I've Been Marking Scripts for 18 Years. I Know When AI Wrote It.

People ask me how I know. Let me tell you exactly how I know.

AI-generated writing has a texture. This is the best word I have for it. When you have read as many student scripts as I have — thousands, across eighteen years, across every grade and ability level — you develop a sense for what a particular student's thinking feels like on paper. The way they reach for an argument, the words they favour, the places where their sentence structure breaks down and then recovers, the specific way they don't quite land the concluding paragraph. Human writing has fingerprints.

AI writing doesn't have fingerprints. It has a surface.

The specific tells

Let me be concrete, because the tells are learnable and students and parents should know them.

Stylistic uniformity. A student's voice is not consistent across a piece of writing. Real student essays have peaks and troughs: a sentence where the expression is genuinely good, followed by a clunky construction, followed by a phrase that's slightly off-register. AI-generated text is uniformly competent. The flatness is the tell. Everything is at the same temperature.

The absence of the student's cognitive fingerprint. I don't mean personal anecdotes, though their absence is also notable. I mean the specific shape of how a particular mind organises an argument. Regular students have recognisable argumentative habits — a particular way of hedging, a preference for examples over statistics or vice versa, a tendency to address the counter-argument at a specific point in the structure. These habits are traceable across submissions. AI doesn't have them.

Vocabulary slightly above the student's register. This is the most obvious tell and the most commonly noted. AI tools consistently produce vocabulary and syntax slightly above the demonstrated level of the student's in-class work and previously submitted assignments. A student who writes at a solid but unremarkable level in class suddenly submits something that reads like a competent adult. The contrast is diagnostic.

Perfect connective tissue, imperfect content. AI is excellent at transitions and structural scaffolding. "Furthermore," "it can therefore be argued that," "having established the above." The connective phrases between arguments are smooth and technically accomplished. The actual substance of the arguments, however, is often surprisingly generic — AI tends toward well-organised statements of the obvious rather than genuine analysis. The prose walks with confidence to say something underwhelming.

Hedging that contradicts the task. Academic AI tends to be cautious. It qualifies, it acknowledges "different perspectives," it avoids strong claims. For examination tasks that require the student to take and defend a clear position, AI-generated work often reads as oddly noncommittal, producing balanced essays for questions that asked for argument.

What the current guidelines actually say

The HKEAA's current AI use guidelines for DSE candidates require disclosure of AI assistance and prohibit using AI to generate assessed work. This is clear in principle. The enforcement reality is more complicated.

DSE examinations themselves — the timed, closed-book, invigilated papers — cannot be AI-assisted in any direct sense. Students are alone in the exam hall with their own knowledge and their own minds. AI does not reach that far.

The problem is everything before the exam hall. School-based assessment components, which contribute to the DSE grade in several subjects, are completed outside examination conditions. Internal assessments, oral preparation, written coursework — these are environments where AI assistance is possible and increasingly undetectable by automated means.

For the internal assessment components, the practical guidance schools have received is to design tasks that are harder to complete with AI assistance and to build in oral defence requirements. A student who used AI to draft a research report should be able to discuss the report in an oral follow-up. Many cannot.

But I want to be honest: the guidelines are running behind the technology. The advice on what constitutes disclosure is genuinely unclear in many cases. A student who used AI to help brainstorm, then wrote the draft themselves, then edited with AI suggestions — is that disclosed AI use? Undisclosed AI use? Use at all? Nobody is quite sure, and different schools are applying different standards.

The crisis beneath the cheating

The cheating question — who's using AI, how to catch them, what the penalty is — is occupying most of the conversation. I think it's the wrong place to look.

The deeper problem is students who have been using AI as a cognitive prosthetic for long enough that they can no longer function without it. Not students who made a strategic choice to cheat on one assignment, but students whose ability to produce sustained analytical writing under their own power has genuinely atrophied.

I have seen this in recent examination cohorts. Students who perform adequately in coursework components but struggle significantly when placed in timed conditions with no access to tools. The gap between assessed and examination performance is not new — nervousness, time pressure, memory lapses all contribute — but the pattern I am seeing is specific. These students are not forgetting knowledge under pressure. They are encountering a kind of blank that happens when the prompting structure is removed. They have learned to generate through prompts; without a prompt, the generation doesn't start.

If you have spent a year and a half writing essays by starting a conversation with ChatGPT instead of with a blank page, the blank page becomes genuinely difficult. The habit of initiating your own thinking — the tolerance for sitting with uncertainty, not knowing how to start, working through the disorganisation of early drafting — that habit can erode.

This matters for DSE specifically because DSE still largely rewards the thing AI is worst at generating: a student's own thinking, traceable, idiosyncratic, evidenced in the working.

What I tell students directly

Don't use AI to draft your answers. Not because examiners can't catch it — though I have just told you that many of us can. Because you need to be able to do this yourself in the exam hall in 2026 or 2027, and the only way to be able to do it then is to practise doing it now.

Use AI the way you use a teacher's explanation: as a source of perspective that you then engage with using your own mind. Ask it to explain a concept you don't understand. Ask it to give you a counter-argument to the position you're developing. Then put it down and write the essay yourself.

The student whose essay is genuinely their own — even if it's less polished, even if some of the connective tissue is clunky, even if the vocabulary is unremarkable — is the student who will perform in the exam hall. The student whose writing has been smoothed by AI has a nice-looking document and a dependency that will be exposed under timed conditions.

Eighteen years of marking has taught me one consistent thing: what comes out under pressure is what was actually learned. Everything else is surface.