Every so often, someone tells me that AI-generated child sexual abuse material is a lesser problem. No child was in the room, they say. No camera, no abuser, no victim. Perhaps, some argue, synthetic material could even act as a safety valve.
The evidence says otherwise. And it points to a group with more power than anyone to change the outcome: the platforms where children spend their lives online.
The pathway nobody wants to talk about Most conversations about child sexual abuse material (CSAM) start with the offender as a fully formed adult. The research tells a more uncomfortable story. Many people who go on to seek out CSAM first encountered it when they were children themselves, and often without looking for it.
Protect Children, the Finnish NGO that surveys people while they are actively searching for CSAM on the dark web, found in its first ReDirection report (8,484 respondents) that 70% first saw CSAM before the age of 18, nearly 40% were 13 or younger, and over half said their first exposure was accidental (Protect Children, 2021). A larger analysis of 11,470 responses found the same pattern: 65.3% first saw the material under 18, 36.7% at 13 or younger, and 50.5% first saw it by accident (Nurmi et al., 2024).
This is not only a dark web finding. In an anonymous survey of 5,512 internet users in the community, 71.2% of those who had viewed CSAM first discovered it before they were 18, most discovered it by accident, and around 45% of those whose first exposure was accidental went on to view it again intentionally (Napier, 2022). Exposure to adult pornography before the age of 14 was a significant predictor of ever viewing CSAM (Napier et al., 2024). Napier's work also suggests that more adolescents are discovering CSAM now than in previous generations.
The harm to the child who stumbles across this material is real in its own right. A 2026 study of 5,642 people who had searched for CSAM found that earlier exposure was significantly linked to greater difficulty coping with stress and greater social withdrawal, and half of those first exposed under 18 described the exposure as accidental (Dieseth et al., 2026). Protect Children's executive director has described exposure under 13 as an adverse childhood experience with potentially far-reaching consequences for wellbeing and development (Yle, 2021).
The most recent and largest study, sponsored by Ofcom and based on more than 20,000 responses, reports that 59% of respondents had seen CSAM by the age of 18 and that many first encountered it accidentally (Protect Children and Ofcom, 2026). Respondents also reported that harmful material was often pushed to them through algorithmic recommendations (MLex, 2026).
These studies rely on self-report, and the dark web samples are self-selected. Most children who see harmful material will never offend, and sexual interest in children remains a strong predictor of continued use (Napier et al., 2024). But a rapid evidence review by the Australian Institute of Criminology identifies accidental first exposure and the online environment as situational factors in the onset of CSAM offending, and concludes that technological interventions matter (Wrightson-Hester et al., 2026). That matters enormously, because situational factors are the ones we can design out.
If children were not stumbling across this material, one of the most consistently documented on-ramps into sexual offending would narrow dramatically. Prevention does not need to wait until someone has become an offender. It can start by making sure a 12-year-old never sees the material in the first place.
What AI changes Generative AI is pouring fuel onto this pathway. In 2025 the Internet Watch Foundation (IWF) assessed 8,029 AI-generated images and videos as showing realistic child sexual abuse. It found 3,443 AI-generated abuse videos, compared with 13 in 2024, and 65% of those videos were Category A, the most severe classification under UK law, compared with 43% of videos involving real children (IWF, 2026a). Girls account for 97% of the AI-generated images, and the fastest growth is in depictions of children aged three to six (IWF, 2026c).
This material is not confined to hidden corners. The IWF reports AI CSAM on both the dark web and the clear web, including mainstream commercial platforms, and has found it on AI chatbot websites (IWF, 2026b). Children are also meeting AI sexual imagery among their own peers: 13% of UK teenagers, the equivalent of around half a million, have had an experience with a nude deepfake (Internet Matters, 2024), and 1 in 17 US teens report having deepfake nudes made of them (Thorn, 2025).
So AI does not only create more material. It creates more material in the places where children already are, and it lowers the barrier to producing it to almost nothing.
"But no child was harmed" This is the claim that does the most damage, and it does not survive contact with the evidence. AI CSAM harms children in at least three ways.
First, it harms every child who is exposed to it. A child does not experience an image differently because software rather than a camera produced it. The developmental harms described above follow exposure, not provenance. When AI material circulates on platforms that children use, every one of those children becomes a potential first exposure.
Second, it harms real children directly. AI models have been trained on real abuse imagery: Stanford researchers identified 3,226 suspected CSAM entries in LAION-5B, a dataset used to train popular image generators (Thiel, 2023). The IWF found a dark web forum hosting 202 models, including models built to generate the likenesses of identified abuse victims, child actors and influencers. One offender shared fine-tuned models for 128 different named victims, and a realistic model of a specific child can be built from as few as 20 to 30 images in as little as 15 minutes (IWF, 2026b). Survivors are being abused again, years later, in new and sometimes more extreme scenarios. AI is also being used in sexual extortion: 21% of reports to the UK Report Remove service involving nude or sexual imagery of young people involved digitally altered or manipulated content (IWF, 2026b).
Third, it feeds the demand that harms children. AI CSAM is shared alongside traditional CSAM, traded, advertised and sold. In the Ofcom-sponsored study, 35% of respondents reported viewing or creating AI CSAM, and half of those were involved in commissioning or producing it for profit (Protect Children and Ofcom, 2026). Researchers caution that AI CSAM may act as a new or enhanced pathway into offending by lowering barriers, desensitising users to progressively extreme content and undermining protective factors, and they warn against claims that it can function as harm reduction (Ó Ciardha, Buckley and Portnoff, 2025). The IWF reaches the same conclusion: synthetic material reinforces sexual interest in children rather than providing a safe substitute, and may increase the risk of escalation (IWF, 2026b).
That escalation is not hypothetical. Around 40% of CSAM users surveyed by Protect Children said they had sought contact with a child after viewing CSAM, and 70% of those tried to make contact online, most often through social media and online games (Insoll et al., 2024).
Platforms are the capable guardian Crime occurs when a motivated offender and a suitable target converge in the absence of a capable guardian (Cohen and Felson, 1979). Online, the convergence happens on platforms. Children are there. People seeking to harm them are there. The material is there. The only actor present at every one of those points, with the visibility, the data and the engineering capability to intervene, is the platform itself.
Parents cannot see inside a recommendation algorithm. Police cannot moderate billions of uploads. Children cannot be expected to protect themselves from material they did not ask to see. Platforms can.
And we know guardianship works. When the reThink chatbot and warning message were deployed on Pornhub in the UK, searches for CSAM fell significantly, and 82% of users who triggered the warning did not appear to search for such material again (IWF, 2024). In the Ofcom study, a third of respondents said CSAM had become harder to access in the past five years, and many who recalled seeing a warning message said it prompted them to reflect on or change their behaviour (Protect Children and Ofcom, 2026). Perpetrators also reported avoiding services with strict age verification and other barriers (Captain Compliance, 2026). Friction works. Design works.
The law is moving in the same direction. The UK's Crime and Policing Act 2026 criminalises adapting, possessing or supplying models optimised to create CSAM, extends the "paedophile manuals" offence to AI-generated material, and creates an offence of administering or moderating services that facilitate CSAM (Home Office, 2026). But law is a backstop. It acts after the harm. Platforms act before it.
A call to work together No single platform can solve this alone, because offenders do not stay on a single platform. They find material in one place, discuss it in another, and contact children in a third. Guardianship has to be as connected as the harm.
I am calling on platforms, AI developers and the wider technology sector to:
Prevent the material being made. Build safety into generative models from the start, remove abuse imagery from training data, red-team for misuse, and refuse to host or distribute models optimised to create CSAM. Prevent it reaching children. Detect known and novel CSAM, including AI-generated material, at upload. Stop recommendation systems surfacing sexual content to young users, and use robust, privacy-respecting age assurance. Interrupt the pathway. Deploy evidence-based warning messages and signpost people to help services such as Stop It Now at the moment of risk, not after arrest. Share what you know. Share hashes, signals and emerging threat intelligence across companies and with hotlines such as the IWF, because a model removed from one service should not reappear on another. Open your systems to independent evaluation. The reThink evaluation shows what transparency can teach us. We need more of it.
The children who stumble across abuse material today should not become tomorrow's case files, either as victims or as offenders. We have the evidence to show where the pathway begins. We have evidence that intervention works. What we need now is the collective will to act as the capable guardian children deserve.
There is no such thing as a victimless image. Let's build an internet that proves we understand that.
References Captain Compliance (2026). Ofcom releases groundbreaking global study revealing how perpetrators access and share child sexual abuse material online, 26 March. https://captaincompliance.com/education/ofcom-releases-groundbreaking-global-study-revealing-how-perpetrators-access-and-share-child-sexual-abuse-material-online/
Cohen, L. E. and Felson, M. (1979). Social change and crime rate trends: A routine activity approach. American Sociological Review, 44(4), 588-608. https://doi.org/10.2307/2094589
Dieseth, T., Insoll, T., Smith, L., Vaaranen-Valkonen, N. and Lundrigan, S. (2026). Early exposure to child sexual abuse material: Prevalence, circumstances and psychosocial functioning. Child Abuse & Neglect, 176, 108070. https://pubmed.ncbi.nlm.nih.gov/42033961/
Home Office (2026). Crime and Policing Act 2026: Explanatory Notes. https://www.legislation.gov.uk/ukpga/2026/20/notes/division/2/index.htm
Insoll, T., Soloveva, V., Díaz Bethencourt, E., Ovaska, A. and Vaaranen-Valkonen, N. (2024). Tech Platforms Used by Online Child Sexual Abuse Offenders. Protect Children. https://www.kinderrechte.digital/fileadmin/kinderrechte.digital/Hintergrundmaterial/MD1795.pdf
Internet Matters (2024). The new face of digital abuse: Children's experiences of nude deepfakes. https://www.internetmatters.org/wp-content/uploads/2024/11/Childrens-experiences-of-nude-depfakes-research.pdf
Internet Watch Foundation (2024). Pioneering chatbot reduces searches for illegal sexual images of children, 29 February. https://www.iwf.org.uk/news-media/news/pioneering-chatbot-reduces-searches-for-illegal-sexual-images-of-children/
Internet Watch Foundation (2026a). Annual Data and Insights Report 2025: Executive Summary. https://www.iwf.org.uk/annual-data-insights-report-2025/executive-summary
Internet Watch Foundation (2026b). Harm without limits: AI child sexual abuse material through the eyes of our Analysts. https://www.iwf.org.uk/media/hl1nvdti/iwf-ai-csam-report-2026.pdf
Internet Watch Foundation (2026c). Annual Data and Insights Report 2025: AI-generated images. https://www.iwf.org.uk/annual-data-insights-report-2025/image-insights/ai-generated-images
MLex (2026). Search engines, porn sites central to CSAM access, UK study finds. https://www.mlex.com/mlex/articles/2454384/search-engines-porn-sites-central-to-csam-access-uk-study-finds
Napier, S. S. (2022). Age and circumstances of onset to child sexual abuse material viewing and characteristics of consumers. PhD thesis, University of Sydney. https://ses.library.usyd.edu.au/handle/2123/29961
Napier, S. S., Seto, M. C., Cashmore, J. and Shackel, R. (2024). Characteristics that predict exposure to and subsequent intentional viewing of child sexual abuse material among a community sample of Internet users. Child Abuse & Neglect, 156, 106977. https://pubmed.ncbi.nlm.nih.gov/39217741/
Nurmi, J., Paju, A., Brumley, B. B., Insoll, T., Ovaska, A. K., Soloveva, V., Vaaranen-Valkonen, N., Aaltonen, M. and Arroyo, D. (2024). Investigating child sexual abuse material availability, searches, and users on the anonymous Tor network for a public health intervention strategy. Scientific Reports, 14, 7849. https://pmc.ncbi.nlm.nih.gov/articles/PMC10991312/
Ó Ciardha, C., Buckley, J. and Portnoff, R. S. (2025). AI Generated Child Sexual Abuse Material: What's the Harm? arXiv. https://arxiv.org/abs/2510.02978
Protect Children (2021). CSAM Users in the Dark Web: Protecting Children Through Prevention (ReDirection Survey Report). https://www.kinderrechte.digital/fileadmin/kinderrechte.digital/Hintergrundmaterial/MD1777.pdf
Protect Children and Ofcom (2026). CSAM Perpetrator Research Report: Findings from a Survey of CSAM Perpetrators on Digital Platform Use and Design (Tell Me More About Tech), 18 March. https://www.suojellaanlapsia.fi/en/post/tmat-csam-perpetrator-research-report
Thiel, D. (2023). Identifying and Eliminating CSAM in Generative ML Training Data and Models. Stanford Internet Observatory. https://purl.stanford.edu/kh752sm9123
Thorn (2025). Deepfake Nudes and Young People, 3 March. https://www.thorn.org/blog/deepfake-nudes-are-a-harmful-reality-for-youth-new-research-from-thorn/
Wrightson-Hester, A., McVey, L., Napier, S., Dowling, C. and Wolbers, H. (2026). Pathways to onset of child sexual abuse material offending: A rapid evidence review. Trends and issues in crime and criminal justice no. 734. Australian Institute of Criminology. https://www.aic.gov.au/publications/tandi/tandi734
Yle (2021). Interview with Nina Vaaranen-Valkonen, executive director of Protect Children, 27 September. https://yle.fi/a/3-12116829