A Teacher's Guide to AI-Written Assignments: How to Spot Them and What to Do

Here's a story that should make every teacher stop and think.

In 2023, researchers at the University of Reading in England secretly entered AI-written answers into real university exams. Nobody was told. The markers were ordinary lecturers doing their normal marking.

The result? 94% of the AI answers weren't spotted. By our sum, that's only about 1 in every 17 caught. And on average, the AI answers scored higher than the real students' work.

These were take-home exams, and the markers weren't using any AI checker. But it shows the problem clearly: AI writing is now good enough to slip past experienced teachers.

So what can a teacher in Nigeria actually do? Let's go through it, step by step.

First, why AI checkers aren't the answer

Many teachers hope a website or app can just tell them, "This was written by AI." Sadly, it's not that simple.

Even the makers of ChatGPT gave up on their own checker

In January 2023, OpenAI, the company behind ChatGPT, released a tool to spot AI writing. On its own tests, it caught only 26% of AI writing, about 1 in 4. And it wrongly flagged 9% of human writing as AI.

By our sum, in a class of 40 honest students, that's about 4 students wrongly accused. In July 2023, OpenAI quietly shut the tool down because it wasn't accurate enough.

Turnitin's own numbers

Turnitin is the checker many universities use. In 2023, Turnitin said it wrongly flags less than 1% of whole documents, but only for papers where it finds 20% or more AI writing. Sentence by sentence, it said its false alarm rate is "around 4%".

Less than 1% sounds small. But here's what Vanderbilt University in America worked out. It sent 75,000 papers to Turnitin in 2022. At a 1% error rate, about 750 student papers could have been wrongly labelled as AI. So in August 2023, Vanderbilt switched Turnitin's AI checker off. It said:

"we do not believe that AI detection software is an effective tool" Vanderbilt University, August 2023

Let's bring that home. Say your school has 600 students, and each writes 10 assignments a term. That's 6,000 pieces of work. By our sum, even at a 1% error rate, about 60 pieces of honest work could be flagged as AI in one term. Each one is a child facing a false accusation.

A little editing beats most checkers

There's another problem. Students can easily fool these tools.

In a 2023 study, researchers ran AI writing through a rewording tool. One checker went from catching 70.3% of AI texts to just 4.6%. By our sum, that's from about 7 in 10 caught to fewer than 1 in 20. That was one checker in one test, but other studies found the same pattern: checkers are easy to fool and often call AI writing "human".

So the student who cheats carefully gets away with it. The honest student who happens to write a certain way gets flagged. That's the worst of both worlds.

The danger for Nigerian students

This part matters a lot for us.

In 2023, Stanford University researchers tested seven AI checkers on essays by people writing in English as a second language. On average, the checkers wrongly called these human essays AI-written 61% of the time. Essays by American 8th-graders, writing in their first language, were judged correctly.

It gets worse. By our count from the study, 97.8% of the second-language essays were flagged by at least one of the seven checkers. Put simply, if you check an honest second-language essay with seven tools, it's almost certain one of them will call it AI.

Why? These checkers look for simple, predictable wording. Many people writing in a second language use simpler words and sentence patterns. So the checker mistakes careful, simple English for a machine. The lead researcher, James Zou, put it bluntly:

"Current detectors are clearly unreliable and easily gamed" James Zou, Stanford University, May 2023

For most Nigerian students, English is a second language. That puts them right in the group these tools treat unfairly.

To be fair to Turnitin, it wasn't one of the seven tools tested. Turnitin says its own testing found no clear bias against English learners for texts of 300 words or more. But that's Turnitin testing itself. Be careful.

Can teachers just tell? Mostly, no

If machines can't tell, surely an experienced teacher can? The research says we're not as good as we think.

In a 2024 study, 200 experienced English teachers looked at essays written by secondary school students and by ChatGPT. They spotted only 37.8% of the AI essays. By our sum, about 6 in 10 AI essays got past them.

And here's the twist: they felt about 80% sure of their answers. By our sum, they were about twice as confident as they were accurate on the AI essays.

Trainee teachers in the same study did no better than guessing. A separate 2023 American study found high school teachers picked the AI essay correctly only 70% of the time, even when told one of two essays was AI.

So "I can just tell" isn't enough to accuse a child.

It's happening in Nigeria too

This isn't a foreign problem.

In February 2025, UNILAG professor Isaac Nwaogwugwu told Nairametrics about an assignment he gave his MBA class. Out of over 100 students:

"about 40% submitted the exact same answers" Prof. Isaac Nwaogwugwu, UNILAG, February 2025

He said the students didn't even know each other. They had all used the same AI tool. In the same report, Dr Felix Echekoba of Nnamdi Azikiwe University said that when students who copy from AI are asked basic questions, "they go blank".

A 2025 study at LAUTECH asked 152 final-year students about their projects. About half had used ChatGPT in some way, most often to write the abstract.

The authorities are starting to act. In February 2025, UNILAG announced it was preparing a policy on ethical AI use. In October 2026, the National Universities Commission (NUC) said it is developing a national AI policy for universities. But we found no rule yet from WAEC, NECO or any national body on AI-written schoolwork. For now, teachers are on their own.

Red flags worth a closer look

None of these proves anything alone. Think of them as reasons to look closer and have a chat, not reasons to punish.

1. A sudden jump in quality. The student who writes simply in class suddenly hands in polished, grown-up work, with no gradual improvement.
2. Nothing personal or local. The essay could have been written about any school, anywhere. No class discussion, no local example, no real experience.
3. References that don't exist. AI often invents sources. In a 2023 study, 47% of the references ChatGPT wrote in medical papers were made up, and only 7% were real and correct. By our sum, that's almost 7 fake references for every accurate one. Newer AI does better, but invented references are still common. Look up two or three.
4. The same answer from different students. Just like at UNILAG. If several students who don't sit together hand in nearly the same words, that's worth a conversation.
5. Left-over chatbot text. A Florida college professor told Gizmodo in 2025 that some students had left their chatbot instructions in the essays they handed in. It happens.
6. Vague, general words. Lots of big words, but little real meaning. Ask yourself, "What exactly is this student saying?"

Signs to ignore: long dashes, missing paragraph indents, or neat formatting. Many students type on phones or use apps that change these things. They prove nothing.

What actually works

1. Ask the student to explain it

This is the simplest and best check. Sit with the student for two minutes and ask:

• "Tell me in your own words what your main point is."
• "Why did you choose this example?"
• "What does this word in your essay mean?"

A student who did the work can usually explain it, even if they're nervous. A student who didn't often can't. The UK's university quality body, QAA, recommends short oral checks like this, even small group chats about written work.

2. Get a sample of their real writing early

In the first weeks of term, have every student write something short in class, by hand, with no phones. Keep it. Later, if something seems off, you have a real sample to compare with.

3. Mark the steps, not just the final answer

Ask for an outline first, then a rough draft, then the final work. Give a few marks for each stage. Students who do the thinking will have the steps. Students who copy from AI usually won't.

4. Set questions AI can't easily answer

AI is good at general answers. It's weak at your class, your town and your students' lives. For example:

• "Interview an older person in your family about how they did their exams. Compare it with yours."
• "Use the prices of three foodstuffs at your nearest market this week to explain inflation."
• "Pick one point from our class debate on Tuesday and argue against it."

The QAA put it well: building these skills over time "avoids the immediate need to deploy unreliable detection software".

5. Do more writing in class

The surest way to know who wrote something is to watch them write it. Short in-class writing, even 15 minutes, is much harder to fake.

If you suspect AI: a fair way to handle it

1. Don't accuse based on a checker score. At most, it's a reason to look closer.
2. Look for more than one red flag. One sign alone means little.
3. Talk privately and calmly. Ask the student to explain their work. Don't start with "You cheated."
4. Listen. There may be an honest reason. Maybe an older sibling helped, or they used AI for grammar and didn't know that wasn't allowed.
5. Focus on learning. Often the best outcome is redoing the work in class, not a punishment. The goal is a student who can think, not a student who's scared.
6. Follow your school's rules, and write down what happened.

Make your AI rules clear

Many students genuinely don't know what's allowed. Tell them, in writing, at the start of term. For example:

Allowed (if you tell me you used it):

• Asking AI to explain a topic you don't understand.
• Asking AI for practice questions.
• Checking spelling and grammar in work you wrote yourself.

Not allowed:

• Getting AI to write your essay, answers or project.
• Copying AI text and handing it in as your own.

Ask students to add one line at the end of their work: "I used AI for..." or "I didn't use AI." Make telling the truth normal and safe. A student who can admit using AI is a student you can teach.

A quick checklist for this term

1. Collect an in-class writing sample from every student in the first two weeks.
2. Write down your AI rules and share them with students and parents.
3. Mark outlines and drafts, not just final work, for big assignments.
4. Add at least one local or personal question to each major assignment.
5. Never punish based on an AI checker score alone.
6. When in doubt, ask the student to explain their work.

If you teach online, Scholarleveling lets you run live classes with a shared whiteboard and class notes, so you can watch your students work things out in real time, not just mark what they hand in.

Read next: what AI is doing to children's thinking, or the right way for children to use AI.