HomeGuides › How to Spot AI in School Essays

How to spot AI in school essays — patterns teachers see daily.

Written for K-12 and college teachers grading at volume. By the time a detection tool finishes loading, an experienced teacher has already noticed the suspiciously even five-paragraph shape, the generic thesis, the polite-assistant tone that no fifteen-year-old uses in class. This guide is the classroom side of the workflow. Five tells you see daily, a five-step read that takes under three minutes per essay, an ESL calibration note that protects your non-native English students from the false-positive trap, and the FERPA-aware way to bring TextSight into the loop when the call is genuinely ambiguous. The score opens a conversation; it does not close one.

Check an essay free See the five tells
5 classroom tells ESL-aware calibration FERPA-conscious workflow
Why this matters

The teacher already knows the student — no detector can do that.

Detectors score a draft. Teachers know the student behind it. The strongest classroom signal in 2026 is the gap between a familiar voice and the page in front of you, and that comparison happens in your head before any tool opens.

The voice you already know is the baseline

Class participation gives you sentence rhythm and active vocabulary. The last in-class write-up gives you a controlled register sample from the same week. Office-hours questions and informal chat show how the student reaches for words when nobody is grading them. By the third week of any term, you carry a richer model of how a particular student writes than any classifier ever will. The classroom tells below all point back to that same baseline comparison.

Conversation-starter, not verdict

A flagged essay is the start of a conversation, not the end of one. The honest job of the read is to surface drafts worth a closer talk with the student, not to declare guilt. Teachers who present a verdict on the strength of one read tend to lose authority when the student pushes back. Teachers who present the pattern and ask the student to walk through their drafting process get more honest answers and keep the relationship intact.

Manual spotting is fast above 250 words

The honest ceiling on classroom reading sits around 75 to 85 percent on essays above 250 words. That is enough to sort a class set into clear-human, clear-AI, and a smaller ambiguous middle. The ambiguous middle is where TextSight earns its keep, with sentence-level highlights you can respectfully share in a one-on-one. Below 250 words, lean harder on the in-class writing sample than on either eye or tool.

Five tells

The five essay patterns teachers see daily.

Any one tell can show up in genuine student work. The signal is when three or more cluster in the same draft. Read each tell as a yellow flag, not a verdict, and weight the opening-phrase and closing-paragraph tells most heavily because they bookend the read.

1. "In today's society" and the polite-assistant opener

The single most reliable school-essay tell. "In today's society," "In today's world," "Throughout history," and "In our modern era" land in the first sentence of roughly half of ChatGPT-drafted school essays. The opener carries the polite-assistant register that every major model defaults to: measured, balanced, gently agreeable. Real students in 2026 rarely open this way. A first sentence that sounds like a TED talk in the voice of a teacher is the loudest tell on the page.

2. The uniform five-paragraph structure

Intro, three body paragraphs, conclusion. Equal lengths to within twenty words. Neat transitions between each. Real student drafts at school level wobble: an oversized intro, a stub of a body paragraph, an abrupt close, a missing transition. AI hits the template on the first attempt with surgical precision. Five paragraphs that all sit within twenty words of each other in length is structural evidence that nobody drafting under deadline produces naturally.

3. The generic thesis that fits any prompt

A thesis that hedges every claim and applies to almost any topic in the unit. Sounds thoughtful, commits to nothing specific, and rarely names a stake the student would actually defend in a discussion. Variants include "While there are many perspectives on this issue, it is important to consider both the positive and negative aspects" and "This essay will examine the various ways in which the author explores the complexity of the human condition." Real student theses are riskier, sometimes wrong, often interesting.

4. Transition-phrase clusters between paragraphs

"Furthermore," "Moreover," "Additionally," "In addition," "Consequently," "Subsequently," "Nevertheless." A real student uses one or two transition phrases across a 600-word essay. AI uses one at the start of almost every body paragraph and often a second mid-paragraph as well. Three or more formal transitions clustering across the same draft, especially in a student whose class voice is plain, is one of the cleanest single signals available without a tool.

5. The polite-assistant closing paragraph

"In conclusion," "To conclude," "In summary," "Overall," followed by a sentence that widens the scope to society, the world, or the human condition. Polished, vague, applicable to every essay on every topic. "Ultimately, this reminds us of the enduring importance of empathy in our world today." Older students were taught to avoid the "In conclusion" signposting years ago. AI brings it back in every model family. A polite-assistant close paired with the opener in tell 1 is the strongest two-tell pairing.

Plans & pricing

Detector and AI rewriter on every tier.

Free covers a few essays per session, enough for a small class set. Teachers grading weekly run on Starter or Pro. Departments running batch review at scale move to Business for the REST API and team seats. Full details on the pricing page.

Free
$0/forever

 

A few essays per session, no card.
  • 10,000 character daily detector quota
  • Sentence-level highlights
  • Authenticity Score on every scan
  • ESL-aware calibration by default
Start free
Starter
$7.49/month

Billed $89.88/year — Save $30

For teachers grading a small class.
  • 20,000 AI rewriter words/mo
  • Chrome extension
  • File upload for essay PDFs
  • Email support
Get Starter
Business
$29.99/month

Billed $359.88/year — Save $120

For departments and writing centres.
  • 150,000 AI rewriter words/mo
  • REST API access
  • 5 team seats
  • Webhook integrations
Get Business

Yearly billing saves 25%. View full pricing →

5-step workflow

The five-step classroom read — under three minutes per essay.

Three eye steps that use the voice you already know, one tool scan on the ambiguous middle, one respectful conversation. The whole sequence fits inside the time you would have spent on a careful first read of the draft anyway.

Step 1 — read the first paragraph for tone

Read the opening paragraph in isolation, before the rest of the essay has a chance to set context. Listen for the register. Does it sound like this student in class, or like the polite-assistant voice every chatbot defaults to? A sudden jump from the student's everyday classroom voice to a measured, balanced essay register is the loudest single signal, and it lives in the first 80 words. The opener carries roughly forty percent of the classroom-read weight on its own.

Step 2 — check the thesis for specificity

Find the thesis sentence and ask whether it commits to anything a particular student would defend. A thesis that hedges every claim and applies to almost any prompt in the unit is a strong yellow flag. A thesis that names a specific stake, a particular scene, a precise claim about cause is the inverse signal. Real student theses are riskier, sometimes wrong, often interesting. Specificity is the strongest counter-signal to AI drafting.

Step 3 — look for uniform paragraph rhythm

Scan the body paragraphs as shapes on the page rather than reading every word. AI lands the five-paragraph template with surgical precision: three body paragraphs within twenty words of each other in length, mirrored topic sentences, evenly weighted transitions. Real student drafts wobble. The rhythm is rarely this clean on a first attempt. Two paragraphs that begin with the same transition phrase or echo the thesis verbatim is another red flag inside this step.

Step 4 — scan with TextSight when the read is ambiguous

For drafts where two or three tells fire but the call is not obvious, paste the essay into app.textsight.ai for sentence-level highlights and an Authenticity Score. The scan takes under a minute. The free plan covers 10,000 characters per day, which is several school essays per session. Pro at $19.99 monthly or $14.99 on yearly billing removes the daily cap for teachers grading weekly. The detector is calibrated to flag ESL writing at roughly 40 percent lower false-positive rates than open-source baselines, which matters in any classroom with non-native English students.

Step 5 — discuss with the student

Treat the read as a conversation-starter rather than a verdict. Ask the student to walk you through their draft paragraph by paragraph. A genuine author can explain why they chose a particular thesis, where they found a quote, what they would change in a second pass. A student who outsourced the draft tends to stumble at the source of the citations and the reasoning behind the topic sentences. Bring the pattern and the highlights, not the percentage. The goal is better writing from the student, not a clean catch for the teacher.

The false-positive that matters most

Non-native English students and the 40 percent calibration gap.

For any classroom with non-native English students, which is almost every classroom in 2026, the ESL caveat is the single most important calibration on this page. Get this wrong and the read produces unjust outcomes regardless of how clean the tells look.

What the research says about ESL false positives

Multiple peer-reviewed studies published since 2023 have shown that off-the-shelf AI detectors flag English-as-a-second-language writing as AI-written at roughly three to five times the rate of native-English writing on the same task. The reason is structural rather than accidental. Learned-second-language English uses more uniform sentence shapes, a narrower active vocabulary, and a more formal register, all of which overlap with the statistical signature classifiers were trained to recognise. The detector is not failing; it is correctly measuring something that happens to mean a different thing for ESL students than for native ones.

TextSight runs about 40 percent lower false-positive rates on ESL prose

TextSight trains on diverse English varieties rather than only US classroom prose, which narrows the structural overlap by roughly 40 percent against open-source baselines. The practical effect is a lower false-positive rate on ESL essays, not a zero false-positive rate. No detector eliminates the overlap; the best ones narrow it. For non-native English students, weight the tells cautiously and lean on the in-class writing sample from the same week as the controlled comparison, since that sample carries the same register constraints the take-home essay does.

What to do operationally in the classroom

Build the calibration into the workflow rather than into the score. For known ESL students, require at least three of the five tells to fire plus a clustered sentence-level highlight pattern in TextSight before treating the read as a high-confidence call. For any high-stakes step, including a failing grade or an integrity referral, never act on the score alone. Lead with the in-class writing sample, the per-sentence evidence, and an honest conversation about drafting process. The student deserves the calibration; your judgement gets stronger because of it.

FERPA-aware workflow

Paste the prose, not the student.

US public schools and many private institutions operate under FERPA, which restricts how third-party tools touch identifiable student records. The TextSight workflow is built so the scan never needs identifying metadata, and the conversation happens face to face.

What the scan does and does not need

TextSight does not require a name, ID number, roster metadata, or any school-system identifier to score a draft. Strip headers and footers, paste the prose only, and the detector returns the same Authenticity Score and sentence-level highlights it would for any anonymised passage. The scan is the smallest unit of evidence; the conversation with the student is where context lives. Reviewing the data-processing agreement before adopting any detector district-wide is the right step for schools subject to FERPA, and the privacy page is the source of truth on what TextSight stores.

Document the pattern, not the percentage

For any case that may travel toward a formal integrity referral, document the tell pattern, the per-sentence highlights, the in-class writing sample you compared against, and the conversation with the student. A single global percentage in a student record is brittle evidence that loses on appeal. A pattern of tells plus a controlled comparison plus a documented conversation is the defensible record. Teachers who lead with the pattern hold up far better than teachers who lead with the score.

Tell students the policy up front

Stating the AI-use policy on day one changes drafting behaviour more than any single catch. Pair the policy with a writing process that includes in-class drafting, so the comparison sample exists when you need it. Transparency turns the scan into a teaching tool rather than a hidden trap, and it gives the student the chance to ask honest questions about acceptable use before the first essay is due. The strongest classrooms in 2026 are the ones where the AI conversation is open, not punitive.

FAQ

Spotting AI in school essays frequently asked.

What is the single biggest tell that a school essay was written by AI?
The opening phrase. "In today's society" and "In today's world" show up in the first sentence of roughly half of ChatGPT-drafted school essays. Paired with a perfectly even five-paragraph shape and an "In conclusion" close, the verdict is settled. Real students rarely open this way in 2026.
Can I accuse a student based on these tells alone?
No. Treat the classroom tells as triage, not verdict. They surface essays worth a closer conversation. Any formal academic-integrity step needs more than a teacher's read: a calibrated detector report, an in-class writing sample for comparison, and an oral walk-through of the draft with the student. The goal is a conversation, not a gotcha.
How do I handle ESL students fairly?
With more care, not less. Learned-second-language English shares structural features with AI prose: uniform sentence shapes, narrower active vocabulary, more formal register. Off-the-shelf detectors flag ESL students at three to five times the native-English rate on the same content. TextSight runs roughly 40 percent lower false-positive rates on ESL writing than open-source baselines, but no detector eliminates the overlap. For non-native English students, weight the tells cautiously and lean on the in-class writing sample as the comparison point.
What about FERPA and student privacy when I paste an essay into a scanner?
Paste the prose only, not student identifiers. TextSight does not require a name, ID number, or any roster metadata to score a draft. Strip headers and footers before pasting. The scan is the smallest unit of evidence; the conversation with the student is where context lives. Schools subject to FERPA should treat detector use the same way they treat any third-party tool that touches student work and review their data-processing agreement before adopting it district-wide.
Do these tells work on Claude and Gemini essays too?
Yes. The five classroom tells are about essay shape and rhetorical defaults, not phrasing tics from a single model. All major LLMs in 2026 default to the polite-assistant tone, the uniform five-paragraph structure, the generic thesis, and the "In conclusion" close when prompted for a school essay. The tells overlap across model families because they share training data on school-essay corpora.
How long should the essay be before I trust a detector?
Around 250 words is the floor where most detectors stabilise. School essays at 1,500 to 4,000 characters sit comfortably above the reliable threshold. Below 250 words, both manual spotting and detection are noisier, and the in-class writing sample carries more diagnostic weight than any tool.
Will TextSight let me check enough essays on the free tier?
Most school essays run 1,500 to 4,000 characters, so the 10,000-character daily detector cap covers several essays per session. The 10,000-character lifetime cap on the free plan is the real ceiling. Pro at $19.99 monthly or $14.99 on yearly billing removes both caps and is the right fit for teachers grading essays weekly.
Should I tell students I check for AI?
Yes. Stating the policy on day one changes drafting behaviour more than any single catch. Pair the policy with a writing process that includes in-class drafting, so the comparison sample exists when you need it. Transparency turns the scan into a teaching tool instead of a hidden trap.
Related

More for the teacher workflow.

Read the next essay in three minutes, scan only the ambiguous ones.

Free to try, no card. Sentence-level highlights, calibrated Authenticity Score, ESL-aware false-positive guardrails on every result.

Check an essay free See pricing
5 classroom tells · ESL-aware calibration · A conversation, not a verdict