Candidates Are Marking Their ACCA Answers With ChatGPT. Here's What It Can't See.

Write an answer. Paste it in. Ask how many marks it would get.

Enough candidates now do this that it has quietly become a standard part of revision for written papers — and it's not a silly idea. It's a sensible response to a real gap: you've written something, you have no way to judge it, and there's a tool sitting there that will give you an opinion in four seconds.

Why candidates started doing this

The gap it fills is genuine. On written papers you can't mark yourself — your own answer always reads well from the inside, because the logic connecting your sentences is already in your head. That's precisely the logic a marker doesn't have.

And ACCA doesn't close the gap either. My Exam Performance gives Applied Skills candidates a breakdown by syllabus area and per-question feedback. Strategic Professional — SBL, SBR, AFM, APM, AAA — returns a mark and nothing more.

So candidates improvised. One who topped a Strategic paper nationally described his method openly: write the answer, paste it in alongside the published examiner's answer, ask for a comparison. The replies under his post were full of people saying they'd started doing the same.

What follows isn't an argument against using it. It's the specific list of things it can't do, so you know which part of its answer to trust.

1. It tends to agree with you

Language models are built to be helpful, and helpfulness pulls toward agreement. The way you frame a question shapes the answer you get back — which matters enormously when the question is "is this good enough?"

A marker has no opinion about what you were hoping for. That indifference is the entire value of external marking, and it's the one thing a conversation can't reproduce.

2. It doesn't have the mark allocation

Written marks aren't awarded for quality in the abstract. They're awarded against a specific allocation for that specific requirement — a set of creditable points a marker is looking for in that scenario.

Unless you paste that allocation in, the model doesn't have it. It's assessing your text as a piece of writing: is it coherent, is it relevant, is it well argued. Those are reasonable questions. None of them is the question the exam asks.

This is why the feedback so often sounds encouraging and general. "Well structured, could be more specific" is a fair comment on a piece of prose. It isn't an answer to "which of the eight available points did I make."

3. It can't tell marks apart

Not all marks on a written paper work the same way, and the differences change what you should do.

On AFM, method marks are available for showing your workings even when the final figure is wrong — which is why abandoning a calculation you're unsure about costs more than attempting it badly. On SBL, professional skills marks are awarded separately from technical content, for things like communication and commercial acumen. On AAA, a point about a risk earns differently from a point about the audit procedure that addresses it.

Generic feedback flattens all of that into "this section could be stronger." A marker doesn't flatten it, because the allocation tells them exactly which kind of mark each part was worth.

4. It doesn't weight the requirement verb

Evaluate, discuss, explain, recommend, calculate — these aren't stylistic variations. They determine what gets credited.

An answer full of sensible recommendations scores close to nothing against a requirement that said evaluate. It isn't wrong; it's answering a different question. This is one of the most common ways strong candidates lose marks, and it's almost invisible in general feedback, because as writing the recommendations are perfectly good.

5. The model answer comparison misleads

The most popular version of this technique is pasting your answer next to the published examiner's answer and asking for a comparison. It's the sharpest use of the tool, and it has a built-in distortion.

Published answers are written without a clock and more completely than a passing script needs to be. Compared against them, almost any real answer looks thin — including answers that would comfortably pass.

The prizewinner who described this method also said he never wrote answers as long as the examiner's. He knew to discount the comparison. Someone reading his method as a recipe won't necessarily know that, and will conclude their perfectly adequate answer is failing.

What it's genuinely good for

None of the above means close the tab. Used for the right jobs it's excellent, and considerably faster than the alternatives.

The line is simple enough. Understanding the material: excellent. Judging whether your script would pass: outside what it can do.

The fair objection: what if I give it the mark scheme?

This is the reasonable pushback, and it deserves a straight answer: yes, it does better. Paste in the published marking guidance along with your answer and the feedback improves noticeably. Anyone doing that is already ahead of most candidates.

Two things still limit it.

First, published guidance exists for past questions. It doesn't exist for the practice you write on your own topics, or for a question you've adapted, and those are a large share of anyone's revision.

Second, having the criteria isn't the same as applying them at a marker's strictness. Real marking is calibrated — markers are trained and moderated against each other so that a borderline point is treated the same way across thousands of scripts. A model reading the criteria fresh has no such calibration, and it stays inclined to give you the benefit of the doubt.