Srovnávací analýza detekce AI generovaných textů v češtině

This bachelor's thesis addresses the issue of automated detection of AI-generated texts in Czech. The theoretical part of the thesis describes the functioning of large language models, with a focus on their application in Czech, and presents two main methodological approaches to detecting machine-generated text: statistical methods and machine learning methods. The following chapter analyzes four commercial detection platforms, GPTZero, Originality.ai, Pangram, and DetekceGPT.cz, and compares their capabilities. In the practical part of the thesis, a dataset of 108 Czech texts was compiled, which were equally divided into three categories: human texts, AI-generated texts, and hybrid texts modified using obfuscation methods. All four tools were tested using standard classification metrics. The results show that the detection reliability between these tools differs significantly. The Originality.ai platform achieved the highest accuracy, but at the expense of the highest false-positive rate. The Pangram and GPTZero platforms ranked second and third respectively. Furthermore, the weakest results were recorded by the DetekceGPT.cz detector. The results confirmed that intentional obfuscation can reduce the detection ability of more sensitive tools.