Can AI Cut Grading Time?

The problem, in numbers

In our reading, marking is the part of teaching that grows with every student enrolled. In AITSL's 2024 workforce survey of full-time classroom teachers, 23% of secondary teachers and 21% of teachers in combined schools reported spending ten or more hours a week on marking and assessment of student work, against 11% in primary schools. Another 36% of secondary teachers reported five to nine hours.^{Source 1} The same survey put median term-time working hours at 50 a week.^{Source 1}

Teachers are already reaching for AI. In the TALIS 2024 survey of 6,040 Australian teachers, as reported by ACER, 66% of lower secondary and 47% of primary teachers said they had used AI in the previous year, with summarising topics and generating lesson plans the most common uses.^{Source 2} In our reading, marking is an obvious next target. The UK government has funded it directly, setting aside £1 million in January 2025 for 16 developers to build AI marking and feedback tools.^{Source 6}

What improves, and why

Two different jobs travel under the name AI marking. One is scoring: putting a grade or band on a piece of work. The other is feedback: telling a student what to fix next. In our assessment, rubric-based work suits both, because the criteria are written down and a model can be given them. A short-answer quiz or a structured paragraph against a clear rubric is close to pattern matching. An extended argument, a creative piece or a practical portfolio asks for judgement about what the student was trying to do.

Speed is the obvious gain. In one peer-reviewed study, experienced evaluators needed about 20 to 25 minutes of total time per essay to write feedback, including their training.^{Source 4} A model produces comments of slightly lower quality as fast as it is asked. That does not translate directly into teacher hours saved, because someone still has to read what the model wrote before a student sees it, and in our reading that checking time has not yet been measured well.

What the evidence actually says

The scoring evidence is mixed. Tate and colleagues had ChatGPT score US secondary essays from three corpora on a 1 to 6 scale, with no example papers to learn from. Two human raters agreed with each other at a weighted kappa of .79; ChatGPT agreed with humans at .52, .23 and .52 across the three samples, which the authors call moderate to fair.^{Source 3} Average scores did not differ significantly, but the model gave fewer top and bottom scores. The authors would not recommend it for high-stakes summative marking or as a replacement for teachers, but see it as useful in the early stages of drafting.^{Source 3}

A larger test was harsher. Researchers at ETS, a US testing organisation, had GPT-4o score 13,121 essays by students in grades 8 to 12. Its score matched the human score about 30% of the time and was within one point 77% of the time, averaging 2.79 against 3.69 for humans. The authors found the scores came nowhere close to industry standards for automated scoring.^{Source 5}

Feedback holds up better than scores. Steiss and colleagues compared 200 pieces of feedback from trained human evaluators with 200 from ChatGPT on the same history essays. Humans were slightly better in every category except criteria-based feedback, and every average difference was within one point on a five-point scale.^{Source 4} The authors concluded that well-trained educators give better feedback, and that AI may be useful for early drafts or when such an educator is not available.^{Source 4}

On time saved, the evidence is thin. The UK figure of up to 50% less time on formative assessment, for some tools, is the developers' own estimate, and the government's claim that curriculum content lifts feedback accuracy to 92% from 67% comes in a press release that does not say how it was tested.^{Source 6} We found no controlled trial that measures how many marking hours a teacher gets back once checking is counted.

Where it falls short

• Uneven by group. In the ETS study, GPT-4o marked Asian/Pacific Islander students down by an adjusted 1.16 points against 0.87 for White students, and when asked to guess a writer's race or ethnicity from the essay it agreed with the recorded answer 49% of the time, against 30% exact agreement with human scores.^{Source 5}

• English learners. In the Tate study, AI agreement with humans was lower for English learners (.40 and .43) than for their peers (.51 and .52), although the regression found no significant interaction.^{Source 3}

• Squeezed to the middle. The model gave fewer of the highest and lowest scores,^{Source 3} which in our assessment risks misjudging the students a teacher most needs to notice.

• Moving targets. Each study tested particular model versions. A tool that scored well last term can change after an update, so one validation is not enough.

Where governance decides the outcome

Australia's schools framework already sets the terms. Teachers and school leaders retain control of decisions and remain accountable for those supported by generative AI; tools are tested before use; and people affected by AI outputs are informed and able to question them. Assessment should allow a clear and unbiased evaluation of student ability, and use should counter unfair discrimination.^{Source 7} A mark that feeds a report or a progression decision is therefore the teacher's mark, whatever drafted it.

Student work is also data. The framework calls for privacy disclosure and care with what is entered into AI tools.^{Source 7} England's guidance goes further: schools must not let students' original work be used to train generative AI models without permission, and the professional who produces the final work remains responsible for it, whatever tools were used.^{Source 8} Before choosing a marking tool, ask the vendor in writing whether submitted work is retained or used for training.

How to think about adoption

Start where the stakes are low and the criteria are explicit: first-draft feedback on structured writing, short-answer quizzes, practice responses. Let AI write the comment and the teacher approve or edit it before a student sees it. Keep summative marks, and anything that reaches a report, with the teacher. Measure the gain honestly: log marking time for a term with and without the tool, counting the time spent checking. Each term, have teachers mark a sample independently and compare it with the AI's marks, broken down by student group, including students learning English. If the gap is wide or uneven, the tool is not ready for that task.