AI Still Cannot Grade University Essays: What the 2026 Research Reveals

The Pressure to Let AI Mark the Papers

After the CBSE On-Screen Marking controversy of 2026, a predictable narrative emerged in Indian education policy circles: if human-led digital evaluation can produce errors at scale, could AI simply replace the human evaluator altogether? The question is not hypothetical. EdTech vendors have been marketing AI essay-scoring tools to universities. A handful of institutions have run pilots. And under pressure to process millions of answer scripts quickly, several university controllers of examinations have asked whether AI scoring could be the logical next step after scanning.

The 2026 research literature now provides a clear answer to that question — and the answer is no. Not yet.

What Cambridge Found

In May 2026, researchers from the University of Cambridge published a study testing three frontier AI systems — including the latest available versions of Claude and ChatGPT as of April 2026 — against a dataset of over 750 undergraduate essays submitted as part of a psychology degree programme across three UK universities.

The headline finding: AI matched the broad degree classification (a first, 2:1, 2:2, and so on) awarded by human examiners between 35 and 65 percent of the time, depending on the model.

That range sounds wide because it is. The AI systems performed better in the middle of the grade distribution and consistently failed at both extremes. Work that human examiners judged to be excellent — the kind of essay that demonstrated original synthesis, deep conceptual understanding, and careful academic reasoning — was systematically undervalued by the AI models. Work that human examiners judged to be poor was routinely overvalued.

The Cambridge team described the core problem as the AI "leaning heavily on surface linguistic features." Sentence complexity. Vocabulary range. Essay length. The AI systems gave higher marks to essays that sounded fluent and well-structured, regardless of whether the academic content was actually sound.

"Homogenized, underestimates brilliance, and favors linguistic style over the substance of sound academic judgment," was how the lead researchers summarized the systematic bias they observed.

What the Cardiff-Melbourne Study Added

Three months later, in August 2026, a follow-on study from Cardiff University and the University of Melbourne extended the scope and reached similar conclusions.

This study focused specifically on mark distribution effects, using generative AI models to grade several hundred undergraduate essays and comparing the resulting distributions against human grades. The findings were striking in their consistency: AI models systematically compressed the mark range. High-scoring essays received lower marks than human examiners gave them. Low-scoring essays received higher marks than human examiners gave them.

The magnitude of the effect was substantial. For some essays, the difference between the AI-assigned mark and the human-assigned mark reached 40 points out of 100. That is not a minor calibration error. That is a disagreement about whether an essay deserves a pass or a distinction.

The study also found that AI models were oversensitive to specific surface-level text properties — particularly essay length. Longer essays, all else being equal, received higher marks. Shorter essays, even when they were more precisely argued, received lower marks. The implication is that a student who writes a well-crafted 600-word argument would be penalized relative to a student who writes a loosely organized 1,200-word response containing many of the same ideas less precisely.

Why These Findings Matter for India

India's university examination system evaluates approximately 15 million answer scripts per semester cycle at the undergraduate level, across languages including Hindi, Tamil, Telugu, Kannada, Bengali, Marathi, Odia, and dozens of others. The scale is an argument for automation. But the research findings identify exactly the failure modes that would be most damaging at scale.

The Language Problem

The Cambridge and Cardiff studies tested AI systems on English-language university essays. Multilingual university examination systems add a further layer of complexity that current AI models handle poorly. A student writing in Telugu or Kannada about thermodynamics is drawing on a different linguistic register than the training data these models have been optimized on. The style-over-substance bias identified in English-language essays would likely be larger in regional language contexts, where the AI's baseline calibration is less reliable.

The Consistency Problem at Exam Scale

In the UK pilot studies, each essay was evaluated by a small number of human raters and a set of AI systems. In an Indian affiliating university examination, a subject paper at a large university like Osmania, Mangalore, or Calicut may involve 50,000 to 100,000 scripts. At that scale, systematic mark compression becomes a structural problem. If AI consistently deflates the top 10 percent of papers by 15 marks, those students face real consequences: degree classification changes, admissions rejections, job application outcomes. The harms are not distributed randomly. They fall on students who did the work.

The Grievance and Revaluation Problem

India's university system already carries a significant revaluation backlog — driven partly by students' reasonable distrust of the original evaluation. If AI grading produces systematically biased outputs that are detectable by statistical analysis, the volume of revaluation applications would not decrease. It would likely increase, now with an AI-versus-human-examiner dynamic that courts and regulators would need to resolve.

What the Research Supports Instead

Neither the Cambridge nor the Cardiff-Melbourne study argues that AI has no role in assessment. The Cambridge researchers were specific about the distinction: AI tools are potentially useful as decision-support for human evaluators, flagging possible errors, highlighting outlier marks, or supporting second-read workflows. They are not ready to be decision-makers for final grade assignment.

This is precisely the architecture that well-designed on-screen marking systems already use. The evaluator — a human teacher with subject expertise — views the digitized answer book and assigns marks question by question. The system automates totalling, cross-checks for unchecked questions, and flags anomalies for review. It does not attempt to replace the judgment of the evaluator. It makes the evaluator's work more reliable, more auditable, and faster.

India's scale problem — 15 million scripts across dozens of languages, evaluated by hundreds of thousands of teachers working in distributed settings — is a logistics problem and an accountability problem. Both of those problems are solved by structured on-screen marking. Neither of them is solved by replacing human evaluators with AI systems that the current research demonstrates cannot reliably assess academic quality.

The Evaluator Training Gap

There is a further implication for how AI tools should be used in evaluation. If AI tools are used to calibrate evaluators — showing them where their marks deviate from a norm, flagging potential inflation or deflation tendencies — the research supports that use. Statistical quality control tools built on AI can help institutions identify evaluator drift without removing human judgment from the final mark. That is a use case the research does not contradict.

Where This Leaves the Indian University System

The CBSE OSM controversy of 2026 produced two kinds of responses in the Indian education policy community. One response was to call for higher standards in the implementation of human-led digital evaluation: better evaluator training, more rigorous calibration, better quality control in scanning and mark entry. The other was to suggest that the solution was AI.

The 2026 research supports only the first of those responses. The problem with India's examination system is not that human evaluators exist. It is that human evaluators have been operating in a system that does not support consistency, accountability, or audit. On-screen marking with properly trained evaluators, structured marking schemes, and digital audit trails is the evidence-based answer to that problem.

AI grading, at this stage of development, is not. The research is clear: AI does not understand academic excellence well enough to assess it reliably. It understands fluency, vocabulary, and length.

Those things are not the same as understanding.

---

Related Reading