Non-consensual porn images and videos, filmed by groups now facing legal action, can live on forever in machine learning datasets used to create deepfake porn
@ruchowdh https://www.vice.com/... Samantha Cole / @samleecole : this “finders keepers” attitude of entitlement in the experimentation on women's images is repeated over and over—deepfakes, deepnude, AI porn, stealing content from sex workers. if I could count the times I've heard this stuff from men... https://www.vice.com/... https://twitter.com/... Rumman Chowdhury / @ruchowdh : In which I talk about ethical AI-generated porn: https://www.vice.com/... But seriously, there's nothing wrong using AI to make porn, but WHY do this based on exploitation and abuse? Thoughtful piece by @samleecole @vice : “It's mad science really, and completely and utterly re-victimizing to the victims.” https://www.vice.com/... Samantha Cole / @samleecole : we spent months researching the origins of machine learning datasets used for AI porn, and found that they're often full of non-consensual videos. new from @emanuelmaiberg @AKoslerova and I: https://www.vice.com/...
Context & Ripple Effects
Samantha Cole had already documented how deepfake pornography had become rampant on mainstream porn sites by mid-2020, with some videos drawing millions of views. This follow-up moves upstream to the training layer: footage from groups now facing legal action persists inside machine learning datasets, meaning prosecution of the original creators does nothing to remove what their cameras captured.
The piece lands as Rumman Chowdhury publicly presses the question of why AI porn gets built on exploitation at all when consensual alternatives exist. The dataset problem it exposes later resurfaces in CivitAI's model-sharing ecosystem, where models trained largely on non-consensual material proliferated through 2023.
First-order effects
- Women whose images sit in these datasets face re-victimization that outlasts any takedown or lawsuit against the groups that filmed them — every derivative model trained on the corpus regenerates the abuse.
- Legal action targeting the original filming groups leaves the extracted training data untouched, so victims' remedy gap shifts from the distributors of videos to whoever curates the datasets.
Second-order effects
- Model-hosting communities like CivitAI inherit the contamination: because popular image models are trained on scraped material taken without consent, one polluted dataset propagates through every downstream fine-tune shared on such platforms.
- The patchwork of state bans in Virginia and California cannot reach dataset composition, which strengthens the case federal lawmakers pursued when scrambling to punish AI-generated nude targeting after cases involving teen girls mounted.
Third-order effects
- Enforcement pressure migrates upstream from distribution sites to training-data curation itself — dataset provenance and consent documentation become the battleground where platform liability is argued.
- If the pattern holds, the Levittown-style prosecutions reported by Bloomberg show criminal law convicting individual perpetrators while the structural enabler — permissively scraped corpora — remains legally unaddressed without a federal framework.
The trend: Non-consensual imagery is migrating from distribution channels into machine-learning training infrastructure, where removals and prosecutions no longer reach it.