Mohamed Moustafa Dawoud (University of California, Santa Cruz), Alejandro Cuevas (Princeton University), Ram Sundara Raman (University of California, Santa Cruz)
Generative AI has enabled the large-scale production of photorealistic synthetic sexual imagery, yet prior work on non-consensual intimate imagery and deepfakes has focused mostly on underground forums and dedicated nudification tools. In this paper, we investigate whether these services have moved into mainstream gig marketplaces, where they benefit from larger user bases and higher trust.
We present the first systematic study of sexually explicit AI generation services (often advertised as AI NSFW services) on a major freelance marketplace, Fiverr. We discover these listings by employing a range of sampling approaches, including keyword searches, sitemap analysis, and snowball sampling, and confirm that they are sexually explicit through an LLM classifier. Through this process we identify 593 AI-enabled NSFW gigs. We also collect a set of control groups from other AI and non-AI categories (n=1,028). We use an LLM to extract each gig’s risk indicators, advertised tools, platform targets, pricing, and seller attributes.
Our results reveal a rapidly emerging market with new NSFW service freelancers joining at consistently higher rates than any other group we observed (74.9% of NSFW sellers joined in 2025). Within the NSFW segment, 82.8% expose deepfake-enabling features and 87.6% violate Fiverr’s policies on pornography and deepfakes. We also uncover a new type of service, not previously documented: custom sexually explicit LoRA/model training. Sellers disproportionately target downstream platforms such as OnlyFans (54.2%), Instagram (29.5%), and Fanvue (24.1%). For the usable security and privacy community, our results reframe abuse-enabling generative AI as a mainstream problem rather than a dark corner of the Internet.