The Commercialization of AI Nudification: Inside the Ecosystem of Image-Based Sexual Abuse and How Researchers Can Study It Ethically

In January 2024, AI-generated sexually explicit images of Taylor Swift spread rapidly on X, with one post exceeding 45 million views within 17 hours. Commentators framed it as a warning about the future, but the authors stress that such abuse is already happening at scale. A broad ecosystem of AI 'nudification' tools, which claim to estimate the body beneath a subject's clothing, has operated openly on the web for years, some claiming operation since 2018.

The researchers set out to understand the infrastructure that enables these images. They acknowledge that studying image-based sexual abuse requires direct engagement with harmful content, a choice made out of responsibility rather than desensitization. Security researchers, they argue, can not only expose how tools of abuse operate but also advocate for those harmed using an informed scientific understanding.

The researchers examined twenty websites as of August 2024. Upon landing on these sites, users encounter non-consensual nude imagery, including celebrities inserted into explicit scenes. Animated demonstrations show clothed women transformed into nude forms. The platforms present their computer vision techniques, training claims, and user numbers (one claiming over 100,000 daily users) as signs of normalcy.

Previously, creating deepfakes required technical skill, computing resources, or access to niche forums such as Mr. Deepfakes, where commissions averaged $87.50 per video. Nudification sites charge an average of $0.64, sometimes requiring only ad viewing or a few cents. Nineteen of twenty sites focused on undressing women. They are advertised on mainstream platforms such as Instagram and app stores. None of the twenty platforms verified that the depicted person consented. Seven claimed subjects must be 18 or older, but buried this in terms of service. Ten allowed browsing without an account, and four never required login. Platforms offered Google, Discord, Apple, or X single sign-on and payment through Visa, Stripe, PayPal, Google Pay, and cryptocurrency gateways, enabling near-anonymous use. Some sites scrape images directly from Instagram, showing that anyone with a face can be targeted.

More than half of the platforms offered functions beyond simple undressing. Nine allowed users to place image subjects into sexual positions or videos through selectable sexual acts. Five offered body modifications such as breast or buttock enlargement and apparent age changes, with age options observed only for adults. Some added stylization filters such as anime or furry transformations.

Research on image-based abuse shows that abusers often seek social currency from sharing images. Nudification platforms exploit this by watermarking outputs or including referral links, turning generated images into advertisements. Affiliate programs paid revenue shares up to 50%, and white-label APIs let buyers launch their own nudification sites with custom branding, pricing, and watermark removal. The authors describe this as an industry blueprint for commercialized synthetic sexual exploitation rather than a democratization of deepfakes.

Studying these systems can cause psychological harm to researchers. During the study, one site depicting well-known women in non-consensual explicit acts caused immediate distress among team members, who responded physically. The team responded with content flags, optional task rotation, a warning system, and access to an on-call mental health professional. The authors emphasize that psychological safety requires a culture in which discomfort is valid and boundaries are respected, especially for early-career researchers.

Academic research does not remove the harm to real people whose likenesses are used. Testing these systems with real or consenting faces risks reproducing the abuse being studied, and consent is complicated by power dynamics, institutional pressure, and unforeseeable downstream model misuse. The authors are exploring fully synthetic human inputs, using AI-generated images with covered faces, to avoid creating new victims. They also stress responsible handling of stored and shared outputs.

The authors recommend: critically examining consent in datasets and team structures; developing internal protocols for physical, financial, and psychological safety, including burner accounts and devices; discussing ethics continuously rather than treating IRB approval as sufficient; recognizing that IRB frameworks may miss long-term psychological, reputational, and replication harms; and treating consent as an ongoing process rather than a final justification.

The abuse documented is technical, intimate, and rooted in power, both the power these tools simulate and the power they grant users. Studying this ecosystem responsibly requires rigor, reflection, caution, and care for both the people depicted in the images and the researchers who study them. The research was supported by the US National Science Foundation and the University of Washington Tech Policy Lab.