Study of 139 CivitAI guides finds AI intimate-image abuse is assembled from ordinary open-weight workflows

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Researchers from New York University, Georgetown University, Arizona State University and Indiana University posted a preprint on October 27, 2026, analyzing 139 sexual-content guides on CivitAI.
Key points
  • The authors find AI-generated non-consensual intimate imagery is "enabled by configuration rather than a distinct set of technical tools," combining sexual and real-identity content.
  • Guides teach these workflows with "little to no discussion of ethical implications or responsible use," and one guide explicitly outlines a workflow for non-consensual imagery.
  • Because most workflows rely on open-weight models run on consumer hardware, the authors say existing technical defenses are limited and easily circumvented.
  • They say model and product safeguards should focus on generating identity content, and that their findings can inform testing against realistic attackers, as new EU regulation requires.
Evidence
Research preprint, not peer reviewed, based on analysis of guides on one platform. It maps what the guides teach; it does not measure how much abusive imagery is made.

See it in the latest record