Cambridge Study Finds AI Not Yet Reliable for Grading University Essays

A University of Cambridge-led team of psychologists and AI researchers evaluated three frontier large language models — including recent versions of Claude and ChatGPT (as of April 2026) — on their ability to grade real undergraduate work. The team assembled a corpus of more than 750 psychology essays submitted at three UK universities and compared AI-assigned grade bands (first, 2:1, 2:2, and so on) against the grades originally awarded by human examiners.
The AI systems matched the human examiners' grade band only 35–65% of the time depending on the model and essay set — far from the reliability required for high-stakes assessment. The researchers identified a specific failure mode: all three systems were oversensitive to surface linguistic features, systematically over-rewarding essay length and vocabulary range regardless of whether the underlying argument was strong. AI graders also routinely undervalued essays that human markers had rated as top-scoring, and overvalued some of the weakest submissions.
When asked to generate written feedback rather than just a grade, the AI systems produced comments three to eight times longer than those written by the original human assessors — verbose but not necessarily more useful to students. The Cambridge team's conclusion was explicitly cautionary: AI is 'not yet good enough' to grade open-ended university essays without significant human oversight, a finding intended to slow down universities considering automated grading before more rigorous safeguards and validation are in place.

Read the full analysis: https://www.cam.ac.uk/stories/ai-university-essay-grading