Bevers, Lisa (2026) Learning to Detect AI-Generated Images: Socio-Cognitive or Visual Processing Factors? Bachelor thesis, Psychology.
|
Text
Final-versionS3458938Group11Thesis.pdf Download (827kB) | Preview |
A thesis is an aptitude test for students. The approval of the thesis is proof that the student has sufficient research and reporting skills to graduate but does not guarantee the quality of the research and the results of the research as such, and the thesis is therefore not necessarily suitable to be used as an academic source to refer to. If you would like to know more about the research discussed in this thesis and any publications based on it, to which you could refer, please contact the supervisor mentioned.
Abstract
Generative Artificial Intelligence (GAI) is improving rapidly, creating increasing difficulty in distinguishing human made images from GAI images. This research explores whether inductive learning can improve GAI detection and whether individual differences in socio-cognitive skills, such as Theory of Mind (ToM), or visual processing skills, such as visual disembedding, contribute to detection accuracy. In an online experiment (N = 267) participants had to identify whether an image of a portrait expressing an emotion (happy, scared or angry) was GAI or human made, with a brief inductive training in the experimental condition. Participants additionally performed a ToM test (Reading the Mind in the Eyes) and a visual disembedding test (Leuven Embedded Figures Test). Results showed that inductive learning strongly improved GAI image detection. ToM showed marginally significant positive association with detection, while visual disembedding was not significant. No interaction was found between inductive learning and ToM. These findings indicate that inductive learning can improve GAI detection accuracy, independent of individual differences in ToM. After accounting for learning, ToM contributes to GAI detection, suggesting the socio-cognitive factors to be influential. This interpretation is further supported by the absence of a relationship between visual disembedding with both ToM and GAI detection accuracy. The results suggest the potential of inductive learning as a scalable intervention to improve GAI detection, however, future research is required to determine the longitudinal effects of training in order to improve humans ability to recognize GAI images from reality.
| Item Type: | Thesis (Bachelor) |
|---|---|
| Supervisor name: | Gutzkow, B. and Gordijn, E.H. |
| Degree programme: | Psychology |
| Differentiation route: | None [Bachelor Psychology] |
| Date Deposited: | 20 Mar 2026 14:50 |
| Last Modified: | 20 Mar 2026 14:50 |
| URI: | http://gmwpublic.studenttheses.ub.rug.nl/id/eprint/6335 |
Actions (login required)
 |
View Item |