"AI nude" means a nude or sexualized image produced by an artificial-intelligence model rather than a camera. The term covers two very different things: fully invented people (synthetic bodies no human ever posed for) and synthetic nudes of real, identifiable people — made by feeding someone's ordinary clothed photo into a model that fabricates what might lie underneath. The first is a legal gray-to-green niche. The second is non-consensual intimate imagery, and it is the reason the term matters.
"AI nude" is a search-term, not a technical one — nobody in the research community calls the technology that. It emerged as the plain-language way people describe what tools like DeepNude (2019) and its hundreds of successors produce: point an AI at a photo, get a nude. The phrase now blankets three distinct technologies, and almost every misunderstanding about the topic comes from treating them as one thing.
| Type | What it is | Who's depicted | Legal position (Nov 2026) |
|---|---|---|---|
| Synthetic-only | Entirely generated bodies — diffusion models inventing people who don't exist | Nobody — invented | Mostly lawful for adults; restricted by platform rules rather than statute |
| Real-person synthesis | A real person's likeness nudified — via inpainting ("undress" tools), face swaps, or identity-conditioned generation | A real, identifiable person — usually without consent | Criminal to share nearly everywhere; criminal to create in UK/Korea; civilly actionable in most US states |
| Minor-directed | Either pipeline aimed at minors | Minors | Prosecuted as CSAM everywhere — no "it's AI" defense |
When people argue about "AI nudes" they are almost always arguing about different rows of this table. A debate about artistic freedom (row one) and a debate about abuse (row two) share the same words and have nothing in common.
The dominant pipeline for real-person images is inpainting: a segmentation model outlines the clothing in a real photograph, and a diffusion model regenerates only that masked region — conditioned on pose, skin tone, lighting, and grain from the untouched pixels. The result preserves the victim's real face, posture, and setting, which is why output reads as "a leaked photo" rather than a render. Face swaps and fully-conditioned generation (LoRA fine-tunes trained on someone's photos) produce the other flavors. How does AI nude technology work? walks each pipeline in depth.
Because of who row two of that table is applied to. Since 2023 the victim pool shifted from celebrities to private individuals — classmates' photos processed in school incidents, partners' photos in domestic-abuse coercion, streamers' and colleagues' images in harassment campaigns. The depicted person's likeness is the content; their consent is nowhere in the pipeline. Researchers classify the output as AI-generated non-consensual intimate imagery (AI-NCII), and the psychological literature treats victimization as equivalent to photographic image-based abuse — the violation is the weaponized identity, not the pixels' origin.
The short version, with the full map on the dedicated law page: synthetic-only imagery of invented adults occupies the lawful end. Real-person non-consensual imagery is criminal to share essentially everywhere in the developed world (US federal Take It Down Act, UK Online Safety Act, EU directive implementations, Korea, Australia) and criminal to create in the strictest jurisdictions. Anything depicting minors is a serious felony in every jurisdiction, synthetic or not.