From a photograph on your desk to a report you can hand to a student, in about a minute and a half.
Six boards and fifty-two papers. You pick once, then mark as many scripts as you like against it — the mark scheme decides everything that follows, so this is the one choice worth getting right.
Before anything is uploaded, ensure all student names and other personal, identifiable details are removed — crop the top of the page, or cover the name. MarksUp also strips a name it finds at the top of a script and tells you what it removed, but the fewer personal details that go in, the better.
Photograph it, drop in a PDF, or paste typed text. A whole class set goes up in one upload, and you are shown which pages have been grouped into which script before anything is spent.
Handwriting is transcribed so it can be marked. A PDF that already contains its text is read straight from the file, which keeps the spelling exactly as written — and that matters, because on several boards the spelling carries marks of its own.
The script goes to our AI model along with the board’s assessment criteria and the marking guidance we have written for that paper. It comes back with a level, a mark, and the sentences that earned them.
The report shows what was credited and where. If you disagree with the mark, change it in one click — yours is the one that counts.
This is the part that makes the difference, so it is worth being plain about.
Every exam board marks differently. The levels are different, the assessment objectives are different, the totals are different, and two boards can use the same words to mean different things. A general-purpose AI asked to “mark this GCSE essay” has no idea which of those it is looking at.
So alongside the script we send the board’s assessment criteria for that specific paper — and, more importantly, marking guidance written by experienced teachers: how the criteria are actually applied, where the boundary between one level and the next really sits, what earns credit and what only looks as though it should. That interpretation layer is written per paper and is the part that took the longest to build.
Every criterion is checked, sentence by sentence, against the board’s published document, so what the marking is measured against is the real thing and not an approximation.
We tested the alternatives on real scripts with known marks.
Marking is not a summarising job. It needs sustained reasoning about a whole argument, held against a grid, without drifting towards the middle. Most models do not do it well: in our testing the cheaper ones marked consistently high and ranked a class in an order no teacher recognised.
MarksUp uses Claude, which came closest to a teacher’s own marks and, just as importantly, put a set of scripts in close to the same order a teacher did. Ranking a cohort correctly is what makes a mark useful; a model that is two marks generous but ranks the class right is more use than one that is accurate on average and shuffles the order.
Two documents from the one submission, because a teacher and a student need different things.
The handout prints straight from the browser. Nothing needs installing, and there is no special file to open at the other end.
MarksUp offers a second opinion. It does not replace your judgement.
Read the report before you hand it on. The mark is a starting point and the annotations are there so you can see the reasoning rather than take it on trust — every judgement points at the sentence that earned it.
If you disagree, change the mark. It takes one click, your mark is what the student sees, and both numbers are kept so you can see over a term how closely the two of you agree.
Scripts are used to produce your report and nothing else. They are not used to train anything. Remove student names and other identifying details before uploading; MarksUp will also strip a name it finds at the top of a script.
MarksUp is for practice essays and mock exams. It is not for coursework or NEA, where exam board rules on authenticating a candidate’s own work apply.
Or look at a marked script first — that costs nothing and needs no account.
Marking
This usually takes 60 to 90 seconds.
0s
Keep this tab open.
Are the names off?