Exams written by ChatGPT 'virtually undetectable' and outperform students

Bolivian coup fails as new army boss orders troops back

University exams written by artificial intelligence are "virtually undetectable" and capable of outperforming the exams completed by humans, a new UK-based study has found.

The findings will likely fuel fears within the education sector around academic integrity, amid the ever-increasing ability of AI tools to mimic the writings of real people.

In the study published in PLOS ONE today, 33 fake students at the University of Reading submitted exam responses that were 100 per cent written by the AI bot ChatGPT-4 for five different subjects in an undergraduate psychology degree.

The AI-written essays made up five per cent of the total papers submitted and the subject coordinators and exam markers were unaware of the experiment.

All but two of the exams (94 per cent) went undetected as forgery by the markers.

Moreover, the chatbot's exam papers received higher marks on average than those of real students, with 83.4 per cent receiving higher grades than a randomly selected group of the same number of submissions from students.

The results demonstrate the rapid advancements that AI tools such as ChatGPT have made in their ability to mimic a real human voice and outsmart devices used to detect them.

"From a perspective of academic integrity, 100 per cent AI written exam submissions being virtually undetectable is extremely concerning," the study authors note.

"Especially so as we left the content of the AI-generated answers unmodified and simply used the 'regenerate' button to produce multiple AI answers to the same question."

They note that the ability of AI to go undetected will only increase over time, as its responses improve in both complexity and abstract reasoning.

"Clearly, we have no way to estimate the proportion of students in our sample who used AI to complete their submissions. However, with the huge media coverage of AI such as GPT-4... we struggle to conclude that its use would be anything other than widespread."