Researchers have conducted an experiment showing that artificial intelligence can successfully pass university exams without being noticed by special programs. Ars Technica writes that ChatGPT’s exam papers score higher than students’ papers.

A team of scientists led by Peter Scarfe from the University of Reading (University of Reading) in the UK conducted a large-scale experiment to check how effectively modern AI systems can cope with university exams. The researchers created more than 30 fake psychology student accounts and used them to take exams using answers generated by ChatGPT. The experiment covered five modules of the bachelor’s degree in psychology, including tasks from three years of study.

The results turned out to be striking; 94% of AI-generated essays went unnoticed by examiners. Additionally, almost 84% of these essays scored half a point higher overall than human student essays. “Examiners were very surprised by the results “, – said Scarff. Moreover, it is interesting that some artificial intelligence work is discovered not because of its robotic nature, but because of its very high quality.

The experiment also highlighted the limitations of existing AI-generated content detection systems. Tools like Open AI’s GPTZero and the Turnitin system perform well in a lab setting, but their effectiveness drops significantly in real-world situations, Scarff said. But not all results were in AI’s favor. In the final course, which required deeper understanding and more complex analytical skills, human students outperformed ChatGPT.

Scarfe emphasized that due to the constant development of artificial intelligence and the lack of reliable ways to detect its use, universities will have to adapt and integrate artificial intelligence into the educational process. “The role of today’s universities is to prepare students for professional careers, and the reality is that students will undoubtedly use various AI tools when they graduate. “said the researcher. In fact, this experiment raises a problem that requires the revision of currently existing training and examination methods.