There is a working takedown infrastructure for pirated adult content — and it does not catch what is replacing piracy. MIT Technology Review reports that Takedown Piracy has fingerprinted over 500 million videos and pulled roughly 130 million from Google search alone. That pipeline works on fingerprint matching against known sources. AI-modified copies of the same performers — with small edits that erase a birthmark or shift body proportions — slip past fingerprinting because the new file no longer matches the original. The infrastructure that exists is the infrastructure for a problem that is shrinking.
The class of content is not face-swap deepfakes — it is body-based. Performers' actual bodies are used as training data and then regenerated with minor alterations. Researchers quoted in the piece estimate that more than 10,000 terabytes of online adult content likely sits in current AI training corpora, though they flag this as a "reasonable assumption" rather than a measured figure (no major lab publishes corpus-level filtering audits at this granularity). Named tools include the nudify apps Crushmate and Grok, and the sexting app Mynx, where scammers deployed AI deepfakes of named performers to defraud paying users. Spicey AI — now defunct — tried to formalize the problem the other way: contracts where performers signed over ownership of their AI likenesses.
The legal layer is forming but uneven. The federal Take It Down Act criminalizes non-consensual intimate imagery publication, which is teeth that did not exist before. Performers' attorneys quoted in the piece flag the dual-use risk: the same statute can be weaponized against consensual performance content under the wrong takedown notice. DMCA chains collapse entirely when hosts operate from Russia, the Seychelles, or the Netherlands. For image- and video-generation platforms shipping today, this matters at two layers: pretraining corpora almost certainly contain scraped adult content unless a deliberate filtering pass was made (and few labs document one), and the downstream detection problem — identifying that a generated output is a body-deepfake of a real person — is not solved by existing fingerprinting tooling.
If you ship image or video generation, the practical implication is that you cannot lean on the existing takedown ecosystem to catch outputs derived from your model. Embodied-harms research cited in the piece distinguishes body-dysmorphia and self-censorship as effects separate from the face-deepfake trauma class — different mechanism, different mitigation. The legal framework will keep moving; the technical detection gap will not close on its own.