Efektivita nástrojů pro detekci AI generovaného obrazu

This work addresses the issue of detecting AI-generated image content within the context of protecting digital authenticity. The primary objective is a comparative analysis of the effectiveness and robustness of selected detection tools against lossy JPEG compression. The theoretical part defines the functional principles of generative artificial intelligence and describes current methods of passive automated detection. In the practical section, six detection tools – three commercial APIs and three open-source models – are tested on a balanced set of 2,000 images from the Defactify dataset, which comprises outputs from five modern generative architectures. The results reveal significant differences in the success rates and internal certainty of the tools; professional systems (Hive, Sightengine) demonstrate high reliability, whereas other tested systems show a considerable error rate. Furthermore, the experiment examines the existence of a compression masking effect, which systematically reduces the success rate of identifying synthetic content for most models. The work provides practical recommendations for digital content verification in the current technological environment.