confirm or rule out AI use is a calm conversation in which the student explains their own work. Try this sequence.

1. Ask them to walk you through their writing process. Where did they start? What did they write first? How long did it take? Students who wrote their own work tell a messy, specific story. Students who didn’t tell a smooth, vague one.
2. Pick one strong sentence and ask them to paraphrase it. If they wrote “Globalization has multifaceted implications for cultural identity,” ask them to say it in their own words. A genuine author can. A copy-paste author cannot.
3. Ask them to define three vocabulary words from their essay. Not the easy ones — the unusually advanced ones. Productive vocabulary they actually used should be vocabulary they can define.
4. Offer a redo with a process requirement. Rather than punishing, ask for a handwritten draft, a one-on-one writing session, or a recorded video of them composing the next assignment. This sidesteps the impossible burden of proof while restoring the assessment’s validity.

Designing AI-Resistant Assignments

The strongest detection strategy is prevention. Redesign assignments so that AI is either useless or has to be used transparently.

In-class drafting with handwritten first drafts

If the first draft of every essay is produced in class, on paper, in your presence, you have a baseline of authentic voice for every student. Subsequent typed revisions can then be checked against that draft. A student whose final version uses vocabulary nowhere in their handwritten draft has questions to answer.

Personalized prompts tied to class discussion

Instead of “Write about environmental protection,” try “Write about the field trip we took to the Daan recycling center on Tuesday, including the comment Mr. Chen made about plastic bottles.” AI cannot generate references to specific shared experiences. The more your prompt depends on classroom-specific context, the less useful AI becomes.

Process-based assessment

Grade the outline, the rough draft, the peer review notes, the revision plan, and the final draft separately. The final essay is worth 30% of the assignment grade, not 100%. A student who can produce a perfect final essay but no outline or draft has just failed the assignment by structure, regardless of AI use.

Oral defense of written work

For higher-stakes assessments, require a five-minute oral discussion of the essay after submission. The student summarizes their thesis, defends a counterargument, and answers two questions about their sources. This makes AI use functionally impossible — a student who didn’t engage with the material cannot fake the conversation.

Allow AI, with required transparency

In many contexts, the most realistic policy is supervised use rather than prohibition. Allow students to use AI for brainstorming, grammar correction, or vocabulary suggestions — but require an appendix documenting which prompts they used and what they changed. This transforms AI from cheating into a skill, and detection from a forensic exercise into a transparency check.

A Calibration Habit That Saves You Hours

Before the term begins, paste your essay prompts into ChatGPT yourself and read the output carefully. You’ll see exactly what the AI tends to produce for that question — the structure it defaults to, the phrases it overuses, the conclusions it lands on. When student work comes in that matches the AI output you generated, your detection becomes pattern-specific rather than vague. This 15-minute exercise per prompt is more useful than any detector tool you’ll ever buy.

A Word on Compassion

ESL students using AI are often not trying to cheat. They’re drowning. Time pressure, exam anxiety, fear of judgment, the gap between their ideas and their ability to express them — these push capable students toward shortcuts they wouldn’t otherwise take. Detection is necessary, but the conversation that follows is more important. Many students who used AI once will not use it again if they understand that the goal of writing class is not the essay but the growth.

The teacher who only punishes loses the student. The teacher who only forgives loses the assessment’s meaning. The teacher who detects calmly, talks honestly, and redesigns assignments thoughtfully keeps both.

Zdroje

• Large language model — Wikipedia
• British Council — English teaching resources
• Mezinárodní asociace TESOL
• Stanford Institute for Human-Centered Artificial Intelligence (HAI)
• Cambridge — language assessment and research

More Worksheets

Postavy X-Menů: Síly a slabiny - lekce angličtiny jako druhého jazykaPrevence syndromu vyhoření učitelů | 8 strategií pro učitele angličtiny jako druhého jazyka založených na důkazech

The Amazon Rainforest | Elementary ESL Reading Worksheet PDF