Introduction
The rapid advancement of generative artificial intelligence (AI) presents a profound challenge to established legal frameworks. One of the most disquieting manifestations of this technology is its capacity to create synthetic yet highly realistic child sexual abuse material (CSAM). This development severs the historical link between such abhorrent imagery and the direct physical exploitation of a real child in its production. Consequently, the traditional legal rationale, grounded in preventing and punishing the abuse inherent in the image's creation, is unsettled. The current UK legal framework, primarily the Protection of Children Act 1978 and the Criminal Justice Act 1988, criminalises the making, possession, and distribution of ‘indecent photographs’ and ‘pseudo-photographs’ of children. While these provisions arguably capture the finished AI-generated product, the ease and accessibility of its creation have prompted calls for intervention at an earlier stage. This raises a fundamental question for criminal justice policy: should the law resort to preventive offences, criminalising preparatory acts, to combat the proliferation of AI-generated CSAM?
This essay will argue that while the threat posed by AI-generated CSAM is severe and demands a robust legal response, the expansion of preventive criminal offences is a disproportionate and principledly unsound solution. Such a move would represent a significant step towards a ‘pre-crime’ society, eroding fundamental tenets of criminal law, including the harm principle, legal certainty, and proportionality, without a guaranteed increase in efficacy. Instead, a more judicious and effective strategy involves the rigorous application and potential refinement of existing substantive offences, a strengthened focus on the distinct harms of dissemination and use, and the reinforcement of non-penal regulatory and technological measures. The role of the criminal law should be to punish culpable wrongdoing, not to police the ambiguous territory of preparatory conduct far removed from the consummated harm.
The Principle and Peril of Preventive Justice
The orthodox function of the criminal law is to respond to harms that have already occurred. Preventive justice, by contrast, seeks to intervene before a substantive offence is completed (Ashworth and Zedner, 2014). This is not a new concept; the criminal law has long embraced inchoate offences such as attempt, conspiracy, and encouraging or assisting a crime. These offences, however, are traditionally tethered to a specific, substantive crime and require conduct that is more than ‘merely preparatory’ (Gullefer [1990] 1 WLR 1063) or demonstrates a clear agreement to a course of conduct that will necessarily amount to or involve the commission of an offence (R v Saik [2006] UKHL 18).
The modern trend, however, is towards a more expansive form of prevention, characterised by stand-alone preparatory offences that criminalise conduct further removed from the eventual harm. A prominent example is section 57 of the Terrorism Act 2000, which criminalises the possession of an article in circumstances giving rise to a reasonable suspicion that the possession is for a purpose connected with the commission, preparation, or instigation of an act of terrorism. This model signifies a shift in focus from act and harm to risk and dangerousness (Zedner, 2007). The justification is pre-emptive: the potential catastrophe of a terrorist act is so great that state intervention is warranted at a much earlier, albeit more ambiguous, stage.
The adoption of this preventive model is contentious because it strains foundational criminal law principles. The harm principle, as articulated by J.S. Mill, posits that the state is only justified in coercively interfering with an individual's liberty to prevent harm to others (Mill, 1859). With preparatory offences, the ‘harm’ is abstract and probabilistic—it is the risk of future harm. This dilutes the principle, potentially authorising state intervention based on speculation about an individual's future intentions. This is closely linked to the principle of autonomy, which respects individuals as rational agents responsible for their choices. Criminalising preparatory acts risks punishing individuals for what they might do, moving uncomfortably close to penalising 'thought crimes' (Simester and Sullivan, 2019).
Furthermore, preventive offences often jeopardise the principle of legality, which demands that the law be clear, certain, and non-retrospective, enabling citizens to know what is prohibited. Preparatory conduct is often equivocal; possessing a piece of software, researching a topic, or writing code can be entirely innocent or the first step towards a serious crime. The mens rea for such offences must therefore do significant work to distinguish the guilty from the innocent, yet proving a future criminal intent can be an exceptionally difficult and speculative exercise for the prosecution. This creates a risk of over-criminalisation and a 'chilling effect' on legitimate activities, as individuals self-censor their behaviour to avoid suspicion (Ashworth, 2009).
Evaluating the Current Legal Framework Against AI-Generated CSAM
Before advocating for new preventive measures, it is essential to assess whether the existing law is truly inadequate. The primary legislation criminalising CSAM is the Protection of Children Act 1978, which concerns ‘indecent photographs of children’, and Part 1 of the Criminal Justice Act 1988, which creates parallel offences for 'indecent pseudo-photographs of children'. An AI-generated image, being synthetic, would fall under the latter category.
Section 7(7) of the 1988 Act defines a 'pseudo-photograph' as "an image, whether made by computer-graphics or otherwise howsoever, which appears to be a photograph". The subsequent subsections clarify that this includes images where a real child's image has been altered or a composite image has been created that includes a representation of a child. Crucially, section 7(8) states an image is to be treated as a pseudo-photograph if the 'overall impression' it conveys is that of a photograph of a child. This legislative focus on the appearance of the final product, rather than its method of creation, appears capacious enough to encompass AI-generated images. The case of R v Fell [2011] EWCA Crim 3014, while predating the current wave of generative AI, confirmed that entirely computer-generated images that appeared to be of children fell within the definition of a pseudo-photograph. Therefore, the substantive offences of making (which includes creating), possessing, and distributing AI-generated CSAM are, on the better view, already established in law.
The problem, then, is not a legal lacuna regarding the final product, but one of enforcement and scale. The argument for preventive offences is that waiting until a finished image is 'made' or distributed is to intervene too late. The technology allows for rapid, large-scale creation, and the damage of normalisation and desensitisation occurs as soon as the material enters circulation. Proponents of a preventive approach would therefore advocate for criminalising acts taken before the image is rendered, such as developing or possessing a specific AI model trained for this purpose, or perhaps even writing detailed prompts designed to elicit such images.
The Case for a Preventive Criminal Law Response
The argument for introducing new preventive offences rests on several interconnected pillars: pre-emption, deterrence, and the expressive function of the law. The primary justification is the precautionary principle. Given the profound societal harm associated with CSAM—including its role in normalising paedophilic ideation, its use in grooming and blackmailing actual children, and its corrosive effect on the value society places on childhood innocence—the state has a compelling interest to act pre-emptively. Waiting for a completed image to be created and disseminated may be seen as an unacceptable risk. The ease with which this material can be created and shared online means that once it exists, containment is exceptionally difficult.
Second, proponents would argue that new preventive offences would have a powerful deterrent effect. Criminalising the possession of specialised tools or the act of training an AI model on illicit datasets would send an unequivocal message that any engagement with this process is prohibited. This could dissuade individuals from exploring the creation of such content by raising the personal cost at a much earlier stage. From an enforcement perspective, it would theoretically empower police to intervene upon discovering not just finished images, but the digital tools and datasets associated with their creation.
Third, such offences would serve an important expressive function. The criminal law does not merely deter; it also declares and reinforces societal values. Creating a specific offence of, for example, 'possessing software for the purpose of creating indecent pseudo-photographs of children' would signal society's absolute condemnation of such conduct at its nascent stages. It would articulate that the wrongdoing is not just in the final product but in the intention and preparation to create it, regardless of whether a real child was involved in the process. This approach seeks to close what might be perceived as a moral gap, ensuring that the law's censure is not confined to the possession or distribution of the image itself.
The Overwhelming Case Against Preventive Offences
Despite the intuitive appeal of pre-emption, the introduction of new preventive offences to tackle AI-generated CSAM would be a dangerous and disproportionate development in UK criminal law. The principled objections are formidable and ultimately outweigh the posited benefits.
The most significant objection is the profound risk of over-criminalisation and the encroachment upon the "thought crime" boundary. The "articles" or "tools" used to create AI-generated material are not analogous to the bespoke components of a bomb or a cache of weapons. They are often powerful, general-purpose, and in many cases open-source, AI models like Stable Diffusion or Midjourney. These tools have countless legitimate and creative applications. Criminalising the possession of such software would be akin to criminalising the possession of a word processor because it could be used to write a terrorist manual. The law would inevitably have to rely on proving the accused's nefarious intent, an evidentiary burden that is both difficult to meet and fraught with the risk of misinterpretation. What evidence would suffice? A user's search history? Saved prompts? Forum discussions? This ventures into policing expression and curiosity, creating a chilling effect on legitimate AI research, digital art, and free inquiry.
This leads directly to the problem of legal certainty. A preventive offence would struggle to define the actus reus with sufficient clarity. At what point does a user's interaction with an AI model cross the line from experimentation to a criminal act of preparation? Is it in the drafting of a prompt? The fine-tuning of a model? The downloading of a specific plug-in? Any legislative attempt to draw this line would be arbitrary and likely technologically obsolete upon enactment. The result would be a vague and uncertain law, granting excessive discretion to law enforcement and creating the potential for arbitrary and discriminatory application, contrary to the rule of law.
Furthermore, such an approach is disproportionate. The substantive offences of making, possessing, and distributing indecent pseudo-photographs already carry severe penalties, reflecting the gravity of the completed harm. To attach similar criminal liability to ambiguous preparatory acts that may never result in a final image is to lose a sense of proportion. The criminal law's gradation of seriousness, from preparation to attempt to completion, is a cornerstone of fair labelling. A broad preparatory offence flattens these distinctions, treating the person who experiments with prompts but goes no further with the same blunt instrument as the person who actively creates and distributes a vast library of material.
Finally, the practical effectiveness of such a UK-specific offence is questionable. The development and availability of AI tools are global phenomena. Criminalising their possession in one jurisdiction is unlikely to halt their proliferation and may simply drive the activity further underground, making it harder for authorities to detect and monitor.
A More Principled and Effective Path Forward
Rejecting new preventive offences does not equate to inaction. A more effective and principled response lies in a multi-pronged strategy that utilises and refines existing legal and non-legal tools.
First, the focus should be on the robust enforcement of the existing substantive offences under the 1978 and 1988 Acts. The law as it stands is adequate to prosecute those who make, possess, or distribute AI-generated CSAM. Resources should be directed towards improving law enforcement's technological capacity to trace the creation and dissemination of this material. The Law Commission could be tasked with a review to confirm the legislative definitions remain fit for purpose in the age of advanced AI, but a radical overhaul through preventive offences is not required.
Second, the legal and regulatory focus should be squarely on dissemination and harmful use. The greatest immediate danger of AI-generated CSAM lies in its distribution across the internet and its use in other crimes. The Online Safety Act 2023 provides the appropriate model. Rather than creating a swathe of new individual criminal offences for preparation, the Act imposes a duty of care on service providers to take proactive steps to prevent users from encountering illegal content, including CSAM, and to swiftly remove it. This regulatory approach targets the platforms that facilitate dissemination, which is a more efficient and effective point of intervention than attempting to police the preparatory actions of millions of individual users. This is a form of prevention, but it operates at a systemic, regulatory level, avoiding the rule of law problems associated with individual preparatory criminal liability.
Third, the law must adapt to the new ways this material can be used to harm real children. A significant threat is the use of AI to create synthetic CSAM of an identifiable child for the purposes of blackmail, coercion, or harassment. This is a direct and egregious harm. The response should be to ensure that existing offences for blackmail, malicious communications (as updated by the Online Safety Act 2023), and grooming are interpreted or amended to explicitly cover this technologically novel method of abuse. Here, the harm is concrete, and the connection between the act and the victim is clear.
Finally, this global problem demands global solutions. The UK should work with international partners to establish norms around the development of AI, promote technological solutions like digital watermarking and robust content filtering, and foster cross-border law enforcement cooperation. A purely domestic criminal law response is destined to be inadequate.
Conclusion
The question of whether the criminal law should use preventive offences to tackle AI-generated CSAM forces a confrontation with the fundamental principles that restrain state power. The impulse to "do something" in the face of such a repellent technological capability is understandable. However, this essay has argued that resorting to broad, stand-alone preventive offences would be a grave error. It would involve a trade-off that is inimical to the traditions of UK criminal justice: sacrificing legal certainty, proportionality, and the principle of autonomy for a speculative and likely marginal gain in pre-emption. Criminalising the possession of general-purpose tools or the ambiguous act of preparing to create a synthetic image would be an overreach, pushing the law into the perilous domain of policing thought and intent far removed from consummated harm.
The law is not powerless. The existing substantive offences for making and possessing indecent pseudo-photographs are fit for purpose and should be enforced with vigour. The most effective route for prevention lies not in expanding individual criminal liability, but in strengthening the regulatory obligations on the technology platforms that form the primary vectors of dissemination, as pioneered by the Online Safety Act 2023. This, combined with adapting existing laws to cover the malicious use of AI in crimes like blackmail and grooming, and fostering international and technological cooperation, provides a robust, principled, and proportionate response. The criminal law must remain a shield against demonstrable harm, not a sword against digital possibility.
References
Ashworth, A. (2009) Principles of Criminal Law. 6th edn. Oxford: Oxford University Press.
Ashworth, A. and Zedner, L. (2014) Preventive Justice. Oxford: Oxford University Press.
Criminal Justice Act 1988, c. 33.
Gullefer [1990] 1 WLR 1063.
Mill, J.S. (1859) On Liberty. London: John W. Parker and Son.
Online Safety Act 2023, c. 50.
Protection of Children Act 1978, c. 37.
R v Fell [2011] EWCA Crim 3014.
R v Saik [2006] UKHL 18.
Simester, A.P. and Sullivan, G.R. (2019) Simester and Sullivan's Criminal Law: Theory and Doctrine. 7th edn. Oxford: Hart Publishing.
Terrorism Act 2000, c. 11.
Zedner, L. (2007) 'Pre-crime and post-criminology?', Theoretical Criminology, 11(2), pp. 261-281.
