How AI marking works

A submission passes through three stages. They are separate on purpose, because the thing that goes wrong with AI marking is letting one model do all three at once.

Step 2 is the whole design. By the time any feedback is written, the grade is already decided and cannot move to suit the prose. A model asked to both judge and justify will quietly adjust its judgement to make the justification read better.

The first call is not asked what grade your answer deserves. It is asked to describe it. How clear are the ideas, how varied are the sentences, which techniques appear and where, how many technical errors per hundred words, how closely the response engages with the question that was actually set.

It also sees what you saw. If the question depends on a source text, an insert, a poem or a photograph, the same material is put in front of the model, with the poems named and attributed rather than run together. A marker who has not read the extract cannot mark a question about the extract, and for a long time that was a real bug here rather than a hypothetical.

Those measurements are stored against your submission. That is what lets the feedback point at specific evidence instead of offering generic advice.

Assessment objectives are not weighted the same across boards, papers or question types. AQA's Language Paper 1 Question 5 splits its marks differently from Edexcel's imaginative writing task, and a Literature essay weights context differently again.

Those weightings live as data, one row per board, subject and question type, transcribed from the boards' own published specifications. When your answer is graded, the row matching your board and the paper the question came from is the one used. Where a board publishes a real level-descriptor grid, the band your response matches sets the range and the measurements position you inside it, so the mark you get comes from that band's own published mark range rather than a generic percentage.

Every grade records which level it matched, which mark band that came from, and the descriptor wording behind it. The marking engine page goes through that in more detail, including what the engine refuses to do.

A one-mark comprehension question and a forty-mark creative writing task should not go through the same process. Running a short factual answer through an essay pipeline is how tools end up awarding a grade 6 to a sentence that happens to read fluently.

That last row is deliberate. A question we cannot mark properly is reported as unmarked. Quietly recording it as nought would drag your total down for a question the system never actually read.

A mark out of the question's real tariff, and on essay-length work a 1 to 9 grade alongside it. Marks lead, because a mark out of forty is the thing a real examiner would write and a grade on a single short question is close to meaningless.

With it: what the response did well, what is holding it back, what the next band up would need, and a breakdown by assessment objective. On Literature and Language alike the feedback quotes your own sentences back at you, because “develop your analysis” helps nobody.

Is this an AI examiner?

Partly, and the distinction matters. AI reads your answer and describes what is in it, and AI writes the feedback at the end. The grade itself is not produced by AI: it is calculated in ordinary code from those descriptions and the exam board's published mark scheme. So the part people worry about, a language model inventing a number, is the part we took the language model out of.

Why does the same answer always get the same grade?

Because the step that decides the grade is arithmetic. Given the same measurements, the same rules produce the same result every time. Submit an answer twice and you will get the same grade twice, which is not true of asking a chatbot to mark something.

How long does marking take?

Usually under a minute for a full essay. Two AI calls run against your answer, and the arithmetic between them is instant.

Does it work on handwriting or photos?

Not yet. You type or paste your answer. Photo submission is on the roadmap but is not built.

Which parts of my answer does it actually look at?

Clarity of ideas, structure, vocabulary range, sentence variety, the techniques you have used, how closely you have answered the question that was set, and technical accuracy. On Literature it also counts how often you draw on context. Those measurements are stored, so the feedback and the grade are explaining the same evidence.