AI-generated answers scored, on average, half a grade boundary higher than those submitted by real students. Remarkably, 94 per cent of the AI essays did not raise any concerns with the markers, making them nearly undetectable
read more
A recent study conducted by researchers at the University of Reading has revealed that artificial intelligence (AI) can outperform real students in university exams. The study found that exams taken by fake students using AI tools were not only better than those taken by real students but also went largely undetected by markers.
The researchers created 33 fictitious students and used the AI tool ChatGPT to generate answers for module exams in an undergraduate psychology degree program, as per a BBC report.
The results showed that the AI-generated answers scored, on average, half a grade boundary higher than those submitted by real students. Remarkably, 94 per cent of the AI essays did not raise any concerns with the markers, making them nearly undetectable. The detection rate of 6 per cent is likely an overestimate, according to the study published in the journal PLOS ONE.
The study stated that this study should be a massive cause of concern considering how AI submissions have robustly gained higher grades than real student submissions. This means that students, if they choose to, could cheat undetected using AI – and in doing so, attain a better grade than those who chose not to.
Associate Prof. Peter Scarfe and Prof. Etienne Roesch, who led the study, emphasized the implications of their findings for educators worldwide. Dr. Scarfe noted that many institutions have moved away from traditional exams to make assessments more inclusive, but the study underscores the need to understand how AI will affect the integrity of educational assessments.
Going back to handwritten exams isn’t also a practical solution. However, the global education sector will need a robust solution and quickly at that, considering how AI is evolving, said Dr. Scarfe.
In the study, AI-generated exam answers and essays were submitted for first, second, and third-year modules without the knowledge of the markers. The AI outperformed real students in the first two years but not in the third year.
This discrepancy is consistent with the researchers’ observation that current AI struggles with more abstract reasoning required for third-year exams. The study is the largest and most robust blind study of its kind to date.
The influence of AI in education has raised concerns among academics. Glasgow University, for example, brought back in-person exams for one of their courses. Additionally, a study reported by The Guardian earlier this year found that while most undergraduates use AI programs to help with their essays, only 5 per cent admitted to pasting unedited AI-generated text into their assessments.
The findings from the University of Reading’s study serve as a “wake-up call” for educators to reconsider how assessments are designed and conducted in the age of AI. As AI continues to advance, the education sector must adapt to ensure the integrity and fairness of academic assessments.
