Dynamický Turingův test zkoumající schopnost vysokoškolských studentů rozpoznat skutečné malby od maleb tvořených umělou inteligencí

Generative artificial intelligence is a type of artificial intelligence designed to create new content, such as text, images, or video. As these models rapidly advance, the boundary between human artistic creation and AI-generated outputs is becoming increasingly blurred, making their distinction progressively more challenging. The primary aim of this thesis is to propose a modification of the Turing test based on the principle of dynamic testing, examining how well university students can distinguish between impressionist paintings created by humans and those generated by artificial intelligence. The dynamic principle further makes it possible to observe whether respondents improve over the course of the test as a result of repeated exposure to the task. Based on a review of relevant literature, a three-round test was designed and subsequently validated experimentally on a sample of 9 respondents. The test enabled the identification of qualitative decision-making strategies employed by participants, as well as the quantitative evaluation of their performance across rounds. The most effective strategy for distinguishing between the paintings was focusing on the complexity of the image. This approach proved successful for respondents, who began to rely on it more frequently as the test progressed. The quantitative results indicate a tendency toward gradual learning, although this improvement was statistically confirmed only partially: significant increase in accuracy was demonstrated between the first and third round. Furthermore, an item response theory analysis showed that the test was relatively easy for respondents and exhibited limited discriminative ability. This thesis presents a documented design and construction of a dynamic Turing test and provides several interpretable qualitative findings. However, to achieve clear interpretability and generalizability of the quantitative results, the experiment would need to be carried out on a substantially larger sample of respondents.