Why auto-grading fails, and co-grading works
Fully automated grading keeps failing in higher education for reasons that have little to do with model accuracy. Here is what actually breaks, and the design that avoids it.
Every few years, someone announces that grading is solved. A model scores highly against human markers on a benchmark, a press release follows, and a pilot starts somewhere. Two semesters later the pilot is quietly over.
This has happened often enough that it is worth asking why — because the answer is not “the models are not good enough yet.” The models have been good enough for some tasks for a while. What keeps failing is the design.
The three failures
1. The accountability gap
When a mark is wrong, someone has to answer for it. A student appeals, a moderation committee reviews, an external examiner asks a question. In every one of those conversations, the institution needs a person who can say: I assessed this, here is my reasoning.
A fully automated system has nobody in that chair. The department head ends up defending a decision they did not make and cannot fully reconstruct. Faced with that, most institutions correctly decide the risk is not worth the saved hours.
2. The teaching moment disappears
The point at which a student learns most is not when they see their mark. It is when they understand why the answer was wrong — which of their assumptions broke, and where.
If a machine handles that conversation end to end, the teacher is absent from the single highest-value interaction in the course. You have optimised away the thing the course exists for.
3. Rubrics are not as complete as we pretend
Any experienced examiner knows that a rubric is a compression of a judgement, not a substitute for it. Real scripts produce cases the rubric does not cover: a correct answer by an unexpected route, a method error that happens to yield the right number, a student who misread the question but reasoned impeccably from their misreading.
A human resolves these in seconds using context the rubric never encoded. An automated system either guesses or applies the rubric literally, and literal application of an incomplete rubric is exactly how a good student loses marks unfairly.
What co-grading changes
Co-grading keeps the machine and moves the boundary. The system does the part that is genuinely mechanical — reading every script, locating the answer to each question, measuring it against the stated criteria, drafting the reasoning — and then it stops.
Everything it produces is a proposal. A faculty member reviews it, changes whatever they disagree with, and finalizes. Nothing reaches a student before that happens.
This sounds like a small distinction. It changes four things:
Accountability stays intact. There is always a named person behind every published mark. Appeals, moderation and audit all work exactly as they did before.
The rubric gap becomes visible instead of silent. When the system is uncertain or the script does not fit the criteria cleanly, it says so and the faculty member decides. The edge cases surface rather than being silently resolved wrong.
Consistency improves where consistency is actually the problem. The measurable unfairness in manual grading is drift — the thirtieth script is not marked like the third, and the script marked at 11pm is not marked like the one at 10am. A machine does not get tired. Holding the same standard across the whole batch is the thing it is genuinely better at.
The saved time goes somewhere useful. Faculty stop spending their evenings locating answers and adding up marks, and start spending that time on the scripts that need judgement — and on the students who need a conversation.
What this looks like in practice
A realistic co-grading loop:
- Scripts arrive as a batch — scanned handwritten answer sheets, typed submissions, or both.
- Each is measured against the rubric the department already uses. Not a new rubric; the existing one.
- The faculty member opens a workspace where each answer is already assessed, with the reasoning written out.
- They agree, adjust, or rewrite. Most answers take seconds. The hard ones take as long as they always did.
- They review the complete set, then finalize.
- Only now does anything become visible to a student — and it arrives with written feedback, not just a number.
The time saving is real but it is not the headline. The headline is that every student in a cohort of three hundred now receives written feedback, which was previously impossible for reasons of arithmetic rather than will.
The honest limitations
Co-grading does not remove the need for subject expertise — it concentrates it. A faculty member reviewing two hundred proposed assessments still needs to be reading critically, and a system that makes review feel frictionless enough to rubber-stamp has failed differently but just as badly.
It also does not fix a bad rubric. If your criteria are vague, the proposals will be vague, and you will spend your time rewriting rather than reviewing. The institutions that get the most out of this are the ones that treat it as a reason to make their rubrics explicit — which turns out to be worth doing regardless.
The test to apply
When you evaluate any assessment tool, ask one question: can a grade reach a student without a human approving it?
If the answer is yes — even as an optional setting, even as a convenience for large batches — you are looking at auto-grading with extra steps, and it will run into the same three walls.
If the answer is no, and that is architectural rather than a checkbox, you are looking at something that can actually survive contact with an examination cell.
GunanQ is built on exactly this constraint: nothing publishes without a faculty signature. See how it works, or book a session where we grade one of your own papers in front of you.
AI-Shala Team
Research & Engineering
Written collectively by the people who build and teach here — engineers, researchers and mentors who spend their week with the problems these posts describe.