"Is an AI undress app safe" splits into two questions the forums and the courts answer differently: safe for whom? This page lays out the consensus on each — the uploader's exposure and the depicted person's exposure — because conflating them is how most coverage of this category goes wrong.
Across the category, submitted photos are retained — standard terms grant operators a perpetual license, uploads feed retraining corpora, and multiple services' databases have leaked with victims' photos and users' payment details inside. Even the user forums treat "assume your upload is public eventually" as baseline advice.
No service in the category verifies or solicits the subject's consent. Every bloc in the position map — including the user community — treats this as the defining property, not an incidental flaw.
There is no jurisdiction and no community where the tools' use on minors is defended. Synthetic CSAM is prosecuted identically to photographed CSAM, and even the tools' own communities police it as the one unforgivable use.
Whether generating without sharing is itself wrongful splits the blocs: criminal in the UK and Korea, civil or unaddressed in parts of the US, "violates the depicted person regardless" per the advocacy bloc, "the harm lives in sharing" per a sizeable minority of users. The legislative trend runs toward criminalization.
Both the research bloc and the platforms agree that classifier scores alone don't convict or clear: regenerated regions carry statistical artifacts and missing provenance, but frontier output and re-compression defeat single-image verdicts. The durable defenses are provenance signing of real media and hash-matching of known material — not "run the photo through a detector."
| Exposure | What the consensus says | Severity |
|---|---|---|
| Legal | Criminal in UK/KR at creation; criminal nearly everywhere at sharing; civil liability in most US states | High & rising |
| Data | Victim's photo + your payment trail held by anonymous offshore operator; documented leaks | High |
| Financial | Scam rate is community-acknowledged: dead domains take credits, "free trials" harvest uploads, refunds don't exist | Moderate |
| Traceability | Card payments and Telegram Stars leave platform-side records; "anonymous" crypto is routinely traced | Underappreciated |
The consensus here comes from the advocacy bloc's casework and the psychological literature, and it is blunt: the harm tracks photographic image-based abuse. The output inherits the victim's real face, pose, and context — the "it's fake" framing describes the pixels, not the violation. Threats of distribution work even when images are provably synthetic, because retraction is impossible once circulated.
The ordering is itself a consensus: communities, advocates, and platform-policy teams independently arrived at the same sequence because each step covers a failure mode of the previous one. Evidence before removal (you'll need it). Hashing after reporting (copies will resurface). De-indexing after hashing (domains will churn). Legal process throughout (it's now realistic).
One further consensus worth recording: the depicted person is never asked to prove the imagery is fake. Both support services and platform policies treat the non-consensual intimate depiction as the harm — the synthetic origin is a technical detail, and advocates report that victims who lead with "it's AI" get the same intake as photographic-NCII cases.