Marking handwritten work with a phone

The workflow

Most student work in most of the world is on paper. If AI feedback is only for typed work, it is for the minority. So: a phone, a stack of scripts, and the model’s ability to read an image.

Photograph the page in daylight, flat, filling the frame. Paste the image with: “Transcribe this exactly as written, including errors, crossings-out and spelling mistakes. Do not correct anything. Mark anything you cannot read as [illegible].” Then give feedback on the transcription using the prompts from earlier lessons.

For thirty scripts, that is thirty photographs, which is a few minutes, and thirty transcriptions, which the free tier of most tools will handle over an evening. Students’ names go under a thumb or a strip of paper before the photograph.

Where it reads well

Clear handwriting in a well-lit photograph transcribes with few errors. Prose transcribes better than maths. English transcribes better than most other scripts, and Devanagari, Arabic, Amharic and other non-Latin scripts are read less accurately: well enough for feedback on the ideas, not well enough for a record of what was written.

The failure that matters

The model does not read; it predicts. When it transcribes, it is producing the most likely text given the image, and “most likely” is shaped by everything it has read. A student’s “recieve” comes out as “receive”. A missing word is supplied. And the check question: a wrong final answer under correct working comes out as the right answer, because in the text the model learned from, correct working is followed by correct answers.

This is a silent correction of exactly the thing you were marking. “Transcribe exactly, including errors” reduces it and does not eliminate it. So the transcription is never the record. The paper is the record. The transcription is what you give feedback on, and before any feedback goes back to a student, you check the paper for the point the feedback is about.

Maths and diagrams

Handwritten maths is where transcription is weakest: 7 and 1, x and ×, a superscript that becomes a coefficient, a fraction bar read as a minus. The consequence is that the model’s feedback on the working may be about working the student did not do.

Two things help. Ask for the transcription in a standard notation you can check at a glance (“write fractions as a/b, powers as x^2”) and read it beside the paper before asking for feedback. And for anything with diagrams, photograph the diagram separately and ask the model to describe what it sees before it interprets. If the description is wrong, stop there.

Feedback on the physical page

The point of all this is what goes back to the student. Two options. Write the model’s two points on the paper by hand, which keeps the feedback in your voice and takes thirty seconds per script. Or print a strip of feedback per student and staple it on. Either way, you have read the feedback and checked it against the paper, which is the step that makes it yours.

What this is not

It is not automated marking. The grade is not coming from the transcription, and the marking lesson said why: a mark you cannot defend to a parent is not a mark. The phone workflow is for the part where every student gets two specific points on their work within a day or two, which was previously impossible for sixty scripts, and it does that.

Privacy, again

A photograph of a script is a photograph of a child’s work, and a child’s handwriting is more identifying than typed text. The setting that stops your uploads training the model, from the first block, is the minimum. If your school has an approved tool, use it. If the rules in your country are unclear, and in most countries, for this exact case, they are, the safe reading is that a photograph with no name, on a tool with training turned off, used for feedback and deleted afterwards, is defensible, and anything more is a question for the school rather than a decision for one teacher.

The honest limit

Transcription accuracy is improving quickly and this lesson’s specific failures will shrink. The mechanism will not: a system that predicts text from an image will always be pulled toward the text that should be there. For the foreseeable future, paper is the record and the model is a fast, fallible reader of it.