AI marking for UK teachers

Dolly marks work it has set, so every answer is marked against a mark scheme written for that exact question. A class goes through five steps:

1. You choose a class and the topics you are teaching from its specification. Dolly writes an extended-response question for that subject with a levels-of-response or points-based mark scheme at 4, 6 or 8 marks, or a quiz, and you can edit any of it before setting it.
2. Students sign in to their own Dolly account and answer there.
3. Dolly marks every answer against the assignment’s mark scheme, and checks each mark against the mark scheme’s structure before saving it.
4. You review the marks, using each mark’s confidence score to decide which to check first, and change any mark or comment you disagree with.
5. You publish, and the whole class gets its marks and feedback at once. Nothing reaches students before that.

Teachers in England can use AI to help mark classwork, homework and practice, as long as a teacher checks the result, pupils’ data is protected and the school has permission to use pupils’ work this way. The Department for Education’s policy paper on generative AI in education (updated 12 August 2025) asks schools to check AI output before relying on it, keep personal data out of AI tools where they can, and not let pupils’ work train models without permission. Its support materials (updated 19 May 2026) add that AI marking can give inconsistent feedback, that AI should not make summative judgements about attainment, and that permission is needed from the rights holder, the pupil or a parent for under-18s, before AI marks or gives feedback on a pupil’s work.

Work that counts towards a qualification is different. Ofqual said in January 2026 that using AI as the sole means of awarding marks does not comply with its rules, and JCQ’s guidance, in force from September 2026, says an AI tool cannot be the sole marker: the teacher must review all the work and stays responsible for the mark.

AI marking accuracy depends on the task. In a 2024 study of short answers, AI agreed with human markers about as often as they agreed with each other, and one company’s 2026 GCSE mock data found the same in maths and science. Studies of university essays found AI less reliable, pulling marks towards the middle. The full summary of the evidence sets out what each study measured and who ran it.

Human marking varies too, especially in English, so the fair test for any tool is whether it agrees with you as often as a colleague would. Dolly has not published an agreement study of its own. What it gives you instead is a confidence score on every mark, so you know which answers to check first, and the final say before anything is published.

Homework and regular practice are where AI marking fits best. Students get specific feedback as soon as you publish, and you see who handed in and what the class got wrong before you plan the next lesson. Dolly also builds personalised flashcards from the topics each student found hardest, for one student or the whole class, and writes reports for a whole class from their marked work for you to edit before publishing.

Dolly writes extended-response questions and mark schemes specific to maths, English language, English literature, biology, chemistry, physics, combined science, computer science, the humanities and modern foreign languages, and covers other subjects too. Every answer is marked against its own question’s mark scheme, and maths answers get method marks for correct working even when the final answer is wrong. Quizzes work in any subject.

Questions are written to the specification you choose from Dolly’s library, which is organised by subject, exam board and level, or to any specification or scheme of work you upload as a PDF.

Dolly stores your data on AWS in London and complies with UK GDPR. Writing and marking run on Anthropic’s Claude through Amazon Bedrock in the EU, and Bedrock does not use your content to train models or share it with model providers.

Students answer in their own accounts, in classes their teacher sets up, and schools can bring classes in from their MIS through Wonde. Students’ work is never used to train models.

Whichever AI marking tool you look at, these questions show whether it fits the guidance above:

• Does it mark against your mark scheme, or a generic rubric?
• Can you see and change every mark before students do?
• Does it show how sure it is about each mark?
• Where is pupils’ data stored and processed, and is it used to train models?
• Does your school have permission to use pupils’ work with it, as the DfE’s guidance asks?
• Is it for practice and classwork, or does it claim to mark exams or non-exam assessment?

For a side-by-side look at a general assistant and a marking tool, see Dolly vs ChatGPT for marking.