AI can create realistic child sexual abuse images without a real child ever being involved, challenging long-established assumptions about victims, harm and the purpose of criminal law.

For decades, child sexual abuse image laws rested on a simple premise: every indecent photograph of a child represented the abuse of a real child – but generative AI has upended that assumption. Today, software can create images that appear indistinguishable from photographs of real children, forcing legislators, prosecutors and courts to confront questions that would once have seemed purely theoretical: what, precisely, is the law seeking to punish? And if there is no victim, is there a crime?

Across the UK, the US and Europe, the answer is proving anything but straightforward. While governments have moved swiftly to strengthen the criminal law, the legal rationale has evolved. Once centred almost exclusively on protecting identifiable victims from exploitation, the debate increasingly encompasses questions of public safety, offending behaviour and whether harm must be directly attributable to an identifiable victim at all. The upshot is a rapidly developing – albeit profoundly distressing – area of law that sits at the intersection of child protection, criminal justice and technology.

A legal framework built around traditional photography

Child sexual abuse image offences were originally designed for a world in which cameras recorded reality on film, with every image depicting an actual child who had been abused in its creation. Criminalising the production, possession and distribution of such material therefore served an obvious purpose: reducing demand for images whose existence depended upon the exploitation of real children.

The arrival of digital technology complicated matters, but prescient legislators anticipated some of the challenges to come. In the UK, the Protection of Children Act 1978 was amended by section 84 of the Criminal Justice and Public Order Act 1994 to include “pseudo-photographs”, extending the legislation to realistic synthetic imagery decades before generative AI entered the public consciousness. The Coroners and Justice Act 2009 added offences relating to certain non-photographic prohibited images of children that fell outside the pseudo-photograph provisions.

Vicky Lankester, associate at London law firm Brett Wilson, which specialises in criminal and media law, explains that existing law effectively treats AI-generated child sexual abuse material (CSAM) in the same way as material created without AI. Section 1 of the 1978 Act criminalises taking, distributing or showing an indecent photograph or pseudo-photograph of a child, as well as possessing such material with a view to distribution, while section 160 of the Criminal Justice Act 1988 separately criminalises possession. Crucially, the 1978 Act defines a pseudo-photograph as “an image, whether made by computer-graphics or otherwise howsoever, which appears to be a photograph”.

“It is irrelevant whether the material is an actual photograph of an actual child, or whether it has been created by technology and/or AI and/or other means if it appears to be a photograph,” Lankester tells GLI News.

But Daniel Jackson, a partner specialising in criminal litigation at BCL Solicitors, identifies a potential grey area. Photorealistic AI material can be prosecuted as a pseudo-photograph, while obviously artificial or cartoon-like material may constitute a prohibited image under the 2009 Act. Advances in generative AI, however, mean that images “increasingly exist on a spectrum between clearly artificial and highly photorealistic”, he says, potentially creating uncertainty over which offence applies to material that is highly realistic but retains recognisably synthetic features.

Recent reforms have therefore been aimed less at filling an obvious gap in the law than confronting the scale, accessibility and sophistication that generative AI has introduced.

The non-existent child

Although public debate often treats AI-generated abuse imagery as a single issue, there are in fact two quite different scenarios.

The first is where AI is used to manipulate photographs of an identifiable child into sexually explicit images. Here, a real child has been targeted, with potential harms encompassing dignity, privacy and psychological wellbeing, as well as the continuing trauma caused by repeated circulation.

The second, and more difficult, category concerns wholly synthetic images depicting children who do not exist. These may appear entirely realistic even though they have been generated from a text prompt rather than a photograph, creating a jurisprudential challenge because no identifiable victim needs to have been abused in their creation.

Yet UK law does not make that distinction. Lankester says there is no meaningful legal difference between an AI-generated sexual image based on a real child, a wholly synthetic image of a child who has never existed and a genuine photograph of a real child. Nor, she points out, are offenders distinguished on that basis when sentenced – although whether they should be is another matter. “In my opinion, the law should draw a meaningful distinction, at least between AI-generated sexual images and real photographs of real children, due to the likely harm caused,” she says.

Where a genuine photograph is involved, Lankester argues, the child will almost inevitably have suffered harm, while an AI-generated image of an identifiable child also creates a real likelihood of harm. With an entirely synthetic child, however, “no harm can be caused to that ‘child’”.

It is the second category that has become the focus of intense legal and philosophical debate – and one that Jackson notes has yet to be squarely tested by the courts. In the 2024 prosecution of Hugh Nelson, who was sentenced to 18 years’ imprisonment for offences involving AI-generated child abuse imagery, the material was based on real, identifiable children whose faces were superimposed on 3D models. The case therefore demonstrated that existing law can deal severely with AI manipulation of real children, but did not test how the courts would approach wholly fictional CSAM.

The UK doubles down

Rather than reopening the question of whether wholly synthetic child abuse images should be criminalised, the UK has largely answered it through legislation. Recently, Parliament has expanded the law beyond the images themselves to address the AI ecosystem surrounding their creation and distribution. The Crime and Policing Act 2026 has introduced offences targeting AI tools specifically designed or adapted to generate child sexual abuse material. It also criminalises possessing, publishing or distributing so-called ‘paedophile manuals’ containing instructions on using AI to create such material, as well as operating websites that facilitate its distribution.

That approach presents difficulties of its own. Jackson points out that many AI models can produce both lawful and illegal material depending on how they are used, making it difficult to distinguish legitimate systems from those intended for offending. “The distinction is likely to become increasingly blurred as AI systems continue to develop rapidly,” he says. The difficulty for lawmakers is therefore deciding where liability should fall: on the technology itself, on those who deliberately design or adapt it to create illegal material, or on those who use otherwise legitimate AI tools for criminal purposes.

There are practical policing concerns too. Generative AI can produce vast quantities of highly realistic material at unprecedented speed, potentially consuming forensic resources and making it harder to identify genuine victims depicted in authentic abuse imagery.

America asks a different question

The United States approaches the issue very differently.

The defining authority remains the Supreme Court’s 2002 decision in Ashcroft v Free Speech Coalition, which struck down provisions of the Child Pornography Prevention Act 1996 criminalising certain virtual child pornography where no actual children had been used. The majority reasoned that the principal justification for prohibiting child pornography lay in preventing the sexual abuse of children during its production. If no child had been exploited, that rationale was absent and the legislation infringed First Amendment protections.

Congress fired back with the PROTECT Act 2003, adopting a more narrowly drawn prohibition intended to survive constitutional scrutiny.

Now, generative AI has revived questions that Ashcroft appeared to leave unresolved. If software can produce images practically impossible to distinguish from genuine...

GLI

Message History

Delivery option(s)