Can AI Mark Your IGCSE Answers? What It Gets Right, and What It Cannot

There is a version of AI marking being sold to students this year that does not exist. It goes: write your answer, paste it in, receive a mark out of six that means something.

The reason it does not exist is not that the models are not good enough yet. It is that an IGCSE mark scheme is not a comparison against a model answer — and almost every tool that claims to mark your work is doing exactly that comparison and printing a number on top of it.

This is worth understanding properly, because the conclusion is not “AI is useless for revision”. It is more specific and more useful than that: AI is weak at the one job everyone points it at, and strong at a different job nobody markets. Getting the two the right way round is worth real marks.

What a mark scheme actually is

Students who have never read one assume a mark scheme is a model answer with a mark attached. It is not. It is a set of instructions to an examiner about what to accept, and it comes in two quite different shapes.

Points-marked questions list more creditworthy points than there are marks. A four-mark question might list seven acceptable points, with an instruction to award one mark each to a maximum of four. It will also carry an accept list and a reject list — specific wordings that do and do not earn the mark, sometimes distinguishing between two phrases a non-specialist would call identical.

Levels-of-response questions — the extended six-markers, the essays, the evaluate-and-conclude questions — do not list points at all. They describe bands of quality, and the examiner places your whole answer in a band based on the character of the argument, then fine-tunes within it. There is no checklist to tick.

Layered on top are conventions that only make sense inside a live marking process:

If you have never sat down with one, our guide to reading a mark scheme walks through a real one. It is the single highest-value hour of revision most students never do, and it is the habit that separates students who reach the top band from students who understand the content perfectly well — see how to get an A* for what that looks like across a whole subject.

Where automatic marking actually reaches

Marking reliability is not one number. It changes completely with the shape of the question, and the gradient is steep.

Notice the awkward shape of that. The questions a machine marks perfectly are the ones you were never really losing marks on. The questions where grades are actually won and lost — the six-markers, the evaluate questions, the extended data responses — are the ones where a confident-sounding score is least trustworthy. Our guide to six-mark questions in Physics shows what that band judgement looks like from the examiner’s side.

“But Cambridge is using AI marking now”

It is not, and this claim is worth killing carefully because it is repeated widely.

The June 2026 series included Cambridge’s first live digital exams — a real milestone, and we covered it in the digital exams explainer. The papers in that sitting were multiple choice. Automatic marking of multiple choice is a lookup against an answer key, the same principle as the optical scanners that have read bubble sheets for decades. Neither Cambridge’s own announcement nor the reporting on the live sitting describes AI-powered marking, and we are not going to repeat a claim its supposed source never made.

Your extended answers are read by examiners. That has not changed for the 2026 series, and no exam board has announced that it will.

What AI is genuinely good at here

Flip the task around and the picture changes completely.

Marking asks: how many marks is this worth? That question needs the scheme, the conventions, and the band descriptors, and it needs them applied consistently across thousands of scripts. It is the wrong question to hand to a model.

Diagnosis asks a different question: why did this answer not earn the mark? You supply the fact that it did not — from the real mark scheme, which is published, free, and sitting next to the paper you just did. The model explains the gap. That is a language task, and language tasks are what these systems are for.

Three uses that hold up:

1. “The scheme wanted X. I wrote Y. What is the difference?”
   This is the highest-value prompt in revision and almost nobody uses it. You are not asking for a score — you are asking why two phrasings that feel identical are not. That is exactly the gap between a B and an A.
2. Mark my working, step by step
   For maths and the calculation-heavy sciences, a photo of your handwriting analysed line by line — which step earns the method mark, where it first goes wrong, what the correct continuation from that point looks like — is far more useful than a total. It also naturally respects error carried forward, because it follows your numbers.
3. Turn the command word into a shape
   “What is an examiner expecting structurally from evaluate as opposed to describe?” The command words guide covers this properly, and it is the thing that moves six-mark answers between bands more than content does.

How this works on PapaMarks, stated plainly

We would rather describe this precisely than sell it loosely, because the loose version is the thing this whole article is arguing against.

So: automatic where automatic is genuinely reliable, self-marking where self-marking is what examiners themselves recommend, and AI aimed at explanation and step-by-step working rather than at inventing a score for a levels-of-response answer. Across the platform that currently sits at 206,639 questions from 6,433 past papers spanning 2001 to 2025, organised across 3,790 topics so you can practise the topic that is leaking marks rather than a whole paper.

The honest summary of the whole category is this: the mark scheme is free and public, and it is better than any automatic marker. What is scarce is the willingness to hold your own answer against it, and a clear explanation when the two do not match.

If you are weighing up tools rather than principles, the AI study tools comparison and the honest revision-site comparison both go through the options, ourselves included. And if the real question underneath is whether to pay for a human instead, does your child need a tutor works through the decision with the same scepticism applied here.

The loop that actually moves grades

1. Do the question under something like exam time
   Untimed practice trains a skill you never get to use. How many past papers you actually need covers the volume question.
2. Mark it against the real scheme, harshly
   Award the mark only if the scheme would. The instinct to be generous to yourself is the single most expensive habit in IGCSE revision. The past papers guide explains where to find schemes for every paper.
3. Ask why, only for the marks you lost
   This is the AI step, and it is third for a reason. You are asking about a specific gap you have already identified, not outsourcing the judgement.
4. Re-answer the same question two days later
   Not a similar one — the same one. If the correction did not stick, you learned nothing yet. Spaced repetition explains the timing.
5. Track which topics keep reappearing
   Losses cluster. Four or five topics usually account for most of the gap to the next grade, and finding them is more valuable than any single mark.

FAQ

The technology is genuinely useful. It is just useful one step later than the marketing suggests — after you already know the mark was lost, when the question becomes why. That question has a good answer, and finding it is what actually moves a grade.