Almost every teacher's primary AI-related challenge today is student use of AI to produce work. "It just upsets the whole dynamic of my classes. Effort and fairness just go out the window. It's not just frustrating—no one seems to have a good answer for this. We can't physically ban AI for students, but asking us teachers to re-invent everything we do is just overwhelming!"
Schools have tried detection tools, watermarking, secure browsers, activity tracking. None of it sticks. The question worth asking is not whether AI wrote it but whether the student can explain it.
When Anthropic announced that text produced by Claude would carry invisible, machine-detectable marks, much of the coverage read it as one more tool for catching students who cheat; the person best placed to explain why that reading is mistaken is, rather neatly, an executive at the company behind Canvas.
Melissa Loble, Chief Academic Officer at Instructure, makes the argument in a Forbes Business Development Council piece published on 5 October. Her starting point is that the watermark exists, by Anthropic's own account, to meet the transparency requirements of the EU AI Act, which asks for AI-generated content to be machine-readable as such; it is a measure for regulators, and turning it into a classroom enforcement device converts a compliance feature into a blunt instrument. (It is worth saying that she is writing as an edtech executive and offering opinion rather than research, and that I have relied on her account of the announcement without checking the announcement itself.)
A mark is not a verdict
The reasoning that follows seems to me sound, whatever one thinks of the company that employs her. A watermark can suggest that a model contributed to a piece of text; it cannot say how, or whether that contribution was permitted. A student who asked a chatbot to check the grammar of a finished paragraph and a student who had the whole assignment written for them can produce text that looks the same to any detector, and a teacher who treats the two as equivalent will be wrong about one of them every time. She makes the further point that no single technology can settle whether a student cheated, which is the same conclusion I reached from the opposite direction in the earlier piece on IGCSE coursework, where detection tools turned out to give false positives against strong writers and false negatives against anyone who lightly edits a draft.
What has changed, and what the watermark debate has obscured, is that the finished product is now a much weaker piece of evidence than it used to be. A polished essay or a tidy problem set can be produced in minutes, which means the thing a teacher collects on the due date no longer tells her much about the route that produced it.
What to ask students to show
Loble's answer is to move the assessment towards the thinking, and the methods she lists are refreshingly modest: reviewing drafts as they develop, asking a student for a short oral walkthrough of how an answer was reached, and asking where they got stuck. The last of these is the most revealing, in my experience, because a student who did the work can nearly always say where it went wrong first, whereas one who did not has nothing to point to. She draws the same distinction between a correct answer reached by luck and one reached by sound reasoning, which a final-product mark cannot separate and a two-minute conversation usually can.
Replace the blanket instruction to "use AI responsibly" with a rule written for each assignment, and put it on the assignment sheet. Loble offers two examples worth adapting: one permits AI only for a grammar and spelling check once a draft is finished, and not for the thesis, outline, arguments, evidence or citations; the other permits AI proofreading of a completed history response but not for generating facts, analysis or citations. Each rule names what AI may do, what it may not, and ties the line to what that particular assignment is meant to teach. A student then has something definite to comply with, and a teacher has something definite to assess against.
Why this matters for Malaysian schools
Schools preparing for the 2027 curriculum reform will find the assessment tiers (Know, Understand, Do and, from Year 3, Apply) assume that students can show what they understand, not merely submit something that looks right, and the exam boards that most international schools here sit under already allow declared AI use in some coursework provided the student's own understanding can be demonstrated. In both cases the burden of proof sits with how the school designs the task, and a detector is a poor substitute for design. The practical advantage of evidence-of-learning methods is that they cost a teacher minutes rather than money: five minutes of conversation per student, or a draft submitted a fortnight early, needs no licence and no vendor.
Want to work through this with your assessment leads?
I run short discovery conversations with heads of department and academic leadership teams on where coursework tasks are genuinely exposed, and what a practical redesign looks like for your specific subjects, not a generic policy.
Book a Discovery CallWhat This Means for You
If your school is weighing a detection tool or reading anything into a watermark, I would suggest asking first what decision it is meant to support, because a probability that a model touched a text is not a finding that a student has done something wrong. The more dependable route, and the cheaper one, is to decide what evidence of learning you want from each assignment and to say so in advance. In short, the question worth asking of a piece of work is no longer whether AI wrote it but whether the student can explain it.