Table 1 Central issues with banning or unrestricting AI at schools

Core idea

Upon the introduction of ChatGPT and comparable AI models, some educational institutions have banned their use for students’ theses and papers

Upon the introduction of capable LLMs such as ChatGPT, some institutions have fully allowed their students to use them for their academic papers and tasks with no or only little limitations

Key reasons

There are some strong reasons to prohibit the use of AI in academic papers:

• Students are often poorly trained in how to use these systems

• There is a high risk that students do not “think for themselves” anymore and hand over the work to machines

• Evaluating what is the proper work of the student and what is the work of an AI is mostly impossible

• There are manifest technical problem such as AI hallucinations leading to the models inventing things that may not be true

There are some arguments leading educators to wholeheartedly accept the full and mostly unrestricted use of AI by their students:

• It is the job of educators to teach students how to use new technologies

• Handing full responsibility to students may be the only way to help them learn to deal with the benefits and challenges of AI

• AI models will become integrated in all spheres of academia, the job market, and daily lives, and as such will be inescapable

• The more students are sheltered from the full scope of AI, the less they might learn its responsible use

• AI is here to stay and hence sooner or later must be dealt with

Key problems

The major problems with prohibiting AI in the work of students is twofold: (1) It is almost impossible for the school to control and make sure that students do not in fact use these systems. Very often, disallowing something with a high demand creates illegal use. (2) Also, since the technology will most likely be integrated into all aspects of people’s lives, it would be valuable to learn its proper and responsible use through the help of their educators

Even if educators provide very generous guidelines and best practices, the incentive to hand over the heavy load of one’s cognitive work to the computer may be very high. This leads to three main problems: (1) It is not clear if students have learned anything. (2) It is challenging to discern if students in fact did any of the cognitive work themselves and how this should be graded (after all, it is neither fun nor useful for teachers to grade a text purely or mostly written by ChatGPT). (3) And it is almost impossible for evaluators to make sure that students did not fall prey to any of the hallucinatory problems that arise from an LLM