Marking SEND Student Work With AI: A Teacher's Guide
Marking SEND student work with AI is one of the places teachers most often pull back from automation — and for good reason. Access arrangements, scribed responses, word processors, extra time, rest breaks, and the simple reality that some students' handwriting is harder to read all sit awkwardly against an automated marker that does not know any of the human context behind the page. Get this wrong and an AI marker quietly punishes students for things the rubric was never meant to penalise.
Done carefully, AI marking can work fairly across SEND papers and save real time. This guide is the working teacher's version: what GradeOrbit sees, what it does not, how to set up the marking criteria so SpaG penalties do not bite the wrong students, and how to handle the physical-paper case where a scribe's handwriting and the student's own writing both appear on the same page.
What AI Marking Sees vs What a SEND Student Produced
An AI marker only sees what is on the page. It does not know there was a scribe in the room, that the student had 25% extra time, that two rest breaks happened, or that the handwriting in the second half got messier because the student was tired. All of that context lives in your head and in the access arrangement paperwork — not in the photo or PDF you upload.
This matters because rubric design assumes a level playing field. If your mark scheme penalises spelling, punctuation, and grammar (SpaG) and you upload a piece of work that was scribed, the marker will judge a scribe's spelling against the student's mark. That is the kind of quiet unfairness that erodes trust in the tool fast. The fix is not to abandon AI marking on SEND work — it is to set the marking criteria up properly before any work goes in.
Marking Scribed and Word-Processed Responses
For scribed work, the rule is simple: turn off SpaG-style criteria before running the analysis. In GradeOrbit, the marking criteria you define drive what the AI grades against. If your criteria do not list spelling, punctuation, or grammar, the AI will not deduct for them — meaning a scribed response is judged on content, structure, and argument, exactly as the exam board would expect under that access arrangement.
Word-processed work is more forgiving because the spelling is the student's own (or autocorrect's, depending on the exam regulations you are working under). Mark it the same way you would mark any typed coursework, with the criteria you would use for the cohort. The bigger watch-out here is formatting — students with assistive technology sometimes produce documents with unusual paragraph spacing or font sizes that look "off" to a human reader but are irrelevant to the rubric. The AI will not penalise that either, as long as it is not in your criteria.
For the broader pattern of marking criteria for marks-based grading, see using grading rubrics to speed up KS3 marking.
Handwriting, Spelling, and the Rubric
Untidy handwriting is the most common SEND marking concern teachers raise. Two separate questions sit inside it: can the AI read the writing, and does the AI penalise it?
On reading: GradeOrbit uses Google Cloud Vision for transcription, which handles a wide range of handwriting quality. Where a passage is genuinely illegible to a human, it will be illegible to the AI too. You can review the transcription before grading — if a paragraph came through as garbled, that is your signal to either re-photograph at a better angle or transcribe that section yourself.
On penalising: handwriting quality is not in the rubric for almost any UK exam board mark scheme (AQA, Edexcel, OCR, WJEC). If your marking criteria do not mention handwriting, the AI will not score it. SpaG marks, where they exist, are about written accuracy of language — not the visual neatness of the script. Set up your criteria to mirror the real mark scheme and the issue mostly disappears.
Extra Time, Rest Breaks, and Comparing Across the Class
When you batch-mark a class, the AI does not know which students had extra time and which did not — and it should not. Every paper is graded on the same rubric, which is the correct behaviour. The fairness work happens in two places: the rubric design (which you have already done by removing penalties for things access arrangements compensate for) and the human review of edge cases (which you do after the AI has graded).
The right workflow for a SEND-heavy class: run the whole batch through AI marking, then read every flagged or borderline grade yourself, with the access arrangement paperwork next to you. The AI gets you 80% of the way there in 20% of the time; your professional judgement covers the last mile where the human context matters.
Scanning Physical Papers With Untidy Handwriting
Physical paper is where most SEND marking happens — end-of-year exams, mock papers, in-class assessments. GradeOrbit's mobile camera upload (via QR code) lets you scan handwritten papers straight from your phone, page by page, without leaving the marking room. The trick with messier handwriting is the photo, not the AI:
- Use indirect natural light, not a desk lamp pointed at the page (kills shadows).
- Photograph one page at a time, square to the paper, with the whole page in frame.
- For pencil work, increase phone screen brightness before previewing to spot any pages that look too faint.
- Redact student names with the in-app black box tool before processing — the AI never needs them and you keep the privacy contract intact.
For the full physical-paper workflow, see how to use AI marking for handwritten student work.
Try GradeOrbit on Your Next SEND Marking Pile
SEND marking should not be the pile teachers leave until last because they dread the time it takes. With marking criteria set up to match the access arrangements your students actually have, GradeOrbit can take the bulk transcription and first-pass grading off your hands and leave you free to spend the real time where it matters — on the borderline grades and the feedback conversations.
Visit the GradeOrbit homepage to set up your first marking criteria and run a batch.