AI-generated child sexual abuse material: what’s the harm?
Open accessThe DOI matches its Crossref record; 32 of the 78 references it cites were found in a registry or an index, and 43 are not checked yet; 9 of the 91 sources its claims cite were checked for support.
Abstract
Examines the harms of AI-generated synthetic child sexual abuse material (AI CSAM) and how generative AI is implicated in the facilitation of grooming, coercion, sexual extortion (sextortion), and lowering offending barriers.
Verified 2026-11-19 · 43 references not checked yet
Overview
CoChat Summary
TL;DR
This narrative review argues that AI-generated child sexual abuse material (AI CSAM) is not a victimless or harm-reducing alternative, identifying seven domains of harm spanning revictimization, coercion, normalization, gateway offending, youth peer exploitation, detection impairment, and commercialization. It cautions that appeals to harmlessness obscure real risks and may contribute to inertia in child-protection, technology, and policy responses.
Key findings
- AI CSAM is growing and spreading: the IWF identified over 20,000 suspected AI-generated explicit images on a single dark web forum in one month, 27% of which were classified as illegal under UK law; 2024 saw a 380% increase in IWF reports containing actionable AI CSAM and a 22% six-month increase in engagement with AI-specific forum threads.
- The paper identifies seven harms: depicting real children, coercion/grooming/sexual extortion, normalization/desensitization, gateway to offending, youth access and peer exploitation, impaired protection/detection capacity, and incentivized production/profit-driven exploitation. In financial sexual extortion reports with identifiable tactics, 11% involved threats using fabricated sexual imagery of the child; 56% of surveyed law enforcement professionals had encountered AI “nudify” tools used on minors; one in ten minors surveyed reported knowing peers who used generative AI to create explicit images of other children.
- Public and victim perceptions matter: 84% of teens and young adults recognized deepfake nude images as causing tangible psychological, emotional, and reputational harm, while 40% of UK adults were unsure whether AI CSAM was legal or believed it to be legal.
- The paper rejects harm-reduction claims, arguing AI CSAM is scalable, customizable, can be more extreme than underpinning CSAM, and may erode protective factors; it concludes these risks substantially outweigh speculative benefits, though carefully controlled synthetic material distinct from AI CSAM could be explored under strict legal and ethical boundaries.
Methods
- Narrative review and conceptual synthesis, not a systematic review, drawing on technical literature, psychological and criminological research, civil society and law enforcement reporting, news media, and sources predating diffusion models.
- The paper also functions as a primer: it defines AI CSAM, describes key generative technologies (GANs, diffusion models, fine-tuning, jailbreaking, hybrid workflows), situates AI CSAM within the Motivation-Facilitation Model and Lawless Space Theory, and examines known and potential harms.
- No datasets were generated or analyzed; the work is a synthesis rather than a primary empirical study.
Limitations
- Stated: the review takes a narrative rather than systematic approach because AI CSAM is an emerging phenomenon and relevant sources are heterogeneous and distributed across industry, civil society, law enforcement, news media, and scientific papers.
- Stated: as an emerging threat, it is not yet clear whether AI CSAM will be associated with distinct patterns of trauma among victims aware of their victimization; research on AI CSAM is still emerging, and evidence linking harm-minimizing beliefs to offending risk is scant.
- Evident from the text: there are no primary data or systematic search procedures, so proposed harm pathways, escalation risks, and causal claims remain inferential and are not empirically tested in this paper.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 3 Not valid · 59 Not checked · 78 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Their ability to produce photorealistic imagery with fine-grained control has made them the dominant tools for generating synthetic content (Yang et al. 2023), including AI CSAM (Thorn 2024a).”
What the sources say
- Not matchedcites Thorn 2024a
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites Yang et al. 2023 source
The Introduction says diffusion models have broken GANs’ long-time dominance in image synthesis. Section 7.4.1 describes photorealistic samples and fine-grained control of synthesis results. The separate claim about AI CSAM is cited to Thorn, not Yang. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Safety features that are more difficult to circumvent (such as ensuring the model is trained on properly curated data) can still be undone with adversarial fine-tuning or other methods (O'Brien et al. 2025).”
What the sources say
- Partly supportedcites O'Brien et al. 2025 sourceExperimental — accuracy not yet measured
The Introduction (p. 5) says none of the tested defenses resisted staged attacks combining fine-tuning with in-context retrieval, supporting the point that curated-data safeguards are not unbreakable. But Section 3.2 (p. 12) reports that the authors’ filtered models resisted adversarial fine-tuning alone for up to 10,000 steps and 300M tokens; the claim overstates the evidence if it means fine-tuning alone undid those safeguards. (LLM, conf=0.83)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Victims commonly experience trauma symptoms including shame, helplessness, anxiety, suicidal ideation, and difficulties trusting others (Ray and Henry 2024; Wolbers et al. 2025).”
What the sources say
- Partly supportedcites Ray and Henry 2024 sourceExperimental — accuracy not yet measured
The Introduction and Impacts and Harms section report shame, anxiety, and suicidal ideation among sextortion victims, but describe suicidal ideation as occurring in some cases rather than commonly. The Reporting and Help-Seeking section describes distrust of adults and professionals, not a general difficulty trusting others. I found no passage describing helplessness. (LLM, conf=0.88)
- Could not verifycites Wolbers et al. 2025 sourceExperimental — accuracy not yet measured
Our copy of 10.52922/ti77789 is not a systematic review in text form: it is a short ACCCE/ThinkUKnow 'sexual extortion response kit' resource page (pp. 1-2) covering prevalence facts, social-media post examples, who the kit is for, and where to report. No passage addresses trauma symptoms, shame, helplessness, anxiety, suicidal ideation or difficulties trusting others; searches on those terms plus 'psychological impacts', 'mental health', 'self-harm', 'depression' and 'trust' returned only the same awareness/reporting sections. The nearest content is (Facts, p. 1) 'Victims fear they will get into trouble if they report', which concerns reporting reluctance rather than the symptom list the claim attributes. The copy is flagged as not truncated, so the claim's point is simply absent from what we hold. (LLM, conf=0.72)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“More recent work indicates that, among users of CSAM, higher frequency of use and exposure to more extreme material are associated with increased likelihood of seeking direct contact with children, particularly in online spaces (Insoll et al. 2022).”
What the sources say
- Supportedcites Insoll et al. 2022 source
Section 5.2 reports that frequency of CSAM use was the best predictor of seeking contact with children online, and that seeking material depicting infants and toddlers was also associated with that outcome. Section 5.4 likewise reports a greater self-reported likelihood of contacting children after viewing CSAM among higher-frequency users. These are associations in an anonymous survey, as the claim indicates. (LLM, conf=0.93)
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, established psychological harms linked to the dissemination of CSAM-including symptoms of post-traumatic stress, shame, anxiety, and self-blame-are almost certain to persist when existing abusive material is multiplied, altered, or rendered more violent through the use of AI tools (Chauviré-Geib and Fegert 2024).”
What the sources say
- Partly supportedcites Chauviré-Geib and Fegert 2024 sourceExperimental — accuracy not yet measured
The Introduction and ‘Consequences and Impact of the Abuse’ describe ongoing victimization when abusive material remains online, and report post-traumatic stress, shame, anxiety, and self-blame among victims. The review does not address AI multiplication or alteration of existing material, or establish that these harms are ‘almost certain’ to persist in those circumstances. Page numbers were not available in our copy. (LLM, conf=0.88)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
75 more claims
Create a free account to see all 148 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 38
- Probably not a substantial claim. 37
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Generative tools allow offenders to tailor CG-CSAM to specific preferences, producing material that could not be obtained through contact offending alone, including content tailored to more extreme or violent preferences (IWF 2024; Ó Ciardha et al. 2026).”
From Generative AI in child sexual exploitation and abuse: views from UK law enforcement (2026) · cites this paper as Ó Ciardha et al. 2026
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Generative AI may intensify these harms by making it easier to alter, remaster, extend or repackage existing material, with the consequence that the distinction between 'real' and 'synthetic' content does not capture the full harm: a generated or manipulated image can still be linked to real children, real survivors and real cycles of re-victimisation (Ó Ciardha et al. 2026).”
From Generative AI in child sexual exploitation and abuse: views from UK law enforcement (2026) · cites this paper as Ó Ciardha et al. 2026
The cited source says what the claim attributes to it.
In the citing paper
“This also extends to recent work by Ó Ciardha et al. (2026), who argue that generative CG-CSAM should not be treated only as material passively consumed by offenders, but as something weaponised in grooming, coercion and sexual extortion.”
From Generative AI in child sexual exploitation and abuse: views from UK law enforcement (2026) · cites this paper as Ó Ciardha et al. (2026)
Citations
Cites
Diffusion Models: A Comprehensive Survey of Methods and Applications
Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., & Yang, M.-H. (2024) · ACM Computing Surveys · doi:10.1145/3626235 · Cited by 1,704 · cited 1×
“Their ability to produce photorealistic imagery with fine-grained control has made them the dominant tools for generating synthetic content (Yang et al. 2023), including AI CSAM (Thorn 2024a).”
The Motivation-Facilitation Model of Sexual Offending
Seto, M. C. (2019) · Sexual Abuse · doi:10.1177/1079063217720919 · Cited by 339 · cited 8×
“Research suggests that some individuals who engage in compulsive pornography use or hypersexual behavior experience tolerance effects, where they seek out increasingly novel or taboo material (Seto 2019).”
InfluentialTypology of Paedophile Picture Collections
Taylor, M., Holland, G., & Quayle, E. (2001) · The Police Journal: Theory, Practice and Principles · doi:10.1177/0032258x0107400202 · Cited by 190 · cited 2×
“Examples shared by NCMEC (2024) demonstrate real prompts seeking to generate material that would correspond to the highest levels of severity on the COPINE rating scale of sexual abuse imagery (Taylor et al. 2001).”
Cyber Sextortion: An Exploratory Analysis of Different Perpetrators Engaging in a Similar Crime
O’Malley, R. L., & Holt, K. M. (2020) · Journal of Interpersonal Violence · doi:10.1177/0886260520909186 · Cited by 158 · cited 1×
“…offenders are characterized by a variety of motivations, including individuals with a sexual interest in children as well as financially-motivated individuals who may engage in the targeting of young people as part of a wider range of cybercrime and scam behaviors (Liggett O’Malley and Holt 2022).”
Understanding online child pornography use: Applying sexual offense theory to internet offenders
Elliott, I. A., & Beech, A. R. (2009) · Aggression and Violent Behavior · doi:10.1016/j.avb.2009.03.002 · Cited by 150 · cited 1×
“Elliott and Beech (2009) describe how online environments provide access to social validation, justification, and reinforcement for CSAM use, which can diminish psychological and social barriers to further offending.”
Klein, V., Schmidt, A. F., Turner, D., & Briken, P. (2015) · PLOS ONE · doi:10.1371/journal.pone.0129730 · Cited by 109 · cited 1×
“While sexual interest in children is a key motivational factor, it is not the only motivating factor for CSAM use (see Klein et al. 2015, who report on the relationship between sex drive and CSAM use).”
Pornography and Sex Crimes in the Czech Republic
Diamond, M., Jozífková, E., & Weiss, P. (2010) · Archives of Sexual Behavior · doi:10.1007/s10508-010-9696-y · Cited by 108 · cited 1×
“…studies have suggested that increased access to pornography including CSAM may correlate with lower rates of sexual offending including child abuse (e.g., Diamond et al. 2011), interpretation of these data is challenging given the use of natural—rather than controlled—experiments for evidence.”
Ray, A., & Henry, N. (2025) · Trauma, Violence, & Abuse · doi:10.1177/15248380241277271 · Cited by 76 · cited 2×
“Victims commonly experience trauma symptoms including shame, helplessness, anxiety, suicidal ideation, and difficulties trusting others (Ray and Henry 2024; Wolbers et al. 2025).”
The implicit theories of child sexual exploitation material users: An initial conceptualization
Bartels, R. M., & Merdian, H. L. (2016) · Aggression and Violent Behavior · doi:10.1016/j.avb.2015.11.002 · Cited by 62 · cited 2×
“Research on CSAM offenders suggests that many justify their behavior not by overtly endorsing abuse but by minimizing harm (Bartels and Merdian 2016).”
Diffusion Explainer: Visual Explanation for Text-to-image Stable Diffusion
Lee, S., Hoover, B., Strobelt, H., Wang, Z. J., Peng, S., Wright, A., Li, K., Park, H., Yang, H., & Chau, D. H. (2024) · Visualization · doi:10.1109/vis55277.2024.00027 · Cited by 57 · cited 1×
“An interactive explainer of this process can be found here: https:// poloc lub. github. io/ diffu sion-expla iner/ (Lee et al. 2024).”
Cited by
A Marketplace for AI-Generated Adult Content and Deepfakes
Ghosh, S., DeVerna, M., & Menczer, F. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3812321 · Cited by 4
“Generative AI has been used to create political disinformation, identity fraud, child sexual abuse material, and non-consensual intimate imagery, often at low cost and high scale [9, 24, 26, 11, 6].”
Gold Standard or Gold-Plated? Human Practices of Triple Verification in CSAM Takedown
Rottier, M., van Eeten, M. J. G., & Zannettou, S. (2026) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3791039 · Cited by 1
Generative AI in child sexual exploitation and abuse: views from UK law enforcement
Demeocq, C., Taylor, A., Ross, B., & McFeeters, A. (2026) · AI & SOCIETY · doi:10.1007/s00146-026-03176-6 · Cited by 1
Mistareehi, H., Dronavalli, M., & Yarali, A. (2026) · AI and Ethics · doi:10.1007/s43681-026-01305-x
Platform-Facilitated Grooming and AI Chatbots: Rethinking Criminal Liability and Regulation
Chawki, M. (2026) · Laws · doi:10.3390/laws15040093 · Cited by 1
“…may facilitate AI-assisted grooming by enabling realistic identity deception, impersonation, and the creation of synthetic child sexual abuse material (CSAM), thereby increasing opportunities for manipulation and exploitation (Gambín et al. 2024; Dushi et al. 2025; Ciardha et al. 2026).”
Metadata
- Authors
- Ó Ciardha, C., Buckley, J., & Portnoff, R. S.
- Affiliations
- University of Kent (GB)
- Venue
- AI & SOCIETY
- Published
- 2026-04-06
- Volume
- 41
- Issue
- 9
- Pages
- 8597-8610
- ISSN
- 0951-5666 (ISSN-L)
- 1435-5655
- OpenAlex
- W4416370986
- Cited by
- 9
- References
- 72
- Citations by year
- 2026: 9
- FWCI
- 74.83
- Type
- journal article
- Language
- en
- Update notices
- correction (2026-08-07)
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- AI & Society (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Funders
- Tech Coalition Safe Online Research Fund (23-EVAC-0015.2-University of Kent)
- Topics
- Social Sciences › Law › Cybercrime and Law Enforcement Studies
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Law › Legal and Social Justice Studies
- Keywords
- generative AI, child sexual exploitation, child protection, online grooming, desensitization, harm reduction, desistance from crime, law enforcement
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: AI-generated child sexual abuse material: what’s the harm? https://app.cochat.ai/papers/10.1007/s00146-026-02932-y
- The paper
- Ó Ciardha, C., Buckley, J., & Portnoff, R. S. (2026). AI-generated child sexual abuse material: what’s the harm? AI & SOCIETY. https://doi.org/10.1007/s00146-026-02932-y
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.