Deepfakes
Deepfakes -- synthetic media created by swapping faces, modifying images, or generating video from AI models -- have introduced a new category of exploitation material. While deepfake technology affects adults broadly, its application to minors creates distinct legal, investigative, and harm-reduction challenges.
Deepfakes in Child Exploitation
Offenders use face-swap tools and image-to-image diffusion models to place children's likenesses into abusive contexts -- often starting from benign photos scraped from social media profiles. Unlike traditional CSAM depicting real abuse, deepfakes may not involve a recorded offense -- but they cause real harm to depicted children and normalize synthetic exploitation.
Detection is harder than hash-matching: each generated image is novel. Platforms and investigators must classify content as abusive without a known hash library entry, often requiring human review under traumatic conditions.
Technical Landscape
- Face-swap models -- Map a child's face onto existing abusive content or benign video
- Image-to-image diffusion -- Modify clothed photos to generate nude or abusive variants
- Video generation -- Emerging capability to produce synthetic video from single images
- Voice cloning -- Synthetic audio used alongside visual deepfakes in grooming and sextortion
Legal and Investigative Challenges
Legal frameworks vary: some jurisdictions treat synthetic depictions of minors as CSAM regardless of whether a real abuse event occurred; others require proving harm to a specific identifiable child. Offenders have raised AI generation as a defense in prosecution. Investigators must determine whether material depicts a real child, a synthetic composite, or a real child's likeness applied to synthetic content.
The Idaho AI-CSAM prosecution (2024), documented in the landscape briefing, establishes early precedent for how courts may handle synthetic material.
Detection Approaches
Research into AI-generated content detection is active but lagging generation capability. Platform trust-and-safety teams, NCMEC, and IWF are developing classifiers for synthetic abuse imagery. Hash-matching (PhotoDNA) fails on novel synthetic content -- requiring perceptual and model-based detection instead.