The loophole a court exposed
On August 20, 2025, Judge Lee Jung-hoon of the Uijeongbu District Court's Goyang Branch acquitted a man surnamed Kim who had distributed AI-generated nude images in a Telegram chatroom. Prosecutors charged him under South Korea's Act on Special Cases Concerning the Punishment of Sexual Crimes — the law colloquially known as the "Deepfake Prevention Law," which carries up to seven years in prison. The judge ruled that nothing in the case file confirmed the images depicted a real, identifiable woman, so the statute — written around real victims — simply did not reach the conduct. The court wasn't wrong on the law; it correctly read a 2020 statute that presumes a human victim, then applied it to a generative-AI reality that statute never anticipated (Business & Human Rights Resource Centre).
What the bill would do
The legislative response arrived fast. On September 11, 2025, Rep. Huh Young and nine other lawmakers from the ruling Democratic Party filed a bill (No. 2212918) amending the sexual-crimes act to add a new provision covering AI-generated sexual content of "virtual persons or representations that could be mistaken for a real person" — with no requirement that a real victim be identified. It carries the same penalties as real-person deepfakes: up to seven years in prison or a 50 million won (~$36,000) fine for creation or distribution, and up to three years or 30 million won for mere possession or viewing. The bill cleared general committee referral and, per Korean reporting in late July, was sent to the Legislation and Judiciary Committee's subcommittee for detailed, article-by-article review — the stage where South Korean bills typically get contested line by line before a floor vote (National Assembly Bill Information System; Korea Times).
The case for closing the gap
Lawmakers have real numbers behind them. Deepfake-specific sex crime cases rose from 31 in 2020 — the law's first year — to 550 in 2024, a roughly 17-fold jump in four years, according to the Korean Women's Development Institute's 2026 gender statistics analysis (Women's News). South Korea's National Police Agency separately reported 963 arrests in a single crackdown running from August 28, 2024 to March 31, 2025 — a 260% jump over the prior comparable period, with more than 90% of suspects in their teens or twenties (Korea.kr). Generative tools have also made the "it's not really her" defense trivially cheap to raise: a defendant can now plausibly claim any image is synthetic composite rather than a photograph, and if prosecutors must forensically prove a real subject, an entire category of humiliation and harassment escapes punishment entirely just because the technology got good enough to blur the line. Treating a photorealistic, sexualized image of what looks like an identifiable person as harmful regardless of whether a name can be attached to it is a coherent, defensible position — the psychological and reputational injury to a viewer, a community, or a look-alike doesn't obviously depend on provable photographic provenance.
Why the fix overshoots
But the bill as drafted doesn't stop at images that could pass for a real person — its "virtual person" language reaches wholly fictional, non-mimetic AI erotica with zero real-world referent, and applies the identical seven-year ceiling used for deepfakes of actual people. That conflates two very different harms: humiliating a real or plausibly-real person, versus generating pornography of nobody at all. During the bill's public comment period (September 17–26, 2025), roughly 18,000 submissions to the National Assembly's own system objected, most on free-expression grounds. Gachon University law professor Choi Kyung-jin identified the doctrinal problem precisely: without a victim, courts have no fixed reference point for what counts as "sexual humiliation" or whose rights were violated — the same ambiguity that produces inconsistent, hard-to-appeal verdicts (Hankook Ilbo). A law that criminalizes purely synthetic, victimless content at the same severity as image-based sexual abuse of real people isn't closing a loophole so much as reviving content-based obscenity policing under privacy-law branding — precisely the kind of proportionality problem South Korea's own constitutional court has flagged in past speech cases.
A narrower fix exists
The legitimate target here is narrow: sexually explicit AI content that a reasonable viewer would take for a specific, identifiable real person — a co-worker, classmate, or public figure — regardless of whether the defendant can be forensically proven to have used that person's actual photo as a training input. That standard preserves the real-victim logic of the 2020 law while closing exactly the loophole Judge Lee's ruling exposed, without also sweeping in generic AI-generated erotica that resembles no one in particular. Pairing a resemblance-based test with today's proposed penalty tiers — heavier for creation and distribution, lighter for possession — would let the subcommittee close the Goyang gap without asking prosecutors, and eventually the Constitutional Court, to referee an offense with no complainant.