When AI Grades Student Work, Who Checks the Feedback?

An elementary school teacher in Lowell, Massachusetts, spent part of her summer reading her students' essays and found scores that did not add up. District leaders alerted the state, and the state eventually confirmed that about 1,400 of roughly 750,000 essays in its MCAS tests had been scored incorrectly in a process that used AI to help mark them. Some essays that deserved a six out of seven reportedly received a zero. The process already included a human second read of 10 per cent of AI-scored essays. The alarm still came from a teacher with a pile of her own students' work. It was a school test, not a university course. The principle travels.

Universities are weighing the same promise: marks in minutes, not weeks. For a part-time student who studies after a shift, quick feedback is worth a great deal, and so is the cost of getting it wrong. A wrong mark delivered fast is still wrong, and the student acts on it sooner. My position is simple. A university may use AI to help assess student work. It may never hand over responsibility for the result. The institution and the faculty who award the degree own every mark, including the ones a machine suggested.

The machine flatters weak work

A study in Assessment & Evaluation in Higher Education, reported by Times Higher Education and Inside Higher Ed, gave 50 undergraduate bioscience essays to two versions of a widely used chatbot under four different prompts. In all but one case the AI's average marks ran higher than the human markers'. In one case an AI mark sat 40 points from the human mark on a 100-point scale. Weak essays were marked up and strong ones marked down. The student who most needed to hear that the argument was thin was told it was fine.

Matching on the average is the easiest trick a grader can pull. A 2024 study of 60 master's-level essays found that GPT-4 landed on the human mean while bunching its marks in the middle range and agreeing poorly with individual markers, and rewording the instructions barely changed that. Validate against class averages and the system passes. Look at individual students and it does not.

Draw the line first

Decide first what AI may touch. Comments on a first draft that a student can bin are one thing. A mark on work that counts towards a degree is another. Europe has drawn that line already: the EU AI Act classes AI systems that evaluate learning outcomes as high-risk, with human oversight among the requirements, and those obligations now apply from December 2027. A university outside Europe has no legal duty to follow. There is no good reason to wait.

Then test the tool on your own students' work. A vendor's accuracy figure describes the vendor's test data. A recent preprint comparing several models found that agreement with human markers varied widely between institutions and that no model or prompt did best everywhere. Test again whenever the tool or its instructions change.

Review that can fail

A human in the loop is no safeguard if the human cannot say no. A lecturer told to approve three hundred AI-written comments by Friday will approve them. So build a review that can fail. Have faculty mark a random sample blind, without seeing the AI's output, and compare the two by grade band, cohort and student group, part-time and second-language students included, so that a drift towards the middle, or against one group, shows up in the numbers. Every failure, every borderline result and every mark that affects progression gets read by a named person.

A student's right to ask

Students who study online often have no corridor in which to ask a lecturer what a comment meant. The written feedback is the conversation. Tell students when AI contributed to a mark or comment, show them the criteria it applied, and give them a route to a human re-mark, with a published deadline and no penalty for asking. Expect appeals to find mistakes. When one does, fix the record and then go looking for the same fault in other students' work, as Massachusetts did by rescoring every essay it found affected. Keep a log of the tool version, the instructions and the output behind each mark. A grade nobody can reconstruct cannot be defended, and a university that cannot explain a grade to a student has not graded the work.

What goes on the page

A mark is a judgement about a person, and a judgement needs someone answerable for it. A responsible framework fits on one page. It says which assessments AI may touch and which it may not. It demands validation before launch and after every change, names one academic who answers for accuracy, and puts a human signature on every consequential decision. Students are told, and students can appeal. One more line belongs there: how many hours of faculty time the university has set aside for checking. If the answer is none, the saving on marking has been taken from students.

AI can take weight off marking. It cannot sign the transcript. Every student should be able to ask who checked their feedback and get an answer with a name in it.