AI GRADING · RESEARCH
Can AI Grade Essays? What the Research Shows, and Why We Built Our Own Model
General AI models are easily impressed by good writing. That is a problem when the grade should reward good thinking.
By Jeroen van Gessel, co-founder of Eduface · November 2026 · 10 min read
Can AI grade essays?
Yes, AI can grade essays, but most AI models you can use today should not do it on their own. A 2026 Cambridge study ran 761 university essays through the latest models from OpenAI, Anthropic and Google and found they matched the human grade band only 35% to 63% of the time.⁴ Their biggest weakness: they reward long words and polished sentences over the quality of the argument. That is the problem we set out to fix. We trained a dedicated grading model on 250,000 academic datapoints. Across several trials at multiple institutions, its grades differed on average by just 4% from the consensus grade a panel of lecturers agreed on together. That is 96% accuracy. A lecturer still approves every grade.
Free 30 min demo
See how Eduface grades your own essays
Bring one assignment and your rubric. We run it through Eduface live and compare the result with your own grade.
Book a demo
→
30 minutes, free, no obligation
My co-founder Menno and I started Eduface as students. We had both handed in essays we had worked on for weeks and got back a number and two lines of feedback. We did not doubt that our lecturers cared. They simply had a stack of 200 papers and a deadline.
So when large language models arrived, the obvious question was whether they could take some of that weight. We tested it the way most instructors do: paste an essay into a chatbot, add the rubric, ask for a grade. The results looked convincing. Confident tone, neat structure, plenty of praise.
Then we compared them with the grades lecturers had actually given, and the picture changed.
Human graders are not the gold standard we pretend they are
Before we get to AI, it helps to be honest about the benchmark.
In a UK study, 24 experienced assessors from 20 universities, all of them external examiners, graded the same student work in history, chemistry, nursing and psychology. Every piece had originally been graded as borderline between a 2:1 and a 2:2. Some came back as a first from one assessor and a third from another. Nine of the 20 assignments were ranked both best and worst by different assessors. Only one was ranked the same by all of them.¹
The authors’ conclusion was blunt: applying assessment criteria is a judgement, and there is rarely a single correct reading of what a criterion means.
That does not make human grading worthless. It means the real question about AI is not whether it beats a perfect grader. Perfect graders do not exist. The question is whether AI plus a lecturer gives students grades that are more consistent and easier to defend than what they get today.
What the research actually found
People have tried to automate essay grading for decades, and the results keep pointing in the same direction.
In 2012, a US demonstration compared nine commercial scoring engines with trained human raters on eight sets of state test essays. The machines came close to human agreement on most sets, though the lead researcher called for more validity research before anyone used them for high-stakes decisions.² Critics soon showed that some of those engines could be gamed with long, fluent essays that said very little.
Large language models changed the technology, but not the pattern. When researchers at UC Irvine asked ChatGPT to score about 1,800 secondary school essays with a rubric and no examples, it landed within one point of the human score 76% to 89% of the time on a six-point scale. Exact agreement was around 40%, compared with about 50% between two trained humans. ChatGPT also played it safe: its scores bunched between 2 and 5, while humans were far more willing to give a 1 or a 6.³
The most recent and most relevant study for higher education came out of Cambridge in May 2026. The OpRaise team tested Claude Opus 4.6, GPT-5.4 and Gemini 3 Flash on 761 undergraduate psychology essays from Cambridge, Nottingham and Manchester Metropolitan. Agreement with the human grade band ranged from 35% to 63% depending on the university, and was best in the middle of the scale, in the upper 50s and low 60s.⁴
The interesting part is why.
The problem nobody warns you about: AI gets impressed by words
All three models in the Cambridge study were, in the researchers’ words, oversensitive to linguistic features. Longer essays, a wider vocabulary and more complex sentences pushed the score up, whether or not the argument was any good.⁴ An essay a human graded 75 was scored several points lower by every AI system. An essay graded 50 was scored several points higher. The best work was pulled down and the weakest work was pulled up.
We had seen exactly the same thing in our own testing. In one comparison described in our guide to AI grading tools, two students in the Netherlands ran papers with known lecturer grades through eight tools, using the official rubrics and the Dutch 10-point scale. A case study the lecturer had graded 6.5 came back from ChatGPT as an 8.1, with warm praise for the very section the lecturer had flagged as the weakest. A law essay graded 4.4 came back from ChatGPT, Claude and Gemini as a 7.1, a 7.2 and a 6.8.
If you have graded for a few years, you know this student. The essay reads beautifully. The sentences flow. And when you look for the actual argument, there is not much there. A good grader notices. A general-purpose model, trained above all to produce and appreciate fluent text, tends not to.
This is not a small quirk. It means the students who write plainly, including many multilingual students, can be underrated for arguments that are just as strong, while students who have mastered academic-sounding prose get a bonus they did not earn.
Why we trained our own grading model
At some point we stopped trying to prompt our way around the problem. A general model asked to “act like a strict grader” is still a general model underneath. So we built a model for one job only: grading academic work.
It is trained on 250,000 academic datapoints, and it was designed around the two failures we kept running into. The first was plain inaccuracy, grades that drifted away from what an experienced lecturer would give. The second was being persuaded by the words themselves: rewarding how something is said instead of what is argued and how well it is supported.
We then tested it the way we wish every vendor would. A single lecturer is not a reliable benchmark, so we compared the model with a consensus grade: the grade a group of lecturers agreed on together for the same work, against the same rubric. Over several trials at multiple institutions, the model’s grades differed from that consensus by 4% on average. In other words, it was 96% accurate.
On that law essay from the Dutch comparison, the one the lecturer graded 4.4 and the general chatbots put around 7, our model returned a 5.5. Not perfect, and it shows why we still want a lecturer in the loop. But it read the essay for what it argued, not for how it sounded.
A few other choices sit around the model. Eduface uses separate subject models for Law, Economics, Social Sciences, STEM, Humanities and Health Sciences, because a strong law answer and a strong nursing reflection are not strong in the same way. Lecturers can tune how feedback is written. And it all runs on our own infrastructure in the EU, without third-party AI APIs such as OpenAI, under a Data Processing Agreement with every institution.
Live comparison
Want to see it next to your own grading?
Bring an essay you have already graded. We run it through Eduface in a 30 minute call and compare the result with your grade, criterion by criterion.
Book a demo
→
30 minutes, free, no obligation
Why a human still signs off on every grade
You might expect a company that sells AI grading to argue the human can step back. We think the opposite.
Even a 96% accurate model is wrong some of the time, and the essays where it is most likely to be wrong are the ones that matter most: the borderline pass, the unusual but brilliant answer, the paper that depends on something you said in class. The Cambridge researchers reached the same conclusion. They suggest using AI as a second pair of eyes, for consistency checks and to flag essays where AI and human disagree, with a human always setting the final mark.⁴
That is how Eduface works. The model reads every submission against your rubric, suggests a grade per criterion and drafts feedback, and nothing reaches a student until you approve it. You can work in two ways. In blind mode you grade first and only then see what the model suggested, which protects you from anchoring on its number. In AI-visible mode you see the suggestion straight away and edit it. Institutions can make either mode mandatory, and essays the model is unsure about are flagged for a full read instead of being quietly averaged out.
The time saving comes from the first read and the drafted feedback, not from skipping the review. Across Eduface implementations, lecturers spend on average 48.5% less time grading.
If you want the full workflow, from rubric to calibration to returning grades in your LMS, our step by step guide to grading papers with AI walks through it.
Can ChatGPT grade essays?
It can give you a number and a paragraph of feedback, and for a quick look at a draft that can be useful. For grades you record, it falls short in three ways.
It is the kind of model the Cambridge study tested, with the same habit of rewarding style. It can give the same essay a different score on a second try, and nothing records how a grade was reached, which leaves you exposed when a student appeals. And pasting student essays into a personal account sends education records to a company your institution has no agreement with. In the US that is a FERPA question, in the UK and EU a GDPR one. Ask your privacy office before you try it with real student work.
Is AI essay grading fair to students?
It can be fairer than what students get today, or less fair. The difference is in the design.
The case for AI is consistency. The ninetieth essay gets read with the same attention as the first, against the same criteria, at eleven at night as at nine in the morning. The case against is the style bias described above, which quietly favours students who already write like academics. That is exactly why a grading model has to be trained and tested against it, and why a lecturer should review the results.
Students also deserve to know what is happening. Tell them in advance that AI provides a first read, that a lecturer approves every grade, and how they can ask for a human review. In the Cambridge study, many students said they would feel cheated if AI marked their work.⁴ Being open about it is the best answer to that worry. In the EU it is also the law: AI used to evaluate learning outcomes counts as high-risk under the AI Act, which requires effective human oversight.⁵
So, can AI grade essays?
A general chatbot, on its own, cannot do it well enough for grades that count. The research is clear about that, and the reason is specific: it is too easily impressed by good writing.
A model built and trained for grading, checked against real lecturer grades and used with a lecturer who approves every result, can. It gives students faster and more consistent feedback, and it gives lecturers back roughly half the hours they now spend grading. The grade stays yours. The first read does not have to be.
Frequently asked questions
Can AI grade essays as well as a teacher?
General AI models come close on average but struggle at the top and bottom of the scale, where they score the best essays too low and the weakest too high.³ ⁴ A model trained specifically for grading does better: across several trials at multiple institutions, Eduface’s grades differed on average by 4% from the consensus grade of a panel of lecturers, which is 96% accuracy. A lecturer still approves every grade.
Why do AI graders give high marks to weak essays?
Because general language models are sensitive to how text reads. The 2026 Cambridge study found that length, vocabulary and sentence complexity pushed AI scores up regardless of the quality of the argument.⁴ A grading model has to be trained specifically to look past fluency, which is what we built ours to do.
Can ChatGPT grade essays accurately?
Not reliably enough for recorded grades. It tends to overrate fluent essays with weak arguments, can score the same essay differently on a second attempt, and keeps no record of how it reached a grade. Using it with real student work also raises FERPA or GDPR questions.
How much time does AI grading save?
Across Eduface implementations, lecturers spend on average 48.5% less time grading. Most of the saving comes from the AI’s first read and drafted feedback. Expect your first assignment to take longer while you check the results against your own grading.
Should students be told when AI helps grade their essays?
Yes. Explain that AI gives a first read, that a lecturer approves every grade, and how to request a human review. It builds trust, and in the EU the AI Act expects transparency and human oversight for AI used in assessment.⁵
Can AI essay graders detect AI-written essays?
That is a different job. Grading judges the quality of the work against a rubric, while AI detection tries to work out who or what wrote it. Our article on AI paper graders vs plagiarism checkers explains the difference.
References
1. Bloxham, S., den-Outer, B., Hudson, J., & Price, M. (2016). Let’s stop the pretence of consistent marking: Exploring the multiple limitations of assessment criteria. Assessment & Evaluation in Higher Education, 41(3), 466-481. https://doi.org/10.1080/02602938.2015.1024607
2. Shermis, M. D. (2014). State-of-the-art automated essay scoring: Competition, results, and future directions from a United States demonstration. Assessing Writing, 20, 53-76. https://doi.org/10.1016/j.asw.2013.04.001
3. Tate, T. P., Steiss, J., Bailey, D., Graham, S., Moon, Y., Ritchie, D., Tseng, W., & Warschauer, M. (2024). Can AI provide useful holistic essay scoring? Computers and Education: Artificial Intelligence, 7, 100255. https://doi.org/10.1016/j.caeai.2024.100255
4. OpRaise project. (2026). AI in university assessment: Evaluating the opportunities and risks of automated marking. ai@cam, University of Cambridge.
5. European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union.
See it on your own essays
In a 30 minute demo we grade one of your essays against your rubric, show the reasoning per criterion and walk you through blind mode. Or start free and try it yourself.