How to Check Student Essays for AI Without False Accusations

To check student essays for AI fairly, treat a detector score as a reason to look closer, never as proof. Screen the essay with a free AI detector such as Essai's, then compare it with the student's earlier work, verify its sources, review drafts or version history, and talk with the student before you reach any conclusion.

This guide is for teachers, professors and TAs who want a process they could explain to a student, a parent or an integrity panel. It draws on guidance from Vanderbilt University, the University of Kansas and MIT Sloan, Turnitin's published statements on false positives, peer-reviewed research and news reports, as of October 2026. For the screening step, Essai's free AI detector is our top pick: it gives you a clear AI probability score from 0 to 100 percent, and a free account adds detailed reports and unlimited scans for a whole class set.

The situation: one essay, one score, and a call you have to defend

You're working through a stack of essays and one of them reads differently. The prose is smooth and evenly paced, but it says almost nothing about the readings or anything discussed in class.

Maybe a detector your school uses has already flagged it. Maybe it's only a feeling. Either way, you now have to decide what to do, and the decision will matter to that student.

Two mistakes are possible. Miss real misuse, and students who did their own work are graded on a different scale from those who didn't. Accuse an honest student, and you damage trust that is hard to rebuild.

False positives aren't hypothetical. When Vanderbilt University announced in August 2023 that it was disabling Turnitin's AI detector, it worked through what Turnitin's claimed 1 percent false positive rate would mean on its own campus. Vanderbilt had submitted 75,000 papers to Turnitin in 2022. Had the detector been running then, around 750 of them could have been incorrectly labeled as partly AI-written.

Turnitin has made the same point from its side. In a May 2023 update, its chief product officer wrote that "behind each false positive instance is a real student who may have put real effort into their original work." A fair process is built for that student as much as for the one who cut corners.

How to tell if a student used AI: what teachers notice

Here is a typical AI-style paragraph, written for this guide in response to a common history prompt:

The Industrial Revolution was a pivotal moment in human history that fundamentally transformed the way people lived and worked. On one hand, it brought unprecedented economic growth and technological innovation. On the other hand, it created significant challenges for workers, including harsh conditions and long hours. Ultimately, its legacy is complex, and its impact continues to shape our world today.

Here's what makes it read as machine-made:

• Nothing specific. No country, decade, industry, source or number appears, even though the assignment almost certainly came with readings.
• Balance instead of an argument. "On one hand... on the other hand" sets up a debate and then walks away from it.
• Stock phrasing. "Pivotal moment," "fundamentally transformed" and "continues to shape our world today" could close an essay on almost any period.
• Even rhythm. The sentences run to similar lengths and follow the same pattern.
• A summary posing as a conclusion. The last line tells the reader the topic is complex and stops there.

None of that is proof. Students write this way too, especially when they're unsure of the material, short on time or writing in a second language. Turnitin's March 2023 FAQ on its AI writing detection, a PDF hosted by the University of Kansas, listed text that "literally repeats itself" and text "paraphrased without developing new ideas" among the kinds of human writing that can produce false positives.

The stronger signals usually sit outside the prose:

• A break from the student's own record. Vanderbilt's guidance suggests comparing a submission with the student's other work for style, tone and level.
• Sources that don't hold up. The same guidance warns that AI text generators can "hallucinate" and invent whole sources, so a citation that doesn't exist tells you more than any score.
• No trace of your course. Note whether the essay engages with your lectures, discussions or assigned readings. Vanderbilt suggests tying assignments to topics discussed in class as one way to ease concerns about AI use.
• A student who can't explain the work. The University of Kansas teaching center suggests asking students to explain any terms or phrasing that seem out of place for the assignment.

Why an AI detector score is a signal, not proof

Are AI detectors accurate?

Not reliably enough to carry a decision on their own. A 2023 study of 14 detection tools, later published in the International Journal for Educational Integrity, covered 12 publicly available tools plus two commercial systems, Turnitin and PlagiarismCheck. Its authors concluded that the available detection tools "are neither accurate nor reliable."

OpenAI, the company behind ChatGPT, shut down its own AI text classifier in July 2023. A note on the tool's announcement page said it was no longer available "due to its low rate of accuracy."

Turnitin's own documentation says much the same. Its March 2023 FAQ stated that the AI percentage "should not be used as the sole basis for action or a definitive grading measure by instructors." Two months later, the company said real-world results differed from its lab tests and that documents where it detected less than 20 percent AI writing showed a higher incidence of false positives. It reports a document-level false positive rate under 1 percent for documents with more than 20 percent AI writing, and it added: "We cannot mitigate the risk of false positives completely."

AI detector false positives hit some students harder

A 2023 study by Liang and colleagues, published in Patterns, illustrates the problem. The researchers ran seven widely used detectors on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth graders. The detectors were near-perfect on the eighth-grade essays but, on average, misclassified more than half of the TOEFL essays as AI-generated. All seven flagged 18 of the 91 TOEFL essays, and at least one flagged 89 of them.

The authors note that their samples were relatively small, but the gap between the two groups was wide. The essays that every detector flagged used more predictable wording, which the authors link to the narrower vocabulary and phrasing common among people writing in a second language. They cautioned against using detectors "in evaluative or educational settings." In a Stanford HAI article, the study's senior author, James Zou, said that for now detectors shouldn't be relied on in educational settings, especially where many students are non-native English speakers.

Turnitin was not part of that study. In its own October 2023 evaluation, the company reported that the difference in false positive rates between English learners and native writers was small and not statistically significant on submissions of 300 words or more, but larger on shorter texts.

Formulaic writing carries risk too. Turnitin's FAQ, cited above, also named lists without much structural variation among the sources of false positives, so take extra care with lab reports, five-paragraph essays and anything written to a strict template. Our explainer on AI detection false positives covers the other common causes.

Low scores don't clear an essay either

Detectors miss AI text as well. The International Journal for Educational Integrity study found the tools leaned toward classifying text as human-written, and that obfuscation techniques made their performance significantly worse. In Liang's study, a second prompt asking ChatGPT to rewrite its own college admission essays in more literary language cut detection rates from as high as 100 percent to 13 percent or less.

AI humanizers, Essai's included, are designed to rewrite phrasing, rhythm and word choice so that a draft reads naturally. That's why a low score can't tell you how a draft was produced. Our guide to detecting AI writing without false positives goes further into why every score needs context.

How to check student essays for AI in 7 steps

Use these steps to check student essays for AI with Essai's free detector, our top pick for this job, or with whichever detector your school provides. Steps 5 to 7 carry more weight than the score.

Step 1: Set the rules before the essays arrive

Fairness starts before anyone submits. Put your AI rules in the syllabus and on the assignment sheet, and say them out loud in class. In its guide AI Detectors Don't Work. Here's What to Do Instead, MIT Sloan's teaching team recommends announcing AI policies "both in person and in writing" and giving concrete examples. In their example, students may use AI to brainstorm or check grammar but not to write large parts of an essay.

Then tell students how you'll handle concerns. Turnitin's advice to instructors in its March 2023 post on false positives is to "know before you go": consider the possibility of a false positive up front, plan your process and share it with students. Ask for process evidence as part of the assignment, such as an outline, a draft, shared version history or a short process statement that explains how the student completed the work and whether they used AI. MIT Sloan also suggests leaving room for students to disclose AI support without penalty.

If your course allows AI-assisted drafting with disclosure, our student-side guide to revising a ChatGPT draft within course rules shows the kind of disclosure note and saved versions you can ask students to keep.

Step 2: Read and grade the essay first

Read the essay as you would any other and grade it against your rubric. If something feels off, write down exactly what: the paragraph, the claim, the citation, the shift in voice. Noting your observations before you run any check keeps a number from shaping what you notice.

Many concerns end here. An essay can be weak, generic or off-prompt without being AI-written, and your rubric already handles that.

Step 3: Screen it with an AI detector

Paste the student's prose into Essai's free AI detector, our top pick for this step. Leave out the title page, the reference list and long block quotations, since none of them shows how the student writes. The detector returns an AI probability score from 0 to 100 percent and works best on 150 or more words, so scan the full essay rather than a single paragraph. Length matters for every tool: Turnitin raised its own minimum from 150 to 300 words in 2023 after finding more false positives on short submissions.

A free Essai account adds detailed reports and unlimited scans, which helps when you're working through a whole class set. Essai runs in your browser with nothing to install, so you paste each essay in, read the score and move on to the next one.

Before you paste anything, check your school's rules on sharing student work with outside tools. Vanderbilt's guidance flagged "real privacy concerns" about entering student data into a detector managed by a separate company. At a minimum, remove names, student IDs and other personal details.

Step 4: Read the result as a signal

An AI probability score is an estimate about the text. It says nothing about who wrote it, what your course allowed or how the student worked. Here is how to read a result from any detector:

Add the context the tool can't see. Is the student writing in a second language? Is the genre formulaic by design? Did your rules allow AI for part of the task?

Scan once, with one tool, and record the result along with the date and the text you scanned. Don't run the essay through detector after detector looking for agreement. In Liang's study, at least one of seven detectors flagged 89 of 91 human-written TOEFL essays, while all seven agreed on only 18. The more tools you try, the more likely one of them flags honest work.

Step 5: Gather evidence a detector can't give you

• Compare with earlier work. Put the essay next to earlier assignments or an in-class writing sample. The University of Kansas teaching center lists style, tone, spelling, flow, complexity, development of argument and use of sources as points of comparison.
• Check every citation. Open each source. Confirm that it exists, that quoted words appear in it and that it supports the claim attached to it.
• Ask for the drafting record. Outlines, notes, drafts and version history show how an essay developed. In Google Docs, only people with edit permission can browse earlier versions, so ask the student to share edit access or walk you through the history. Google also notes that revisions may occasionally be merged, so a thin history proves nothing on its own.
• Look for your course in the essay. Note whether it uses your lectures, discussions or assigned readings, and whether it gets them right.

Step 6: Talk to the student about AI use

Meet privately, and open with the work, not the score. "Walk me through how you wrote this" invites an explanation. "The detector says this is AI" invites a defense. KU's teaching center suggests having students explain their research and writing process and talk you through the subject matter.

Be honest that detectors make mistakes. Turnitin warns that failing to acknowledge a possible false positive leads to "a far more defensive and confrontational interaction." Its advice is to assume positive intent and, when the evidence is unclear, to assume students will act with integrity.

Keep the meeting a conversation rather than a quiz. MIT Sloan's teaching team points out that oral presentations are especially stressful for students with anxiety and for non-native English speakers, so keep the tone calm and give the student time to answer. Vanderbilt's guidance adds that "some students will just admit the use upfront if approached correctly."

Step 7: Decide, document and follow policy

If doubts remain after the conversation and your policy allows it, KU's teaching center suggests a second chance: have the student redo the assignment, with specific feedback on the problem areas. If you're certain a student misused AI, follow your institution's academic integrity procedure and be ready to document how you reached that conclusion. KU's advice is blunt: you are "far better off erring on the side of caution" than pouring time and energy into a misconduct claim "that may not hold up."

Write down what prompted the concern, which tool you used, when and on what text, the comparison and citation evidence, what the drafting record showed and what the student said. Your record should stand on its own without the score.

A fairness checklist before you raise a concern

• Your AI rules for this assignment were in writing before students submitted.
• You read and graded the essay before running any detector.
• You scanned only the student's own prose, at least 150 words, with personal details removed.
• You ran one detector once and recorded the tool, the date and the text scanned.
• You compared the essay with the student's earlier or in-class writing.
• You opened every cited source.
• You asked for drafts, notes or version history.
• You considered language background, genre and anything else that makes writing more formulaic.
• You talked with the student before deciding anything.
• Your notes would persuade a colleague who never saw the score.

Mistakes to avoid

• Treating the score as the finding. It's an estimate about text. The decision is yours, and it needs other evidence.
• Shopping for a second opinion. Trying detector after detector until one matches your hunch raises the odds of a false positive.
• Scanning fragments. A single paragraph, a list or a short answer gives any detector too little to work with.
• Reading a low score as a clean bill. Revised or reworded AI text can score low.
• Opening with an accusation. Leading with the detector result, especially in an email, makes an honest explanation harder to hear.
• Changing the rules after submission. Judge the work by the policy students had when they wrote it.
• Forgetting who gets flagged. Non-native English writers and students working in formulaic genres face more risk of false positives.

Policy, privacy and fairness

Your institution's academic integrity policy comes first, and it may rule out detectors altogether. Vanderbilt disabled Turnitin's AI detector in 2023 and concluded, "we do not believe that AI detection software is an effective tool that should be used." If your school has taken a similar position, follow it. This guide is for settings where a detector is allowed as one input among several.

Privacy deserves the same care. Use tools your school has approved, strip identifying details, and read any detector's privacy policy before you paste student work, Essai's included. If you have questions about how Essai handles the text you paste, you can reach the team through the contact page or at [email protected] before you start.

Fairness also means making disclosure safe. When students know they can report AI help without automatic penalty, you get honest information instead of guesswork. MIT Sloan adds that making disclosure normal lowers the stigma for students who may need accommodations.

Keep the tool in proportion. KU's teaching center argues that "detection software will never keep up with the ability of AI tools to avoid detection." Clear rules, thoughtful assignment design and honest conversation will do more of the work than any score.

Why Essai is our top pick to check student essays for AI

Essai's free AI detector is our top pick for the screening step in this process. It does the one job a detector should do here, giving you a clear signal about the text, and it leaves the decision to the evidence and the conversation in Steps 5 to 7.

• A score you can read and record. You get an AI probability score from 0 to 100 percent, easy to log with the date and the text you scanned in Steps 4 and 7.
• Free for a whole class set. The detector is free to use, and a free account adds detailed reports and unlimited scans, so you never ration checks or count credits.
• Suited to full essays. It works best on 150 words or more, which matches the advice to scan the student's full prose rather than a single paragraph.
• Nothing to install. It runs in your browser. Paste the essay, read the score and move on to the next one.
• No paid plan required. Paid plans start at $14.99 a month, or $9.99 a month billed yearly, as of October 2026, and every one includes unlimited AI detection, but you don't need one to screen essays. See Essai pricing for the details.

Here is how free and paid access compares with three detectors that market to schools and educators, as of October 2026:

Essai gives you the most free detection and the lowest yearly price of the four. A class set of 30 essays at 1,000 words each adds up to 30,000 words, three times GPTZero's free monthly allowance, while a free Essai account covers it with scans to spare. Our GPTZero review, Originality.ai review and Winston AI review cover each tool's plans in detail.

"I've tested many alternatives, but nothing comes close to the quality Essai provides." (Priya Sharma, Essai user)

Like any detector, Essai gives you a probability about the text, not a verdict about the student, so its score opens Step 5 rather than replacing it. If your school already licenses a different detector, the same seven steps apply. Our reviews of Copyleaks and ZeroGPT cover more detectors, including what each company says about its own accuracy, and our ranking of the best AI detectors, where Essai takes the top spot, compares nine tools side by side. Editors who review freelance copy will find a similar workflow in our guide for content teams.

The bottom line

An AI detector can tell you where to look. It can't tell you what happened. To check student essays for AI fairly, set the rules early, read first, scan once, gather evidence a detector can't give you and talk with the student before you decide anything. That process protects honest students and leaves you with a record you can stand behind.

For the screening step, Essai's free AI detector is our top pick: a clear AI probability score from 0 to 100 percent, with detailed reports and unlimited scans on a free account. Set one up before your next stack of essays arrives.