AI Checker for Teachers: What the Scores Mean and What to Do Next

An AI checker for teachers estimates how likely it is that a piece of writing came from a chatbot. It cannot prove that it did. Use the score to decide which essays deserve a closer look, then decide with evidence the score can't give you: the student's earlier writing, their drafts, and a conversation about the work. Never let a detector score be the only reason for a lower grade or a cheating report.

That advice isn't caution for its own sake. OpenAI shut down its own AI text classifier in 2023 because it wasn't accurate enough, and independent studies have found that detectors flag honest writing, especially writing by English learners. Turnitin, whose detector is common in schools, says its AI writing report shouldn't be the sole basis for action against a student.

This guide covers what an AI checker actually measures, what the research says about accuracy, how to read a score, and a step-by-step process for a flagged essay that is fair to the student and still takes AI misuse seriously.

What an AI checker actually measures

AI checkers don't find a hidden watermark or match the essay against a database of chatbot output. Most of them are classifiers: models trained on large sets of human writing and AI writing that learn the statistical differences between the two.

The most common signal is predictability. Large language models write by choosing likely next words, so their prose tends to be smooth, even, and low in surprise. Researchers call this low perplexity. Human writing usually varies more from sentence to sentence. A detector looks at those patterns across a passage and outputs a probability that the text was machine-generated.

That design explains most of the problems teachers run into:

• Clear, simple writing can look like AI. A student who writes short, plain sentences, follows a template closely, or is still learning English produces predictable text. So does a strong student who writes very clean prose.
• Short text is unreliable. A paragraph or a short answer gives the model too little to go on. Most vendors warn about results on short submissions.
• Light editing changes the score. Running AI text through a paraphrasing tool, or having a student rewrite it in their own words, can move a score from "AI" to "human." Detection is easy to evade for students who know how.
• The number means different things in different tools. Depending on the checker, 60% can mean the model is 60% confident the whole essay is AI, or that it thinks 60% of the sentences are. Check what your tool means before you read anything into the number.
• Handwriting is outside its reach. A checker reads typed text. Work written by hand in class is the one format it can't touch, which is also why in-class writing is useful evidence.

None of this makes a detector useless. It makes it a screening tool, the same way a quick spelling check tells you where to look but doesn't grade the essay.

How accurate are AI detectors?

Accuracy depends on the tool, the length of the text, and who wrote it. The evidence so far points the same way: detectors catch a lot of unedited AI text, they miss edited AI text, and they flag some human writing.

OpenAI's own classifier. When OpenAI released its AI text classifier in January 2023, it reported that the tool correctly labeled 26% of AI-written text as "likely AI-written" and wrongly labeled human writing as AI 9% of the time. OpenAI retired the classifier in July 2023, citing its low rate of accuracy.

English learners get flagged more. A Stanford study published in the journal Patterns tested seven popular detectors on essays written by non-native English speakers for the TOEFL exam. The detectors labeled more than half of those essays as AI-generated, with an average false positive rate of 61%, while they were close to perfect on essays by US eighth graders. Nearly all of the TOEFL essays were flagged by at least one detector. The authors traced the gap to simpler word choice, which reads as predictable.

Independent testing of 14 tools. A team of academic integrity researchers tested 12 public detectors and two commercial systems, including Turnitin. They concluded that the tools are neither accurate nor reliable enough to serve as evidence, that they lean toward calling AI text human, and that light obfuscation makes them much worse.

Vendor numbers. Turnitin reports a document-level false positive rate below 1% for documents it scores above 20% AI, and it marks scores below 20% with an asterisk because they are less reliable. That figure comes from Turnitin's own testing.

Who gets accused. A 2024 Common Sense Media survey of teens found that 20% of Black teens said their schoolwork had been wrongly flagged as AI-generated, compared with 7% of white teens and 10% of Latino teens.

Why a 1% error rate still matters in a classroom

A false positive rate that sounds small adds up fast. If you teach 150 students and check every essay, a 1% rate means one or two students who did their own work get flagged on each assignment. Over eight essays in a year, that is about a dozen honest essays that look suspicious. Vanderbilt University ran the same math on its 75,000 papers a year, estimated that around 750 could be wrongly labeled, and turned off Turnitin's AI detector in 2023.

The cover image of this article shows that math: 150 essays, two of them flagged, both written by the student. If the score alone decides the outcome, those two students pay for the tool's error.

Which AI checker do teachers use?

Most teachers use whatever their school already pays for. A nationally representative survey by the Center for Democracy and Technology found that a majority of middle and high school teachers had used an AI detection tool. The common options fall into three groups:

• Detectors inside plagiarism checkers. Turnitin's AI writing report comes with many school and district licenses, so it shows up in the same place teachers already check for copying.
• Standalone detectors. Tools such as GPTZero, Copyleaks, Originality.ai and Pangram let you paste text or upload files. Several sell educator plans and LMS add-ons, and many offer a limited free tier.
• Detection built into grading tools. Some AI grading tools, GradeWithAI included, can check each submission while they grade it, so the score sits next to the feedback instead of in a separate tab.

Rankings of "the best AI detector" change constantly and are often written by detector companies. Rather than chase the top of a list, judge any AI checker for teachers by the questions that matter in a classroom:

• Does it publish its false positive rate and how it was measured? A vendor that won't say is asking you to take the risk on trust.
• Does it show which passages drove the score? Sentence-level highlighting helps you read the essay; a single number doesn't.
• What does it do with student work? Check how long it keeps submissions, whether it trains models on them, and whether it has signed your district's student data privacy agreement.
• How does it handle short or non-English text? Look for a minimum length and clear warnings rather than confident scores on a paragraph.
• Does it fit where you grade? A checker you have to copy and paste into will get used inconsistently, which is its own fairness problem.

How to read an AI detection score

Treat a detection score the way you'd treat a smoke alarm: worth checking, often a false alarm, and never the same thing as a fire. The useful question is what the score should change about what you do next.

GradeWithAI's AI detection, for example, gives each submission a score from 0 to 100% and one of four labels. Here is a reasonable way to act on each:

• Human Written (below 25%). Grade the work as usual.
• Mostly Human (25% to 50%). Also grade as usual. Formulaic assignments, short responses, and very polished writing often land here.
• Mostly AI (50% to 75%). Look closer before grading. Read the essay beside the student's earlier work and check the draft history.
• AI Generated (75% and up). Hold the grade, gather process evidence, and talk with the student before you decide anything.

Two habits keep the score in its place. First, read the essay before you look at the score, so the number doesn't color how you judge the writing. Second, compare the score with what you know: a student who has written in this voice all year is not suddenly suspect because a model says so.

What to do when an essay gets flagged

A flag is the start of a process, not the end of one. The steps below protect a student who did their own work and still give you solid ground if they didn't.

1. Pause the grade. Don't enter a zero, don't comment on it in front of the class, and don't mention the score yet. Mark the assignment as pending if your gradebook allows it.
2. Compare it with the student's earlier work. Put the essay beside two pieces the student wrote before, ideally one written in class. Look at vocabulary, sentence length, typical errors, and how they build an argument. A sudden, total change in voice is meaningful. A student who simply had a good week is not.
3. Check the drafts. Google Docs keeps a version history, and Canvas keeps earlier attempts when students resubmit. A document that grew over several sessions, with deletions and rewrites, looks very different from one that appeared in a single paste. Outlines, notes, and annotated sources count too.
4. Talk it through. Meet with the student privately and ask about the work, not about AI. "Walk me through how you got to this thesis." "Why did you pick this quote over the others?" "What would you change if you had another day?" A student who wrote the essay can usually explain it, even if they're nervous. Some will tell you what happened on their own once the conversation is about the writing.
5. Decide by your school's policy, and write it down. Follow the academic integrity process your school uses, involve whoever it says to involve, and note what evidence you relied on. If the evidence beyond the score is thin, give the benefit of the doubt. A redo or an in-class rewrite is often a better outcome than a penalty either way.

What you should not do is ask a chatbot whether it wrote the essay. Chatbots will happily say yes to text they never saw.

Make AI use visible before you need a detector

The most reliable fix is in the assignment, not the software. MIT Sloan's teaching team, in a guide titled AI Detectors Don't Work. Here's What to Do Instead, recommends clear policies, open conversation with students, and assignments built so the work itself shows the thinking. In an English classroom that can look like this:

• Say what's allowed on each assignment. "AI can help you brainstorm, but every sentence you submit must be yours" is clearer than a blanket ban. Students cross lines they can't see.
• Grade the process, not just the product. Collect a thesis statement, an outline, and a rough draft before the final essay. Each checkpoint is a small grade and a sample of the student's real voice.
• Get a baseline in class. A short in-class writing sample in September gives you something to compare against all year, and it's handwritten, so no detector is needed.
• Ask for what a chatbot can't know. Prompts tied to class discussion, a local issue, or the student's own experience are harder to outsource and more interesting to read.
• Add a short reflection or conference. A paragraph on "the hardest part of writing this" or a two-minute conversation makes ownership obvious.
• Give students a legitimate way to get feedback. Much AI misuse is a student stuck at 11 p.m. A feedback tool such as our free essay checker, which scores thesis, evidence, organization and mechanics and points to specific sentences, gives them help that keeps the writing theirs.

These changes take some planning, but they also make better assignments. Process grades, conferences, and personal prompts were good teaching before chatbots existed.

Using AI detection inside GradeWithAI

If you want an AI checker that sits inside your grading instead of in another tab, GradeWithAI includes AI detection on the Pro plan. Turn it on when you set up an assignment, and each typed submission gets a 0 to 100% AI score and a label next to its grading feedback. Photos and scans of handwritten work are not checked. It works the same way for assignments synced from Google Classroom or Canvas and for files you upload.

Our detector has the same limits as every other detector described above. We show the score so you can decide where to look closer, and the grade and feedback stay drafts until you review them. If you want to see how the grading side works, our guide to AI essay grading walks through it, and Can ChatGPT Grade Essays? covers how to review AI feedback before it reaches students.

Frequently asked questions

Can an AI checker prove a student used ChatGPT?

No. A detector reports a probability based on how predictable the writing is. It can be wrong in both directions, and no detector vendor claims its score is proof. Evidence of misuse comes from the writing process: drafts, version history, comparison with earlier work, and the student's ability to explain what they wrote.

What AI checker do most teachers use?

Most teachers use the detector their school already licenses, most often the AI writing report inside a plagiarism checker such as Turnitin. Standalone tools like GPTZero are also common, and some grading tools include detection. The tool matters less than how you use the score.

Are free AI detectors accurate enough to grade with?

No detector, free or paid, is accurate enough to decide a grade on its own. Free tiers often add tighter length limits and less detail about which passages drove the score. Use any detector as a prompt to look closer.

Do AI checkers work on short answers?

Poorly. Most detectors need several hundred words to produce a stable result, and scores on a paragraph or a short answer swing widely. For short responses, comparing the answer with what the student says in class is more useful than any score.

Is it fair to run English learners' essays through an AI detector?

Be very careful. Research has found that detectors flag essays by non-native English writers far more often, because simpler word choice reads as predictable. If you use detection at all, never act on a score for an English learner without strong evidence from their drafts and earlier work.

Can students get around AI checkers?

Yes. Paraphrasing tools, translation, and light rewriting all lower detection scores, and researchers found that this kind of obfuscation makes detectors much less accurate. That's why assignment design and process checkpoints do more to discourage misuse than detection does.

Should I tell students I use an AI checker?

Yes. Explain your AI policy, tell students how you use detection, and say that a score alone won't decide anything. Students who know the process is fair are more likely to ask for help than to look for a shortcut.

Sources

• OpenAI, New AI classifier for indicating AI-written text (2023, with the July 2023 retirement note)
• Liang et al., GPT detectors are biased against non-native English writers, Patterns (2023)
• Weber-Wulff et al., Testing of detection tools for AI-generated text, International Journal for Educational Integrity (2023)
• Turnitin, AI writing detection update from Turnitin's Chief Product Officer (2023)
• Vanderbilt University, Guidance on AI detection and why we're disabling Turnitin's AI detector (2023)
• Education Week, Black students are more likely to be falsely accused of using AI to cheat (2024)
• Center for Democracy and Technology, Up in the air: educators juggling the potential of generative AI with detection, discipline, and distrust (2024)
• MIT Sloan Teaching & Learning Technologies, AI Detectors Don't Work. Here's What to Do Instead