Sunday evening, a stack of essays, and a lesson to plan for 8 a.m. — every teacher knows the tableau. It is no wonder that AI marking tools are spreading so quickly. The workload case is genuinely compelling. The Department for Education's Teacher Workload Survey (2019) found teachers in England working around 49.5 hours in a typical term-time week, with roughly one in four exceeding 60 hours, and marking consistently among the heaviest burdens outside the classroom itself. If a machine can hand back several hours a week to an exhausted profession, only a purist would refuse to look at it.
So let me say the quiet part first: automating elements of marking is not, in itself, unethical. The ethics depend entirely on how it is done.
The case for caution
Start with what we know about humans and automation. In aviation and medicine, researchers have documented "automation bias" for decades: operators of reliable automated systems gradually stop checking them, accepting the machine's output even when it is wrong — errors of omission (missing the fault the machine missed) and errors of commission (following the machine into a mistake). A systematic review by Goddard, Roudsari and Wyatt (2012) found the effect across study after study, and — the uncomfortable finding — expertise is no reliable protection against it. There is no reason to think teachers with LLMs are exempt. The teacher who lets AI mark a hundred essays will not review a hundred essays. They will skim ten, find them plausible, and stop. That is not a character flaw; it is how human attention works around competent machines.
The second problem is subtler. Marking is not only an output; it is how a teacher maintains their grip on the material. When I mark a set of essays, I learn what my class actually understood, where the misconceptions cluster, which of my explanations failed. Outsource that reading and you outsource your own situational awareness. Teachers who hand their thinking to LLMs for long enough lose their mental grip on the content they teach — and with it, eventually, their intellectual authority in front of the class. Students can tell, sooner than you would like, whether the person at the front is on top of the material.
The third problem is the student on the receiving end. Feedback is one of the most powerful levers in education — Hattie and Timperley (2007) placed it among the strongest influences on achievement, and the EEF's Teaching and Learning Toolkit estimates that effective feedback buys around six additional months of progress. But the same research is clear that feedback works when it is specific, timely, and trusted — and trust is relational. A generic paragraph from a model that has never met the child is not the same intervention as three precise sentences from a teacher who watched that child wrestle with paragraph two on Thursday. Strip the human out of feedback and school becomes, in the strict sense, less useful; the teacher-student bond, which carries much of the motivational weight of the whole enterprise, frays.
A workable ethic
My own position, then, is neither refusal nor surrender. It runs something like this. AI may draft; the teacher must decide. Use it to produce a first pass on routine, low-stakes work — then review every comment before it reaches a student, rewrite what is wrong, and personalise what is generic. Keep the high-stakes marking — exams, coursework, anything that shapes a child's future — firmly in human hands, with AI at most as a second reader. And never let the tool mark work you have not read yourself, because reading the work is where your own expertise is maintained.
In other words, the machine's output should arrive at the student's desk packaged in a human way: checked, edited, and owned by a teacher who could defend every word of it. That is not an arbitrary scruple. It is the only arrangement in which the classroom remains a connected place — where the teacher knows the students' minds, the students trust the feedback, and the saved hours go back into planning and people rather than into simply producing more throughput.
Is it unethical to automate marking with AI? No. It is unethical to abdicate it. The distinction is one sentence long, and I suspect the profession will spend the next decade learning to hold it.