User Issues and Concerns in Generative AI: A Mixed-Methods Analysis of App Reviews

Generative AI models such as ChatGPT and Stable Diffusion have become easily available to end users through various apps. Research has identified several safety risks and limitations of generative AI, but the experiences and issues faced by real users of this technology in the wild have not been systematically investigated. In this paper, we identify user issues related to trustworthiness dimensions of generative AI, by analyzing user reviews of AI apps using a hybrid approach that combines unsupervised topic modeling and manual qualitative analysis. The results revealed user issues related to the validity, reliability, safety, security and privacy of the AI. Validity-related issues, such as incorrect output, were often found, but these issues appeared to result from high expectations about the capabilities of the technology, rather than an accurate reflection of its limitations. Concerns about safety issues, such as bias and the handling of inappropriate content, also appeared frequently, although users had conflicting expectations on how these should be handled. On the other hand, the user reviews contained fewer instances of concern related to the security and privacy of the AI itself. Overall, the results suggest that real users of generative AI have inadequate information about the characteristics and limitations of these models.

@InProceedings{BLTLP24chira,
  author    = {Vanessa Bracamonte and Sascha Loebner and Frederic Tronnier and Ann-Kristin Lieberknecht and Sebastian Pape},
  title     = {User Issues and Concerns in Generative AI: A Mixed-Methods Analysis of App Reviews},
  booktitle = {8th International Conference on Computer-Human Interaction Research and Applications (CHIRA 2024)},
  year      = {2024},
  keywords  = {privacy, machine learning},
}