Why You Shouldn't Use LLMs to Grade Student Essays (And the Ethical Way to Evaluate Written Work)

The Allure of the Automated Red Pen

Every educator knows the dread of the Sunday evening grading pile. You are staring down fifty essays on Macbeth or the socio-economic impacts of the Industrial Revolution, knowing that each one requires your full cognitive energy to evaluate fairly. In a world where Claude 3.5 Sonnet can analyse complex text in milliseconds, the temptation is obvious: why not build a prompt, feed in your marking rubric, upload the PDFs, and let the model do the heavy lifting?

It sounds like a triumph of efficiency. Some systems even promise to write custom feedback for each student, matching the teacher’s tone. But outsourcing the actual evaluation of student writing to a Large Language Model (LLM) is an ethical and pedagogical mistake. It damages student trust, misinterprets the nature of writing, and accelerates a destructive feedback loop that threatens to make human writing obsolete in education.

The Fallacy of the "Objective" AI Rubric

To understand why AI grading fails, we have to look at how LLMs read. When you feed an essay to GPT-4o, the model does not "understand" the student's argument. It calculates statistical probabilities. It looks for semantic patterns, structural markers, and vocabulary choices that correlate with the criteria in your rubric.

This leads to a phenomenon known as "surrogate grading." LLMs are notoriously biased toward a specific brand of fluent, highly structured, yet deeply generic writing. If a student writes a technically perfect, formulaic five-paragraph essay that says absolutely nothing new, an LLM will likely award it an A. If another student writes an ambitious, slightly messy, highly original essay that breaks traditional structural rules to make a brilliant point, the LLM will often penalise it for structural deviations.

By grading with LLMs, we inadvertently reward compliance over creativity. We teach students to write for the machine: predictable, middle-of-the-road, and devoid of personal voice.

The Homogenisation Loop

There is a second, even more existential issue at play: the feedback loop. Students are already using LLMs to draft their essays. If educators use LLMs to grade those essays, we enter a bizarre, post-human cycle where machines are writing text for other machines to evaluate.

When a student suspects their teacher isn't reading their work, their motivation to write genuinely vanishes. Why spend ten hours pouring your soul into an essay when you know it will be processed by an API in three seconds? The student will simply use a prompt generator to write the essay, submit it, and receive a grade generated by the exact same underlying technology. Education becomes a transaction between two instances of a cloud-hosted model, with human beings acting as mere copy-paste conduits.

Understanding how a student’s mind works—and finding the exact, messy spark that makes their intellectual engine tick—is the entire point of teaching. When we outsource grading, we outsource the very moment where mentorship occurs.

The Ethical Alternative: AI as a Pedagogy Partner

This does not mean AI has no place in the humanities classroom. Used ethically, LLMs can be incredibly powerful tools to support writing instruction, provided they are kept far away from the final grading pen. Here is how to use them without violating pedagogical trust:

1. The Rubric Stress-Tester Before you hand a rubric to your students, feed it to an LLM. Ask the model: *"What are the loopholes in this rubric? If a student wanted to get a top mark while writing a completely superficial essay, how would they do it?"* This helps you tighten your criteria, ensuring you are rewarding deep thought rather than structural mimicry.

2. The Interactive Socratic Peer Instead of having the AI evaluate the final product, have students use it during the drafting phase as an interactive sounding board. You can design prompts that instruct the LLM to act as a critical reader who asks challenging questions about the student's thesis without writing any of the text for them. This keeps the cognitive load—and the authorship—firmly on the student.

3. The Diagnostics Assistant If you have a massive cohort, you can use LLMs to perform anonymous, high-level diagnostic scans on your class's submissions. For example, you can upload a batch of essays and ask the model to identify common conceptual misunderstandings or grammatical weaknesses across the group. The prompt might be: *"What are the top three historical misconceptions present in these submissions?"* This allows you to tailor your next lecture to real student needs while ensuring you still read and grade individual essays yourself.

Protecting the Sacred Space of Feedback

Writing is an act of vulnerability. When a student submits a piece of work, they are showing you how they think. To hand that thinking over to a statistical engine for judgment is a quiet abdication of our responsibility as educators.

By keeping the human teacher at the centre of assessment, we preserve the value of the written word. We tell our students that their thoughts are worth a human being's time. Use LLMs to design better assignments, to clarify your rubrics, and to brainstorm lesson plans—but when it comes to reading and valuing a student's voice, do the work yourself.