Experts warn about pedophiles using generative AI to create CSAM, leading to thousands of images on dark web forums; some question whether laws cover the images
Drew Harwell / Washington Post : Twitter: @drewharwell and @richardhanania Twitter: Drew Harwell / @drewharwell : New: AI-generated images of child sex abuse are being made and traded by the thousands in the worst parts of the web. Photos of victimized children have been used to train the AI. DOJ tells me they're illegal, even when the child depicted doesn't exist https://www.washingtonpost.com/ ... Richard Hanania / @richardhanania : “Justice Department officials who combat child exploitation say such images still are illegal even if the child shown is AI-generated, but they could cite no case in which a suspect had been charged for creating one.” https://www.washingtonpost.com/ ...
Context & Ripple Effects
This June 2023 Washington Post report is the earliest alarm in a documented arc: months later, the Internet Watch Foundation confirmed thousands of AI-made abuse images breaking UK law, and by mid-2025 it counted 1,286 AI-generated abuse videos in six months, up from two a year earlier — the trajectory the Post's reporting foreshadowed.
The supply side was already contaminated: Stanford later found 1,008+ CSAM instances inside LAION-5B, the dataset behind Stability AI's models, and sites like CivitAI were hosting image models largely trained on non-consensually scraped material. The unresolved question in the piece — whether law covers images of children who don't exist — is the hinge everything downstream turns on.
First-order effects
- The DOJ asserts AI-generated CSAM is illegal even with no real depicted child, yet could cite no prosecution for creating such images — leaving enforcement authority declared in principle but unproven in practice.
- Dark web forums now host thousands of synthetic abuse images alongside real ones, forcing moderators and investigators to triage content where victim identification no longer applies.
Second-order effects
- Dataset and model distributors inherit the exposure: LAION-5B's CSAM contamination and CivitAI's consent-free training corpora put Stability AI-class companies under pressure to audit training data or face the same scrutiny as hosts of the output.
- Reporting pipelines built for photographic evidence strain first — Stanford's warning that synthetic images could overwhelm the federally authorized CyberTipline means tip volume grows while each report's forensic value to victim identification drops.
Third-order effects
- If DOJ can cite no prosecutions, the gap between stated legality and enforceability pushes lawmakers toward statutes explicitly covering fully synthetic depictions — the same scramble already visible in the push to punish AI-generated nudes targeting women and teen girls.
- Child-protection infrastructure shifts from photo-matching against known victims to detecting synthetic content at generation time, making dataset provenance and model-level filtering a compliance requirement rather than a research concern.
The trend: Generative AI is converting child sexual abuse material from a scarce, traceable commodity into a synthetically scalable one, dragging dataset governance, detection infrastructure, and pre-digital-era law into a reactive catch-up that is still unresolved.