Artificial intelligence tools typically award higher marks on average than humans and cannot be relied on to give an accurate indication of a students’ performance, a paper has claimed.

With AI increasingly being explored by universities as part of the marking process to relieve pressure on time-strapped academics, the , published in the journal Assessment & Evaluation in Higher Education, found that generative AI tools such as ChatGPT “do not reliably reproduce human judgement in the marking of extended written work”.

The researchers uploaded 50 undergraduate bioscience essays to two versions of ChatGPT and asked it to grade the essays against seven assessment criteria under four different prompting conditions, finding there were “significant discrepancies” between AI-marked essays and those marked by humans.

In all but one case, the AI models typically returned higher average marks than humans. In one case the difference between an AI grade and a human one was 40 marks for an essay where the top score was 100.

ɫ��ֱ��

Lower-scoring essays tended to receive inflated marks, whereas higher-scoring essays received lower marks compared with human assessment. AI was more aligned with human markers on mid-grade essays.

The study says that the large language models (LLMs) tested “varied considerably” in marks awarded, and “were inadequate predictors of the human mark awarded to essays”. 

ɫ��ֱ��

“Despite being relatively consistent at producing similar marks when using the same prompt on the same essay, when used to mark a range of essays of differing standards, there was poor alignment between the LLM-provided marks and the marks assigned by the original human marker. Although the ‘overall’ mark for the essays were relatively well aligned on average with human marks, differences for individual criteria were considerable, reflecting the aggregating effect of a composite score,” it says.

The research was primarily motivated to evaluate whether generative AI could be used as a formative benchmarking tool for students, in addition to feedback from academics, rather than whether it could replace human markers.

Report co-author William Kay, senior lecturer in statistics (teaching and scholarship) at Cardiff University, said he had never quesitoned that responsibility for marking should remain a human one, but the study underlined that fact: "At present, GenAI is unable to reliably assign a mark to a subjective piece of written work comparable to that of humans – even with extensive training of the LLM,” he said. 

“While there is interest across the sector in whether the pattern-recognition capabilities of LLMs could facilitate objective grading of students’ work, making marking more efficient and relieving pressure on staff, the findings of this research indicate that at present this is not advisable. Aside from the ethical issues of submitting student work to GenAI tools without express consent, LLMs cannot, and should not, be relied upon for assigning grades to extended written work by students.

ɫ��ֱ��

“As LLMs become more sophisticated, it is possible that in the future their ability to mimic human judgement may enhance. But, as we find in this study, aligning marks between humans and GenAI may be hard to achieve,” Kay said. 

The paper notes that LLMs might be better equipped to predict “extreme” marks if marking criteria is more detailed and uses “objectively distinct descriptions for each mark bracket”, rather than relying on terms like “good, excellent, outstanding”, which “may be challenging for LLMs to differentiate”.

Although it concludes that “as LLMs become more sophisticated, their ability to mimic human judgement may enhance,” it adds that “high levels of inter-rater variation among human markers” could make this alignment “hard to achieve”. 

juliette.rowsell@timeshighereducation.com