Quick answer: a May 2026 Cambridge-led study found that AI models matched human degree bands only 35–65% of the time on university essays, and tended to push marks toward the middle (University of Cambridge). A separate preprint found that which examples you show a model changes its scores (arXiv preprint). The study’s own recommendation is a workable plan: use AI as a second pair of eyes, flag big disagreements for review, and keep the final mark yourself.
Contents
What the Cambridge Study Tested
Researchers tested Claude Opus 4.6, GPT-5.4 and Gemini 3 Flash on 761 undergraduate psychology essays from three UK universities, using human marks as the benchmark (University of Cambridge). The essays came from 125 students, 2022–2025, across 50 modules and 87 assignments, including coursework, open-book at-home exams and invigilated exams. The best prompt for each model was chosen on a 20% calibration subset of 153 essays.
The models matched the human degree band 35–65% of the time depending on institution: Cambridge 63%, Nottingham 53%, Manchester Metropolitan 35%. The researchers suspect the variation reflects grade range: the Cambridge essays had the narrowest range and were all invigilated exams, while the Manchester Metropolitan essays had the widest range and were all coursework.
How the Cambridge Team Built and Tested Its Prompts
It helps to know what the Cambridge team actually was, because it shapes how much weight to give it. The University of Cambridge news story (22 May 2026) describes a report from the OpRaise project, titled “AI in University Assessment: Evaluating the Opportunities and Risks of Automated Marking.”
The page does not name a journal, so it is a project report and not a peer-reviewed paper. Its named authors include Dr Deborah Talmi, who leads OpRaise, Dr Yael Benn and Dr Alexandru Marcoci. The human marks it used as the benchmark were moderated marks on a 0–100 scale from expert assessors following routine institutional processes.
The team did not rely on one prompt. It varied prompts along three dimensions: criteria specificity, calibration intervention, and scoring strategy. The prompts ranged from a basic examiner-role instruction to a full rubric with the expected mark distribution, where the model justified its evaluation before giving a mark. The best prompt for each model was chosen on a 20% calibration subset of 153 essays and then applied to the whole set.
What a teacher can borrow (my reading of those dimensions, not an instruction from the study). Paste your full rubric instead of a one-line request. Compare the AI’s marks with your own on a small sample before you trust it. And ask for the reasoning before the mark, as the rubric prompt below does by requiring a quoted sentence first. Doing the sample step at the same 20% proportion would mean calibrating on 6 essays out of a batch of 30, since 20% of 30 is 6, and using the rest for the real check.
Adding Up the Match Rates: About Four in Ten Essays
The study’s three universities contributed 133 essays (Cambridge), 172 (Nottingham) and 456 (Manchester Metropolitan), which add up to 761. Combine those counts with the reported match rates and you can work out how often the models landed on the right band overall. This is my arithmetic from rounded percentages, so treat it as approximate.
- Cambridge: 63% of 133 is about 84 essays.
- Nottingham: 53% of 172 is about 91 essays.
- Manchester Metropolitan: 35% of 456 is about 160 essays.
Those add up to about 335 matches out of 761 essays, or roughly 44%. In plain terms, about four in ten essays got the same degree band from the AI as from the human assessors, and the weight of the 456 coursework essays pulls the overall figure toward the lowest institution’s result.
The encouraging reading is that the same study found strong consistency: repeated runs on the same essays gave the same or similar marks, so a large gap between your mark and the AI’s is a stable signal that is worth a second look, even though the AI mark alone is not a verdict.
Patterns Worth Planning Around
- Central tendency. A human-awarded 75 was scored several points lower and a 50 several points higher. Human and AI marks aligned most often in the upper 50s to low 60s, and AI was least accurate for the best and worst essays and at grade boundaries such as First vs. 2:1 or pass vs. fail (University of Cambridge).
- Style over substance. All systems were oversensitive to linguistic features, giving higher marks for longer essays, wider vocabulary and more complex sentences. The authors say this can “underestimate brilliance.”
- Consistency. Repeated marking of the same essays gave the same or similar marks, and the models agreed with each other much more closely than with humans. In other words, a consistent AI score is not the same as an accurate one.
- Feedback length. AI-written feedback was three to eight times longer than the original assessors’ feedback. When word counts were matched to human feedback, focus groups struggled to tell the two apart.
Note that this was university psychology essays at three UK institutions, and the study tested specific model versions as of April 2026, so results may not hold for later models, and the study covered university psychology essays only.
The Study’s Own Recommendations
- Use AI as a “second pair of eyes” for error detection and consistency checks.
- Use AI to triage feedback for students.
- Flag large gaps between AI and human marks for further human review.
- Use AI to automate labor-intensive parts of marking, freeing staff for direct student engagement.
- Keep humans responsible for final marks.
The researchers say AI alone is too shallow and inconsistent to grade undergraduate work, and that many students said they would feel cheated if AI marked their work (University of Cambridge). That second point is a reason to tell students how any AI assistance is used.
Examples Change the Score
In a preprint testing GPT-3.5 and GPT-4 snapshots on 72 TOEFL11 essays, the choice of example essays in the prompt affected results, GPT-3.5 more than GPT-4. The authors recommend evaluating each model version separately, and note limits: a small sample, one dataset, and a preprint that is not yet peer reviewed (arXiv preprint). For a teacher, the takeaway is to keep the same anchor examples for a whole batch.
A Cautious Workflow
These steps are my suggestions, drawn from the findings above rather than tested as a package:
- Pick a few past essays you have already marked, spanning low, middle and high scores, and compare the AI’s marks with yours before using it on a real batch.
- Mark the batch yourself first, then use the AI score only to flag large gaps, as the study recommends.
- Re-read the essays at the extremes and at grade boundaries yourself, since that is where the study found the AI least accurate.
- Use the same rubric and the same anchor examples for the whole batch.
- Trim AI feedback to the few points you agree with; it can run far longer than human feedback.
- Keep the final mark with you, especially for borderline or high-stakes work.
Student Privacy: Use Tools Your School Has Approved
Common Sense Media’s summary of FERPA explains that the Department of Education has interpreted “school officials” to include edtech vendors that have a legitimate educational interest, perform a function the school would otherwise do, and are under the school’s direct control with respect to use of student records (Common Sense Media). The practical route that follows is to run student work through tools your school or district has contracted for, because the arrangement that makes them permissible is set up at the school level and not by an individual teacher’s account.
A Rubric-Based Prompt
If your school permits an AI tool, this prompt keeps the AI inside your rubric (my suggestion):
Here is my rubric:
[paste the rubric]
Here is one student essay with the name removed:
[paste the essay]
For each rubric criterion:
1. Quote the sentence from the essay that is your evidence.
2. Give a score from 1 to 4.
3. Give a one-sentence reason.
Then list every criterion where the evidence is weak.
Rules:
- Use only the rubric and the essay.
- Do not give an overall grade.
- Do not reward length or vocabulary on its own.The quote requirement makes disagreements easy to check against the text, and withholding an overall grade leaves that decision with you.
An invented example. Say a rubric has four criteria scored 1 to 4. You mark the essay 3, 3, 2 and 4, a total of 12 out of 16. The AI scores 3, 2, 2 and 2, a total of 9. Comparing criterion by criterion, the gaps are 0, 1, 0 and 2.
Using a simple rule of mine, not one from the study, you re-read any criterion where the gap is 2 or more, which here is the fourth. You read the quoted evidence for it, decide whether the AI missed something or you read too generously, and keep the final mark yourself. The rubric, scores and essay are made up.
What Students and Staff Said, and How to Reply
The study also asked staff and students what they thought, and the answers give you a plan for communicating with your class. In the focus groups, staff and students said human grading and feedback are fundamental to the “social contract” between academics and students. Many students said they would feel cheated if AI marked their work.
Staff warned that over-reliance could weaken trust, motivation, professional judgement and human engagement. On feedback, AI-written comments were three to eight times longer than the original assessors’ feedback, and when they were trimmed to a comparable word count, participants had difficulty telling them apart from human feedback; once authorship was revealed, not everyone valued the AI-generated insights.
- Say what the tool does and does not do. For example: “I mark every essay myself; an approved tool helps me spot inconsistencies.” This wording is my suggestion.
- Keep the comments yours. Trim any AI-drafted feedback to the few points you agree with, as the length finding suggests.
- Keep the human in charge of the mark. This is the study’s own closing recommendation.
Taken together, the findings point to a reassuring conclusion: students and staff accept AI that helps a teacher check consistency, and the support drops when AI replaces the teacher’s judgement. A plain, honest sentence at the top of an assignment is a small step that fits what they said they wanted.
What a U.S. Study Found With Middle and High School Essays
Because the Cambridge study used UK university essays, here is U.S. evidence at school level. A team led by Tamara Tate of the University of California, Irvine, with Steve Graham of Arizona State University among its members, had ChatGPT score about 1,800 history and English essays by middle and high school students against a six-point rubric, as reported in the Hechinger Report via KQED: AI essay grading could help overburdened teachers. ChatGPT landed within one point of the human graders 89% of the time in a batch of 943 essays, 83% in a batch of 344 English papers and 76% in a batch of 493 history essays, but matched exactly only about 40% of the time, compared with roughly half for trained human raters agreeing with each other. It clustered scores between 2 and 5 and was less likely than humans to give a 1 or a 6, which is the same “push toward the middle” pattern the Cambridge team reported. The study was presented at the 2024 American Educational Research Association meeting and was still under peer review when the article ran. The researchers’ practical advice: use AI scores only for low-stakes purposes such as a preliminary grade on a first draft, keep teachers in the loop because reading student work shows the mistakes that should guide instruction, and ask vendors for student-level exact-agreement and within-one-point rates before adopting a grader.
What the Largest U.S. English-Teacher Group Warns About
The NCTE Position Statement on Machine Scoring opposes machine scoring of student writing. Its reasons include that computers can’t judge logic, accuracy, evidence, audience appeal, humor or irony; that some systems estimate vocabulary sophistication from word length and frequency; that grammar and mechanics get overweighted, which penalizes nonnative speakers; and that students can learn to game the software. It recommends portfolio assessment scored by teams of teachers and balanced, teacher-designed assessments instead. That position predates today’s models, so treat it as a statement of values and concerns rather than a test of a current tool.
My View on AI Grading Tools for Teachers
This closing part is my opinion. I think AI grading is most useful as a consistency check on your own marks, and least useful as a replacement for your judgment on the essays that matter most.
AI Grading Tools for Teachers: Sources
- University of Cambridge: AI not yet good enough to mark university essays (May 2026)
- The Impact of Example Selection in Few-Shot Prompting on Automated Essay Scoring Using GPT Models (arXiv)
- Common Sense Media: 50 Years after FERPA’s Passage, Ed Privacy Law Needs an Update for the AI Era
- Hechinger Report via KQED: AI essay grading could help overburdened teachers
- NCTE Position Statement on Machine Scoring
For a related view on keeping people in charge of the judgment, see learning to code with AI.
Leave a Reply