AI Generated Images for Education and Work Life: On Bias and Guardrails

Abstract

The rapid multimodal development of Generative AI (GenAI) tools has opened up possibilities for content creation in many fields. This paper presents a study that had a focus on image generation for educational contexts and professional development in a course on artificial intelligence for education and work life. In an assignment, participants used different GenAI tools for image generation. Moreover, course participants analysed and discussed their AI generated images in essays. In the wide variety of GenAI tools two different image generation software were suggested, one from a big established IT company and the other from a small independent software developer. Both these tools were chosen because they are free to use without any licence fees, and that there are not any complex login procedures before using them. This was seen as important criteria for a group of course participants with relatively low pre-knowledge of GenAI and image generation. On the other hand, course participants with earlier experience of were allowed to use other and more advanced tools.  Images could be generated individually or in groups, but the final analysis essays had to be written individually without any AI assistance. Four portraits should be generated with each tool of two world-wide well-known persons, one locally known person and a portrait of the course participant that wrote the essay. 61 essays were thematically analysed with the use of open coding and axial coding. Results were divided into the categories of Age, Gender and Ethnicity, Language, Guardrails, and Training data, with Training data as the central or axial category. Findings show that the results to a certain degree are depending on prompting and the language, but that the found bias is depending on the training data.