Výzkum schopností lidí rozpoznat výstupy z nástrojů využívající generativní AI: Možnosti zneužití a vliv školení

This bachelor's thesis investigates how people perceive, evaluate, and recognize visual content generated by artificial intelligence in the context of fake news on social media, and how a short educational intervention on the principles of generative AI affects their decision-making. The research is designed as a multi-phase qualitative study combining a literature review, a secondary analysis of three documented cases of misuse of generative AI (Hurricane Melissa 2025, the fake Pentagon explosion 2023, and the spread of AI visuals following the announced capture of Nicolás Maduro 2026), and semi-structured interviews with ten participants divided into a control group and an experimental group; the experimental group completed an approximately twenty-minute training session on the indicators of generative AI. The secondary analysis identified four recurring mechanisms contributing to the perceived credibility of synthetic content: an information vacuum, thematic compatibility with the event narrative, the exploitation of the sharer's authority, and rapid cascading across platforms. The thematic analysis of the interviews showed that contextual cues (source, platform, accompanying text) dominate the decision-making process and often override visual analysis itself; the five-category typology of visual implausibilities (Kamali et al., 2024) was not fully intuitively accessible to untrained users — anatomical and sociocultural categories appeared even without instruction, while physical and stylistic implausibilities required deliberate inspection that the training was able to activate. The short training led to a measurable shift both in identification accuracy (94% versus 60% in the control group) and, more importantly, in the nature of the decision-making process: it introduced a structured vocabulary, a contextual layer of judgment, and greater resistance to falsely confirming signals — particularly the blue verification badge. The thesis contributes to an understanding of how users actually reason when evaluating visual content under conditions resembling everyday media exposure, and proposes linking the expert typology of visual implausibilities with contextual analysis as the foundation of effective media-literacy interventions.