Recognizing AI-Generated Student Work

A Practical Guide for Teachers

A TeachSpark resource

---

Before You Start: What This Guide Is (and Isn't)

There's no reliable way to prove a piece of writing was generated by AI — detection tools like Turnitin's AI checker or GPTZero have real false-positive problems, and a wrongly-accused student is a worse outcome than a missed one. This guide won't give you a verdict. It'll give you the same thing an experienced teacher's gut already does after years of grading: a feel for what doesn't sound like your student.

Use these as prompts for a conversation, not a courtroom. The goal is catching disengagement early and redesigning the assignment — not building a case.

---

The Tells: What to Look For

1. Too smooth, too fast Real student writing — especially from students below advanced levels — has friction: a clunky transition, a sentence that runs a little long, a word repeated because they didn't think of a better one. AI writing is often suspiciously even in quality from the first sentence to the last, with no rough patches.

2. No personality, no specificity Genuine student work usually has some trace of this student, this class: an inside reference to something you discussed Tuesday, an opinion phrased the way they talk, even a complaint about the assignment itself. AI-generated answers tend to be generically "correct" — technically responsive to the prompt but disconnected from anything specific to your class, your discussion, or your local context.

3. Vocabulary or sentence complexity above their normal level If a student who typically writes short, simple sentences suddenly turns in something with varied sentence structure and words like "however" and "furthermore" used correctly and often — that jump is worth noticing. This is the single most useful signal if you have anything to compare against (see "Building a Baseline" below).

4. Structural uniformity AI-generated essays often follow a very predictable shape: an intro that restates the prompt, three body paragraphs of roughly equal length, a conclusion that restates the intro. Real student writing is messier — paragraphs of wildly different lengths, ideas that don't wrap up neatly.

5. Confident but shallow AI-generated answers are frequently correct-sounding without being specific. Ask a follow-up question ("what did you mean when you said X?") and a student who didn't write it often can't explain their own reasoning, or explains it in language that doesn't match what's on the page.

6. No errors at all This sounds counterintuitive, but zero spelling errors, zero grammar slips, and perfectly placed commas from a student who doesn't usually write that cleanly is itself a signal. Real first-draft student writing almost always has something imperfect.

7. Finished far faster than the work should take TeachSpark already tracks how long a student spends on an activity. A response that comes back much faster than the activity should reasonably take is worth a second look — though a quick, confident student isn't automatically a red flag either. Use it alongside the other signals here, not on its own.

---

Building a Baseline (Your Best Tool)

The single most reliable thing you can do costs nothing and takes five minutes: get a low-stakes, in-class, handwritten or supervised writing sample from every student in the first week or two of the year. A journal entry, an "introduce yourself" paragraph, anything.

This gives you something to compare later work against — not to run through software, just to glance at side-by-side. A sudden jump in vocabulary, sentence complexity, or tone from that specific student's baseline is far more meaningful than a generic "does this sound like AI" judgment, because you're comparing the student to themselves, not to an abstract pattern.

---

Designing Assignments That Are Harder to Outsource

This matters more than detection. If AI can't easily answer the question well, you spend a lot less energy playing detective.

• Reference something only your class would know. A question tied to a specific class discussion, a shared reading, or a local example is much harder for a generic AI tool to answer convincingly than an open-ended prompt any search engine could handle.
• Ask for reasoning, not just answers. A short "explain your thinking" or "what was the hardest part of this for you" addition, even two sentences, raises the bar significantly — and low-effort reflections are often a giveaway on their own.
• Build in checkpoints for longer work. An outline, then a draft, then a final version creates a paper trail of the student's own thinking process, and makes last-minute AI-generation much harder to pull off convincingly across every stage.
• Favor in-class, supervised work for high-stakes assessment. This isn't about distrust — it's about knowing where your assessment weight actually needs certainty versus where lower-stakes practice work is fine either way.

---

If You Suspect Something: How to Approach the Conversation

• Lead with curiosity, not accusation. "Can you walk me through your thinking here?" opens a conversation. "Did you use AI for this?" closes one.
• Ask about a specific detail, not the whole piece. "What made you choose this example?" is much harder to fake convincingly than a general defense.
• Treat it as a teaching moment first. Many students don't see using AI to generate an answer as different from looking something up — they may not fully understand why it matters for their own learning until you explain it directly.
• Know your school's policy before the conversation, so you're not improvising consequences on the spot.

---

A Quick-Reference Checklist

Use this as a mental checklist when something feels off about a submission:

• Does the tone match how this student normally writes or talks?
• Is there anything specific to our class, our discussion, or a local reference — or could this answer have been given by any student, anywhere?
• Is the vocabulary/sentence complexity consistent with their earlier work this year?
• Are there any small imperfections, or is it suspiciously clean?
• Did they finish much faster than the activity should reasonably take?
• If I asked a specific follow-up question about their own answer, could they explain it?

None of these alone is proof of anything. Together, they're usually enough to know whether a conversation is worth having.

---

This guide is included with your TeachSpark subscription. We built TeachSpark around the idea that teachers should be the ones using AI — not students — and we know that trust runs both directions. If you have a suggestion for this guide based on something you've seen in your own classroom, we'd love to hear it.