Can AI Actually Mark Handwritten Paper Exam Scripts?
Can AI mark handwritten exam scripts? Yes, but only if two things go right: the handwriting has to be read accurately, and the right mark scheme has to be applied to what was read. Most UK secondary marking is still on paper. GCSE and A-Level mocks are handwritten because the real exams are, and so are most KS3 and KS4 internal tests. A tool that expects a pasted essay or a Word document doesn't help with a pile of 30-page mock booklets.
This guide explains how GradeOrbit handles paper scripts: how it reads them, where it still makes mistakes, how a whole class set gets in and out, what it costs, and how close the marks come to a senior examiner's.
How the AI reads a handwritten script
GradeOrbit doesn't run a separate OCR step and then mark the typed-up text. The AI model reads each page image directly, and it transcribes and marks in the same pass. This matters for messy handwriting. A model reading in context can use the question and the surrounding words to make out a scrawled word, much as you do when you mark. Our article on whether AI can read messy student handwriting covers this in more detail.
Long scripts are handled in batches of pages behind the scenes. The results are then joined back into one result per student, so a 24-page mock marks as one script, not as several fragments.
You can always see what the AI read. On an individual account, each student's result has an Original / Transcript toggle. The transcript can be edited, and Regenerate Feedback re-marks the script from your corrected text. On a school account, the result viewer has a transcription panel you can show alongside the marks.
Where handwriting reading still goes wrong
It would be misleading to say handwriting recognition is solved. It isn't, and the weak spots are the ones you'd expect: faint pencil, heavy crossing-out, cramped writing squeezed into margins, and answers that continue on another page without a note saying so.
GradeOrbit doesn't give each word a "confidence score". A model's own confidence is not a reliable guide to whether it misread something. Instead, it flags questions for you to check:
- Maths answers get a second reading. When handwriting is unclear, it's easy to read the answer you expect: a pupil's 28 can look like the 27 the mark scheme wants. So in maths, a second reading of the answer is done without seeing the mark scheme. If the two readings disagree, the question is flagged as Handwriting hard to read so you know to look at the scan.
- Comments that don't match their mark are flagged in every subject. If the comment says "full marks" on a question scored 0, for example, the question is flagged Comment and mark don't match. The mark is never changed for you.
For other subjects, the safeguard is the transcript. On a script where the handwriting is hard going, compare the transcript with the scan before you accept the marks.
Getting a class set in
There are two ways to get paper scripts into GradeOrbit, and both work on individual and school accounts.
- Your phone, via a QR code. The upload step shows a QR code. Scan it with your phone and your camera links to that upload session, so each photo appears on your computer as you take it, with no transfer step. This suits marking at your desk with the pile in front of you.
- Files from a scanner or photocopier. Most school copiers can scan a stack straight to PDF. Upload the PDFs, or JPEG, PNG or WebP images.
For a whole class, turn on Bulk upload and enter how many pages each student's script has. Select every file at once and GradeOrbit splits the pages into students automatically, counting PDF pages one by one. One practical point: the split works by page count, so every script in the batch needs the same number of pages. If a pupil used a continuation sheet, upload that script on its own so their extra page doesn't shift every script after it.
Two more practical points. iPhone photos in HEIC format aren't accepted, so set the camera to Most Compatible (Settings, then Camera, then Formats). On an individual account, one student's script can run to 50 pages, which is enough for a full mock booklet. School accounts have no page limit.
Giving it the right mark scheme
The marks can only be as good as the scheme behind them. GradeOrbit marks against the criteria you give it, so it works for AQA, Edexcel, OCR, Eduqas, WJEC or your own department's scheme.
- Points-based schemes, common in maths, science and geography: the AI works out which marking points the answer meets and suggests a mark from them.
- Levels-based schemes, the norm for extended writing in English, history and the social sciences: the AI reads the answer against each level descriptor and suggests a band and mark, with its reasoning.
For a past-paper mock, upload the board's published mark scheme as it is; you can photograph it page by page if you only have it on paper. Handwritten maths mark schemes, with your answers written onto the question paper, are read directly from the scans, including circled M1, A1 and B1 marks. If you choose an exam board for a maths scheme, marking follows that board's published rules on things like misreads, more than one attempt and dependent method marks.
How close does it get to an examiner?
Exam boards publish exemplar scripts: real candidates' answers, marked by senior examiners. We removed every examiner mark and comment from the pages, ran the scripts through the same marking system teachers use with the board's official mark scheme, and compared the results with the examiners' marks. The results, measured on 18 October 2026, are on our accuracy page:
- GCSE Maths (Pearson Edexcel), 69 marks across 54 scripts: 96% within 1 mark of the examiner and 99% within 2, with an agreement score (QWK) of 0.92.
- GCSE English (Eduqas), 98 marks across 73 scripts: 83% within 1 mark and 97% within 2, with an agreement score of 0.96.
Read these with their limits. We ran the study ourselves and it has not been independently reviewed. The samples are small. Exemplar scripts are chosen to show each mark level, so they aren't a random sample of a real class. And the figures cover only these two subjects and boards. They are evidence the approach works, not a promise about every script you upload.
What a mock-length script costs
Mock scripts are long, which matters for tools that charge by the page: a 30-page booklet costs 15 times as much as a 2-page quiz. On an individual GradeOrbit account, marking costs one credit per student, however many pages the script has (up to the 50-page limit). A 30-page mock costs the same as a 2-page quiz. Plan prices are on the pricing page, and our comparison of AI marking tools sets this against how other tools charge.
School accounts work differently. Schools pay a per-teacher subscription with unlimited marking, so a department isn't charged per script at all.
Names, privacy and where the scans go
This is the biggest difference between individual and school accounts, and it decides which one suits a department's mocks.
On an individual account, scripts are anonymous. Students appear as "Student 1", "Student 2" and so on. Any names you add stay in your browser and the AI never sees them. Before anything is sent for marking, you can black out names and candidate numbers on your own device. Draw a box once and apply it to every page of every student if the name sits in the same place on each cover. Work is processed to produce the feedback and not kept afterwards.
On a school account, the school signs a Data Processing Agreement, and scripts are linked to real pupils. Class lists come from your MIS through Wonde (read-only) or from a CSV import. When you upload a class set, GradeOrbit reads the name on each cover page and matches it to the class list. A clear match links automatically; otherwise you pick the pupil. Scans are kept while you need them and removed when the marking run archives. That is 90 days by default, and the school can set anything from a week to a year. Teachers get a week's warning before scans are removed.
After marking: reviewing a class set
Every script gets a suggested mark per question, a comment against the mark scheme and a summary. You review it, change anything you disagree with, and approve it. Nothing is final until you say so. Structured questions usually need little change. Answers on the boundary between two levels take more of your judgement, as they should.
On a school account, a marked class set also gives you:
- a question-by-question analysis table with class averages, which you can export as a CSV
- sorting by surname, total mark or any single question
- a PDF report for each pupil, full or a two-page summary
- a class markbook you can export
On an individual account, each student's marked script exports as a PDF.
For the wider mock-season workflow, from planning the marking load to feeding back to students, see our guide on how to mark mock exams faster.
What makes the difference
Three things decide whether AI marking of a pile of paper scripts saves you time. The first is clear scans: flat, well-lit pages, with continuation sheets uploaded alongside their script. The second is the board's full published mark scheme rather than a summary. The third is checking the right scripts: anything flagged as hard to read, anything near a grade boundary, and a sample of the rest. Get those right and your evening goes on checking marks rather than writing the same comment 30 times.