Why NTU dropped its AI detector, and what that means for your child
Picture the moment a teacher looks at a child’s essay and thinks: this is too good. For a couple of years, the answer was supposed to be software. Paste the text in, get a percentage back, and the percentage tells you whether a machine wrote it.
In August, NTU said out loud what many teachers had suspected. The percentage does not tell you that.
What NTU actually decided
In an email to faculty on 13 August, reported by CNA and Mothership, NTU said it would deactivate its institutional AI detector by 31 December 2026. Associate Vice-Provost Tan Seng Chee wrote that current tools are “fundamentally unreliable” and “lack empirical validity”. He also wrote that detector scores “undermine the psychological safety” of students.
That last line matters. NTU had a reason to feel it. In 2025, three of its students were failed for using generative AI in an assignment, disputed the finding, and said they had been “falsely accused”. NTU’s own guidance already warned that detection was unreliable and biased against “non-native writing patterns”.
By September, the question had reached Parliament. One MP’s question noted that two autonomous universities had announced they were discontinuing AI detection tools over reliability concerns. MOE’s written reply on 10 September did not defend the tools. It said: “Where such tools are used, detection scores alone do not determine academic misconduct.” Instead, educators look at the student’s work and can ask for “evidence of the learning process, such as drafts or earlier work.”
Why the detectors kept failing
This is not a Singapore quirk. The people best placed to build a detector struggled to make one work.
OpenAI, the company behind ChatGPT, released its own classifier in January 2023. On its test set it correctly flagged 26% of AI-written text and wrongly labelled human writing as AI 9% of the time. In July 2023 it withdrew the tool “due to its low rate of accuracy.”
A Stanford study the same year found something worse for a multilingual city. Seven detectors were near-perfect on essays by American eighth-graders, but classified more than half of essays written by non-native English speakers for the TOEFL test (61.22%) as AI-generated. Only 2 of the 91 TOEFL essays escaped every detector.
In a multilingual city, that should give any parent pause. A tool that mistakes plain, careful English for a machine is not a neutral referee.
So what does count as proof now?
Strip away the software and the system has two answers. The first already reaches primary school.
The supervised room. On 7 October, replying to a question about checking what students can do without AI, MOE said: “National and school-based examinations are typically proctored and disallow the use of AI”. Where some AI use is allowed, such as in certain coursework, students must say how they used it and remain responsible for the work. MOE added that it will keep reviewing assessment as AI improves. Earlier in the year it said AI is introduced progressively from Primary 4, under teacher supervision, and that it keeps assessing students’ foundations “to ensure they can apply them without using AI.”
The visible process. Drafts. Earlier versions. A conversation where the student explains what they did. That is the evidence MOE described for universities, and NTU said it would lean on assessment built around “transparent process-oriented documentation”, with a standard AI disclosure form for faculty to use.
Put together, this is the long view worth holding onto. For most of schooling’s history, the finished piece of work was the proof of learning. That link is now broken, and no detector is going to repair it. Proof is moving to two places a machine cannot easily stand in: the exam hall with no devices, and the messy trail of how a piece of work came to be.
[[photo:1]]
What this means at your kitchen table
None of this is a reason to panic about AI homework. It is a reason to change what you ask to see.
- Ask for the draft, not just the final copy. A crossed-out paragraph tells you more about your child’s thinking than a polished one. Keep the scrappy versions; they are exactly the kind of evidence MOE describes.
- Have them explain one thing out loud. “Why did you start the composition there?” takes thirty seconds. If they can answer, they did the thinking, whatever tools were open.
- Practise without the tool, on purpose. The test that counts is AI-free. A weekly piece of unaided writing or problem-solving is not old-fashioned; it is rehearsal for the room they will actually sit in.
- Make declaring normal. If school allows AI on a task, help your child write one honest line about how they used it. We wrote about building that habit in show your working with AI.
MOE’s September reply to a reader’s letter in the press put the goal in one line: students are taught “not only how to use AI, but also to recognise when AI use may hinder their learning.” That is a skill no detector can measure. You can, though, a bit, every time you ask your child to show you how they got there.
Mentus AI makes a mentoring app for children, so we have a stake in how AI is used for schoolwork. This piece reports public statements from NTU and MOE and does not concern our product.
Was this useful?
Sources
- NTU to stop using AI detectors · Mothership (reporting CNA)
- Guidelines for Institutes of Higher Learning on Use of AI-writing Detection Tools · Ministry of Education, Singapore
- Measures to Assess Students' Ability to Perform Key Learning Tasks without AI Assistance · Ministry of Education, Singapore
- Correlation between Students' Use of AI and Cognitive Skills Decline · Ministry of Education, Singapore
- Schools Guide Students to Use AI Purposefully for Learning · Ministry of Education, Singapore
- New AI classifier for indicating AI-written text · OpenAI
- AI-Detectors Biased Against Non-Native English Writers · Stanford HAI