05. The "AI Detectors" Myth, Authentic Assessment & Honest Answers to Teachers' Doubts

Published 2026-09-14

Core Theme: "When students use AI to write their essays, chasing them with 'AI Detectors' is a futile war. Only a teacher's wisdom — never a piece of software — can truly measure understanding in a human mind. And here we give transparent, honest answers to the 9 deepest questions teachers ask in the AI era."

What do AI detectors actually measure?

Teachers across the world, and increasingly in Sri Lanka, have begun running student assignments through software such as ZeroGPT, GPTZero, CopyLeaks, or Turnitin's AI writing detector. The assumption is that these tools detect AI authorship. They do not, and understanding what they really measure is the difference between fair assessment and a serious injustice.

These are not detectives uncovering the truth. They measure two statistical properties of text: its predictability, known as perplexity, and the variety of its sentence patterns, known as burstiness. Human writing tends to be uneven — a long sentence, then a fragment, an unusual word choice, a slightly awkward construction. Machine writing tends to be smooth and even.

Follow that logic to its conclusion and the flaw becomes obvious. If a document is written in careful, orderly, grammatically correct prose, the software concludes it was AI-generated — because correctness and evenness are precisely what it has been built to flag.

Figure: The Technical Truth About AI Detectors (The False Positive Dilemma) — Idasara Academy

The best writer in the class

Now consider the human cost of that technical weakness.

The most gifted English essayist in the class — dictionary at her elbow, grammar book open, two late nights of honest work — hands in a flawless, polished essay. The teacher runs it through a detector. The verdict: "87% AI-generated."

What actually happened is that her essay was too correct. By the perplexity index, disciplined grammar and standard sentence patterns are machine markers. The better a child writes, the more likely she is to be branded a cheat — which inverts the entire purpose of assessment.

And there is a second group at even higher risk, which matters enormously in a Sri Lankan classroom: children learning English as a second language. They have been carefully taught standard sentence patterns and they reproduce them faithfully, which is exactly what good second-language instruction produces — and exactly what these tools read as machine output. The students most likely to be falsely accused are the ones who worked hardest to write correctly.

This is the principal reason universities around the world have suspended disciplinary action based on AI detectors.

⚠️ A hard warning for teachers: Never accuse a student of cheating, and never cut marks, on the strength of a percentage score from an AI-detector program! You may be destroying — permanently — the self-respect and the academic life of an innocent child who genuinely did the work.

Consider also the asymmetry of the error. A student who did use AI and is not caught loses one assignment's worth of learning. A student falsely accused loses their standing, their trust in the teacher, and sometimes their relationship with the subject permanently. When a test's false positives are that costly, the test should not be used for accusations at all.

Assess the process, not just the product

If detection does not work, what does? The answer is to stop grading only the final product and start grading the learning process — which removes the incentive rather than policing it.

Figure: Process-Based Authentic Assessment — Idasara Academy

The mechanism is simple. When a student must show their first rough notes and each stage of the build, not just a finished essay, there is nothing to paste from ChatGPT in an instant. A finished product can be generated in seconds; a trail of development cannot.

The four components

The idea map, 10%. Handed in handwritten, in class, on the day the assignment is set. Handwritten and in class is the whole design: it exists before any opportunity to generate anything.

Source list and draft, 30%. The first draft at mid-week, with an AI use disclosure stating openly at which step, if any, AI was used. Thirty per cent of the marks now attach to work that must have happened days before submission.

The final piece, 40%. The completed essay — still the largest single component, but no longer the only evidence.

The oral defence, 20%. The two-minute conversation described below.

The most valuable property of this scheme is what it does to the ethics of the situation. AI use stops being theft. The child's conversation with AI becomes a transparent, declared part of the learning process — exactly like an adult borrowing a library book and citing it. Nobody has to be policed, because nothing has to be hidden.

The two-minute oral defence

After collecting the assignment, call the child over and ask one of three questions.

"This important concept you've written on page 3 — explain it to me in your own words, without looking at the paper."

"What was the hardest thing to understand while doing this assignment? How did you work through it?"

"If someone argued the exact opposite of the position in your essay, what would your answer be?"

A child who understood the subject smiles and explains it. A child who printed an AI's words without reading them stands speechless.

This is the most accurate AI detector in existence, and it costs two minutes and no software. It works because it tests the thing you actually care about — whether understanding is present in the child's mind — rather than a statistical property of the text. Note also that it produces no false positives: the strong writer from the previous section answers the question easily, because she wrote the essay.

Write a classroom AI policy, not a ban

Figure: The Classroom AI Policy — Idasara Academy

A classroom that bans AI reproduces exactly the Shadow AI condition that Core Module 06 describes in government offices: the use does not stop, it goes underground — secret, unguided and unteachable. The ban removes your visibility, not the behaviour.

So instead of a prohibition, agree a short written policy with your students on day one. Four clauses cover it.

• "Using AI to understand a concept, to ask questions, and to practise is fully approved."
• "Submitting something AI wrote under your own name is plagiarism — the same as copying from a book."
• "Whenever you use AI, state at the end of the assignment which step you used it for — there is no penalty whatsoever for saying so."
• "In the exam hall and in classroom tests, AI use is banned — there, what we are measuring is your own mind."

The third clause is the one that makes the policy function. A disclosure rule with a penalty attached is a confession rule, and nobody uses it. Removing the penalty is what makes honesty the easy option — and honest disclosure is the only way you ever learn how your class is actually working.

Nine questions teachers ask — answered honestly

The same nine questions surface in every Idasara Academy workshop with Sri Lanka's O/L and A/L teachers. They deserve direct answers rather than reassurance, including where the honest answer is uncomfortable.

Q1: "Will children stop coming to my class?"

Answer: No, sir! Students come to your class for your teaching, your guidance, and the trust their parents place in you. Idasara is simply your class's digital instrument — a child joins through your Join Code, under your name. Remember: patients flock to the skilled doctor who owns a good MRI scanner!

Q2: "I have no time to learn new technology."

Answer: There is nothing new for you to learn. Student registration and system setup are done by our team. All you do is glance at the weekly report in your phone's browser — about as hard as reading a WhatsApp message!

Q3: "Are online questions as good as my printed tute?"

Answer: Your tute is your strength, sir — keep it. Keep the theory section in your own printed tute. Idasara's questions are built to the NIE syllabus and past-paper patterns, and you can also load your own questions into the app as a quiz. The difference is not in the question — it is in the instant marking and feedback. A sheet of paper will never tell you which child got the question wrong!

Q4: "What does it cost after the pilot?"

Answer: With complete transparency: teachers have a free package (up to 3 classes and 15 students). The Classroom package — Rs. 4,500 per month — has no student or class limits. The Insight package — Rs. 11,000 per month — adds the lessons needing attention and per-student analytics. Even charging 40 students just Rs. 500 each, you keep a net profit of Rs. 9,000 a month. Students can start free.

Q5: "You sell directly to children, don't you? Doesn't that make you my competitor?"

Answer: Honestly: yes — a child can use Idasara alone. But your class has something no stand-alone app can offer: a living teacher who reads the report, recognizes the symptom, and acts! Your student joins under your Join Code. What we want is for your class to post the best results in the area.

Q6: "Our children don't have good phones or internet."

Answer: Idasara runs in the browser of an ordinary smartphone — one shared phone at home is enough, and the child can practise at whatever hour suits. Whether it fits your students is exactly what we find out together in the 30-day free pilot.

Q7: "Parents won't pay an extra fee."

Answer: Parents pay for results, sir. When the child's weekly progress report lands on their phone, the value is self-evident. At the end of the pilot, show parents those reports first — and only then discuss the extra fee!

Q8: "What if the AI gives a wrong answer?"

Answer: On essays, the AI mark is only a proposal — the final mark is approved by you. The AI study assistant answers only from NIE syllabus content. And since you are the subject expert, any error you spot can be corrected on the spot.

Q9: "I do the hard work building the hub — and the teacher in the next room reaps the benefit."

Answer: Every teacher's classes and data are kept fully separate. And the first teacher to join earns the name and the head start of the area's first "AI-Powered Hybrid Learning Hub"!

Question 5 is the one worth rereading, because it is answered with a yes. A teacher weighing this decision is entitled to know where the genuine tension lies rather than being told there is none — and the answer to it is also the answer to question 1. A child can use the app alone; what the app cannot supply is the teacher who reads the report, recognises what it means, and does something about it on Monday morning.

In the final chapter: student data privacy, the 30-Day Free Pilot, and the unified teacher vision of Idasara Academy.