The master's thesis addresses the problem of detecting AI-generated text in universities. It focuses on solving this problem with a proposed interface that enables transparent insight into the process of creating texts using large language models (LLMs). The empirical part was carried out in three phases: in the first phase, a literature review of existing detection methods and an analysis of the EU AI Act and the guidelines of the University of Ljubljana regarding the use of detectors were conducted. In the second phase, the solutions of foreign universities in this field were examined, and in the final phase, a theoretical plan and a prototype interface for prompt traceability and transparent proof of authorship were created. A comparative analysis was also conducted to determine whether such a systemic solution offers a better alternative for the transparent use of LLMs than existing detection methods. The research confirms the unreliability of classic detection methods, as they do not take into account the context of use, show systemic bias, and do not yield reliable results with simple editing or cleaning of texts. The proposed interface proved to be an effective possible solution, which, through insight into the entire interaction, accurately distinguishes between content generation and mere linguistic editing of the text. This reliably protects students from unjustified accusations, while offering educators an objective basis for fair evaluation of student work.
|
