Detecting Metadata-Conditioned Response Drift in AI-Generated Visual Information Systems: A Two-Phase Human–Computer Interaction Framework

AI-generated visual information systems often present images with authorship metadata, yet such labels may shape the evaluation data they are intended to contextualize. Existing work has seldom offered a system-level procedure for measuring metadata-conditioned response drift under unchanged visual input. This study develops a two-phase evaluation framework combining blind presentation, attribution-label overlay, trial-level logging, drift computation, and mixed-effects validation. Ninety-two participants rated four AI-generated images first without labels and then with attribution labels, producing 732 valid observations. Linear mixed-effects modeling, supported by ordinal sensitivity analysis and leave-one-image-out checks, identified a significant Phase × Label interaction within the implemented attribution-label assignment scheme. Under this assignment scheme, ratings for stimuli assigned to the human-labeled condition increased after disclosure, whereas ratings for stimuli assigned to the AI-labeled condition decreased. These findings are interpreted as evidence that the proposed framework can detect disclosure-associated response drift, rather than as a fully counterbalanced causal estimate of pure label effects.

@article{Jia2026,
  title     = {Detecting Metadata-Conditioned Response Drift in AI-Generated Visual Information Systems: A Two-Phase Human–Computer Interaction Framework},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Yu Jia and Shaojie Zhang and Na Li},
  doi       = {10.14569/IJACSA.2026.0170711},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170711}
}