A 16-fold jump that reframes the AI abuse debate
AI-generated deepfake pornography is no longer a fringe abuse pattern. A new study, “The Backbone of Abuse,” published in the Journal of Online Trust and Safety on September 30, 2026, estimates that AI-generated non-consensual intimate imagery videos rose from about 500,000 in 2023 to 8 million in 2025 . That is a 16-fold increase in two years, and it shifts the public discussion from isolated “misuse” of generative AI tools to an industrialized distribution system.
The study was led by Sarah Morgan of Lancaster University, with co-authors Dr. Sophie Nightingale, also of Lancaster, and Professor Hany Farid of Dartmouth College . The researchers use the term AI-NCII, meaning AI-generated non-consensual intimate imagery, rather than treating the material as ordinary pornography. That distinction matters: the central harm is not explicitness alone, but the use of a real person’s likeness in sexual or intimate contexts without consent.
Lancaster University’s release says the material mostly targets women and girls, while the sites examined by the researchers primarily focused on Hollywood and Hindi-film actors, K-pop idols, and well-known feminists and activists . The finding confirms what victims’ advocates have argued for years: deepfake sexual abuse is heavily gendered, reputationally destructive and difficult to remove once indexed, mirrored, downloaded or reposted.
What the researchers actually measured
The study did not simply count viral posts on social platforms. It mapped the infrastructure behind websites hosting AI-NCII. According to reporting that summarized the methodology, the researchers gathered 400 candidate web addresses through keyword searches and Google Alerts over a six-week period in February and March, then narrowed that set to 88 active sites hosting the targeted material . The distinction is important because 312 of the 400 candidate addresses were not deepfake abuse sites; they were unrelated websites using abuse-related keywords to attract search traffic .
Of the 88 confirmed sites, 38 exclusively hosted AI-generated non-consensual intimate imagery . Some sites hosted hundreds of thousands of videos, according to the university summary of the paper . The researchers also found that 93% of the identified sites were easily discoverable through Google Search . In other words, the problem is not hidden only in obscure encrypted spaces; much of it remains reachable through ordinary search behavior.
The paper also identified five recurring infrastructure providers across the studied sites: Cloudflare, Google, Namecheap, WordPress and Proton . Their roles varied. Cloudflare appeared most often in services such as hosting, content delivery, domain-name services and analytics; Google was identified in connection with SSL certificates and advertising; Namecheap appeared prominently as a registrar; WordPress was tied to content-management systems; and Proton appeared in mail-server infrastructure .
The infrastructure question
The most important contribution of the study is its focus on the “backbone” rather than only the visible surface. Earlier debates often centered on whether a user, model developer or social network should be blamed after an abusive image spreads. This research asks a different question: which ordinary internet services allow dedicated abuse sites to stay online, be found, monetize traffic and continue distributing victims’ likenesses?
The researchers are not saying that every provider knowingly supports AI-NCII. Superpower Daily’s summary notes that the authors do not allege that providers intentionally enabled the abuse . That caveat is essential. Large internet infrastructure companies serve millions of domains and cannot manually pre-screen every piece of content. But the study argues that once sites are verified as hosting non-consensual sexual deepfakes, providers should have clearer mechanisms to suspend, block or limit services.
Morgan’s proposed response is preventive, not only reactive. The Lancaster release says infrastructure providers should block sites that exclusively host AI-NCII and suspend services for mixed-content sites until good-faith efforts are made to remove the abuse . The researchers also recommend safeguards such as perceptual hashing, which creates digital signatures of confirmed abusive material so it can be detected and blocked when reuploaded .
This is a sharper intervention than traditional notice-and-takedown. A takedown system puts the burden on victims to find every copy, prove harm repeatedly and wait for each platform or site to respond. A hash-based or infrastructure-level approach aims to reduce redistribution before another victim has to file another report.
Search, monetization and discoverability
The study’s finding that 93% of the identified sites were easily discoverable through Google Search is one of its most consequential points . Search visibility transforms deepfake sexual abuse from a private harm into a public reputational attack. If a victim’s name is linked to fake explicit content in search results, the damage can reach employers, family members, classmates, journalists and strangers.
Google’s role is complicated because the company can appear in more than one layer of the ecosystem. According to Superpower Daily, Google told 404 Media that it needed specific domains to investigate and said its policies prohibit non-consensual explicit imagery, including AI-generated images . Google also said violations can lead to blocked ads or suspended advertiser accounts, and that people can request removal of such images from Search .
That response highlights a recurring enforcement gap. Platforms and infrastructure providers often require specific URLs, while victims may face a fast-moving web of copies, mirrors, renamed files and new domains. Researchers and civil-society groups can identify patterns, but companies often need confirmed addresses before acting. The paper’s implied challenge is to turn scattered reports into a trusted, verified pipeline that infrastructure providers can use quickly.
WordPress, software and responsibility
The study also triggered a debate over what counts as “infrastructure.” Superpower Daily reported that WordPress.org disputed the framing, saying it distributes open-source software and does not host or control independently run websites . That is a meaningful distinction: publishing software is not the same as operating a hosting service.
Farid, one of the study’s co-authors, countered that Automattic’s WordPress.com, Jetpack and content-delivery tools can host sites or serve their images even when a site is not hosted directly by Automattic . This disagreement matters because modern websites often depend on layered services. A site may have one registrar, another hosting provider, a third content-delivery network, separate analytics, third-party advertising and email infrastructure. Abuse can persist because no single company sees itself as the owner of the whole problem.
The policy question, then, is not simply “Who published the image?” It is also “Who can interrupt access, indexing, payment, delivery, storage or rediscovery once the abuse is confirmed?” The study’s answer is that infrastructure providers have more leverage than they have historically used.
Why the 8 million figure matters
The 8 million estimate is not only a shocking number; it changes the scale of response required. A problem measured in thousands can be handled case by case. A problem measured in millions requires automated detection, cross-platform coordination, legal clarity and trusted reporting pipelines.
The study also points to earlier evidence of enormous audience demand. Lancaster University’s summary cites a 2024 Channel 4 News investigation finding that almost 4,000 famous people were listed on five of the most visited deepfake pornography sites, as well as an industry analysis showing more than 4 billion views across 40 of the most visited deepfake pornography sites . That audience scale explains why individual takedowns often feel futile: the abuse is not just produced by lone perpetrators, but circulated through attention markets.
TechXplore’s republication of the Lancaster University report emphasizes that creation and sharing have accelerated because the material can be generated and distributed through websites, social platforms and private messaging services, some using end-to-end encryption . This means enforcement cannot rely on one chokepoint. If public websites are pressured, sharing can move to private groups; if a domain is suspended, a mirror can appear elsewhere. The response must be layered.
Law is necessary, but not sufficient
The study arrives as governments in several jurisdictions are criminalizing the creation, sharing or solicitation of AI-NCII. Lancaster University’s release specifically notes laws being implemented in countries including England and Wales, but Morgan argues that legislation alone is not enough . Her point is practical: a criminal law can deter some perpetrators and punish others, but it may not stop a website from remaining searchable, monetized and technically available.
This is the central policy tension. Law moves through courts and national borders; deepfake abuse moves through domains, search engines, content-delivery networks and encrypted channels. Victims need legal rights, but they also need operational systems that remove or suppress abusive content quickly, consistently and across services.
The strongest version of the study’s argument is not that infrastructure providers should become global speech police. It is that confirmed AI-generated sexual abuse should trigger a narrow, auditable and victim-centered response. That could include verified reporting by trusted organizations, rapid preservation of evidence for law enforcement, hash-sharing of confirmed abuse files, demotion in search, advertising bans and suspension of services for sites built around AI-NCII.
A digital security problem, not just an ethics problem
The surge from 500,000 to 8 million videos should also be understood as a digital security warning. A single public photo can become raw material for sexualized impersonation. The victim does not need to have shared an intimate image. The attacker needs only access to a likeness and a distribution channel.
That reality undermines basic assumptions about identity, consent and reputation online. It also creates risks beyond celebrity targets. If famous women, activists and performers are heavily represented because their images are abundant and searchable, ordinary people may be exposed through school photos, workplace headshots, social media posts or stolen image archives.
The study’s most urgent message is therefore structural. Generative AI lowered the cost of producing abuse; internet infrastructure lowers the cost of distributing it. If the response remains focused only on deleting individual files after victims discover them, the gap between harm and remedy will keep widening.
The publication of “The Backbone of Abuse” makes the next step clearer: regulators, platforms, infrastructure providers and researchers need a shared mechanism for identifying confirmed AI-NCII sites and cutting off the services that make them searchable, stable and profitable . Without that shift upstream, the jump from 500,000 to 8 million may be remembered less as a peak than as an early warning.