Policy#AI safety#Image generation
Denmark’s Supreme Court: 18 months in prison for AI-generated child sexual abuse imagery, with no discount for fictional subjects
Denmark’s Supreme Court upheld 18 months in prison for a 31-year-old with over 36,000 AI-generated CSAM images: fictional subjects are no ground for leniency.
On October 15, Denmark’s Supreme Court (Højesteret) upheld an 18-month unconditional prison sentence for a 31-year-old man convicted of possessing and distributing more than 36,000 AI-generated child sexual abuse images. The case number is 98/2025, and the court’s summary, the anonymized judgment and a statement from the Danish Prosecution Service were all published the same day. The sentence matters less than one line in the court’s reasoning: when sentencing possession and distribution of sexual material depicting people under 18, no account shall be taken of whether the depicted persons may be assumed to be fictional.
Key points
- Coverage: the court held that § 235(1) and (2) of the Danish Criminal Code already covers AI-produced visual representations, and that the defendant’s fictional-subjects defense fails.
- Aggravating facts: of the 36,209 images, more than 12,500 showed sexual activity (category 2) and 113 showed particularly serious sexual conduct (category 3); distribution was systematic and partly for profit.
- Operation Cumberland: Europol reported on February 28, 2025 that a Danish-led operation across 19 countries identified 273 suspects and arrested 25 — Europol does not name this defendant as the operation’s main suspect.
What the ruling actually decides
The man, from Randers, admitted before the Randers city court that he generated 36,209 images with AI and distributed them online, but denied breaking the law. His appointed counsel, Lasse Hummelhof Frandsen, summarized the defense to Ritzau: § 235’s term for visual representations means real, existing people. The Supreme Court had to settle that semantic question.
Its answer comes from the 2003 preparatory works. When Denmark amended § 235 that year to criminalize “fictional” child pornography, the legislative notes spelled out what “representations” would cover, including an example that reads as if written for this case: a computer-generated image identical to a photograph apart from the fictitious motif falls inside the provision. The court held that § 235 therefore covers AI-produced material even when the depicted persons may be fictional, provided the output appears the same way as, or approximately the same way as, a photograph. It then went a step further than past case law on sentencing: the fictional nature of the subjects is simply not to be weighed.
The judgment records three defense arguments and rejects them. First, that the text reaches only real persons. Second, that the defendant should walk or see the penalty reduced as excusable ignorance or mistake of law under § 83 and § 82(1)(4). Third, that computer-generated material should carry a lighter sentence than material of a real abuse. The prosecution’s answer to the third point is the sentence most worth keeping on file: criminalizing fictitious material aims to protect children generally and counter the sexualization of children, and neither the wording nor the preparatory works distinguish AI-generated material from other material. The court also noted a practical observation of its own — as the technology advances, deciding whether the people shown in sexual material are real or fictional becomes steadily harder.
Three courts, three different sentences
The escalation across three instances is the part of this file that reads less like precedent and more like a live disagreement about proportionality.
| Court | Judgment date | Sentence | Suspended |
|---|---|---|---|
| Randers City Court | January 8, 2025 | 1 year 3 months | 3 months to serve, rest suspended |
| Vestre Landsret (Western HC) | June 12, 2025 | 1 year 6 months | none |
| Supreme Court | October 15, 2026 | 1 year 6 months | upheld |
The high court aggravated the sentence citing the volume of images, the scale and systematic character of distribution, and the profit motive. Prosecutors had asked for 1½ to 2 years unconditional; the Supreme Court left the 18 months in place. One reporting conflict is worth flagging: IT之家, citing Bloomberg, put the first-instance sentence at “15 months in January 2025,” while Kristeligt Dagblad and DK AI both report one year and three months with only three months to be served. Neither the Supreme Court’s summary nor the judgment states a first-instance sentence at all — the judgment records only that court’s name and date (Retten i Randers, January 8, 2025) — so the two Danish newsrooms are the sole source for that figure, and this entry follows them.
The court also refused the mitigation request under § 82(1)(9) — the provision the defense invoked for confessing the factual elements — and held that the large number of duplicates among the images did not disturb the finding of particularly aggravating circumstances.
The amendment and the enforcement
Two other Danish threads run alongside the case. On legislation, Act no. 722 of June 20, 2025 (in force July 1, 2025) inserted “or material equivalent thereto” into § 235(1) and (2), brought production into § 226’s scope, and allowed police to distribute fake sexual material of minors during investigations. That amendment postdates the high court’s judgment and predates the Supreme Court’s, but the court applied the version of the Criminal Code in force when the offense was committed, so the new text played no role.
On enforcement, Europol’s February 28, 2025 release on Operation Cumberland gives the operational context: 19 countries, 273 suspects identified, 25 arrested, 33 house searches, 173 devices seized. The main suspect, a Danish man arrested in November 2024, ran a platform where users could pay a symbolic fee to reach the material he produced. Europol called Operation Cumberland one of the first cases involving AI-generated CSAM and said investigators were hampered partly by “the lack of national legislation addressing these crimes,” with member states then discussing a common regulation proposed by the European Commission. Europol does not state that the Randers defendant was that main suspect, and this entry does not treat the two as the same person.
Third parties focused on exactly this gap. Save the Children Denmark welcomed the ruling; Pernille Spitz, who heads the organization’s anti-abuse work, said stricter law was the point because “technological development is constantly overtaking us.” The organization’s figures show the curve: 2,770 reports in 2024, 6,558 in 2025, and 4,069 in the first half of 2026, with 377 reports forwarded to police in 2026 so far. Christel Schaldemose, a vice-president of the European Parliament, called the decision “a completely correct ruling” in a written comment to Ritzau and wrote that protecting children has to keep pace with technology, adding that the EU has already banned AI tools developed to generate such material and that the companies behind them must be held accountable.
Our take
For anyone building image or video generation models, the value of this ruling is not the sentence. It is that the court made photographic realism the decisive feature — not whether a real victim exists. That is the same direction of travel as eight US states suing OpenAI and Florida’s bid to halt frontier model work: regulators on both sides of the Atlantic are regulating what a model can output, not only what was recorded. The two checkable points for model builders are training data and output filters. The prosecution argued in the first-instance proceedings that these models were trained on real images of abused children; the defendant disputed that, and the judgment drew no conclusion on it.
The strongest counter-view is on record too. Sten Schaumburg-Müller, professor of law at the University of Southern Denmark, told Ritzau that the ruling is one of principle — the Supreme Court has settled how § 235 should be read — but added that criminal policy could reasonably discuss whether a difference should exist. The court’s own answer to that is its harder observation: if nobody can reliably tell real from fictitious, a fictional-subjects discount is unusable in practice.
What is still open is how far this travels. The reasoning rests on Denmark’s 2003 preparatory works, and an EU-wide rule remains under discussion, so other member states may not follow the same route. Meanwhile the capability keeps shipping — Kling 4.0 went fully live on October 15, and this page’s cover was itself generated by a model.