The Brief
Society & Ethics 5 min read

39 Million Essays Later, AI Still Can't Replace a Human Marker

NAVION

Share

There is a version of university education where a student submits an essay, an algorithm scores it, and no human ever reads a single word. It is not science fiction. The tools to make it happen already exist, and some of them have been commercially available since at least 1998. The more pressing question is not whether this scenario is technically possible. It is whether universities understand what would be lost if it became normal.

A Technology Far Older Than the ChatGPT Moment

Most public discussion of AI in education treats the technology as something that arrived recently, disrupting a stable system. The history is more complicated. Educational publisher Pearson reports that its specialised trained systems have scored 39 million responses, spanning essays, short answers and equations. That is not a pilot programme. That is a mature commercial product with a long track record, developed well before large language models became household names.

What has changed is accessibility. For most of that history, AI marking was an expensive specialist tool, out of reach for many institutions. Openly available models such as ChatGPT and Claude have altered that equation. A teacher can now upload an essay alongside a grading rubric and receive a score and feedback within seconds, at no significant cost. The barrier to adoption has dropped sharply, even if the underlying logic of the technology has not changed much.

That logic is worth understanding clearly. These models do not evaluate work the way a human does. They predict what mark is likely, based on textual patterns compared against large bodies of existing text. If the model has been trained on material similar to the task being assessed, its predictions tend to be closer to human judgements. If not, the gap widens. This is a fundamental distinction: prediction is not comprehension.

Where the Research Points

A study conducted by researchers at the University of Melbourne and Cardiff University examined whether later versions of ChatGPT could reliably replicate human markers’ scores on 50 undergraduate bioscience essays. The findings were instructive. There was some alignment at the level of overall essay marks, but the researchers concluded that ChatGPT could not be relied upon to assign and explain marks accurately. Performance was particularly weak on specific marking criteria, such as the appropriateness of content and the quality of supporting evidence. The model also tended, on average, to grade essays higher than human markers did.

This upward drift is not a bug in the conventional sense. It reflects how the technology works. A system that predicts rather than evaluates will tend toward outputs that look plausible and acceptable, which in grading terms often means generous. That has real consequences for academic standards and for the meaning of a grade.

Australian universities are navigating this carefully, though not always transparently. Deakin University prohibits teachers from uploading student assignments into openly available AI tools, citing privacy concerns related to offshore data storage and Australian privacy regulations. The University of Melbourne permits teachers to use approved AI tools to support their evaluation of student work, but explicitly prohibits relying on AI alone to determine marks or grades. The pattern across institutions is consistent: human oversight is required, and responsibility for judgements remains with the person, not the system.

What Gets Lost When the Human Steps Back

Here is what most coverage of AI marking misses. The debate tends to focus on accuracy: can the algorithm match the human score? That is the wrong frame. Assessment is not only a measurement exercise. It is a relationship.

Research into student and teacher perspectives on feedback suggests that AI can create distance between the two. Students who once might have approached a tutor with a question now turn to AI first. That shift is subtle but significant. Feedback from a human teacher carries a signal that a score from an algorithm cannot: it tells the student that their learning matters to another person.

There is also a feedback loop that runs in the other direction, from student work back to the teacher. If a large proportion of students in a cohort all misunderstand the same concept, a teacher who marks their essays will notice. That pattern becomes an opportunity to improve how the subject is taught. An algorithm that processes the same essays will score them, but it will not flag a systemic teaching problem.

The University of New South Wales has piloted a different approach in some subjects, removing numerical marks entirely and replacing them with competency assessments against defined outcomes. The university describes this as shifting focus toward student learning. The pilot was not designed as a response to AI, but it points toward a broader question: if the goal of assessment is learning rather than ranking, what role should any marking system, human or automated, actually play?

In Short

AI marking is not new, and it is not going away. What is new is how easy it has become to use. The evidence available suggests current AI tools cannot reliably replicate human judgement on complex written work, particularly at the level of specific criteria. Universities are responding with policies that keep humans accountable for final grades. The deeper issue is that reducing assessment to a scoring problem misses what assessment is actually for: understanding whether learning has happened, and strengthening the relationship between teacher and student that makes learning possible. Technology can support that process. It cannot substitute for it.

Based on reporting from The Conversation.

Read the original study.

Written by

NAVION