Generative AI in child sexual exploitation and abuse: views from UK law enforcement
Open accessThe DOI matches its Crossref record; 55 of the 55 references it cites were found in a registry or an index; 11 of the 48 sources its claims cite were checked for support.
Abstract
Empirical insights from semi-structured interviews with seven UK law enforcement practitioners are used to expand the scope of AI safety evaluation methods and propose changes to regulatory guidance on generative AI and public safety, specifically regarding multi-turn AI tools facilitating grooming, lowering barriers in image generation, and interaction-centred misuse patterns.
Overview
CoChat Summary
TL;DR
This paper reports semi-structured interviews with seven UK law enforcement practitioners on how generative AI facilitates child sexual exploitation and abuse (CSEA), identifying misuse patterns, a “testing gap” in AI safety evaluation, and structural barriers to enforcement. The authors argue that safety testing must move beyond single-prompt evaluations to multi-turn, interaction-centred, practitioner-informed methods.
Key findings
- Participants described four recurring patterns of AI-facilitated CSEA: AI-assisted grooming and victim targeting, synthetic and manipulated image-based abuse, system-of-systems misuse (e.g., LLMs used with image generators), and locally run or offline model use.
- The paper identifies a “testing gap”: current AI safety evaluations overlook real-world offender practices such as multi-turn grooming, prompt chaining, and evasion; one participant said firms “clearly hadn’t tested the software in a way that an offender would.”
- The paper cites an Internet Watch Foundation report of a 380% increase in reports containing AI-generated CSAM in 2024.
- Participants called for cross-institutional and international changes, including data-sharing frameworks, legally enabled evaluation environments, and broader regulatory definitions covering interactional and facilitative harms, not only generated outputs.
Methods
- Semi-structured interviews with seven UK-based law enforcement professionals actively engaged in investigating and prosecuting CSEA; recruitment used professional networks and a LinkedIn advertisement from February 2025 to early April 2025.
- Interviews lasted 45–70 minutes, took place via Microsoft Teams, were audio recorded and transcribed, then anonymised; six pre-designed questions guided discussion across four interview parts.
- Analysis used an abductive approach with an emic orientation, involving multiple transcript readings, group data sessions, and whiteboard mapping; an external expert panel from IWF, INHOPE, Terre des Hommes, and Ofcom reviewed the manuscript.
Limitations
- The authors state the sample of seven UK-based practitioners is small and not representative of all law enforcement perspectives within the UK or internationally, and that findings are shaped by UK legal, institutional, and technological conditions; they describe the work as agenda-setting rather than exhaustive.
- The authors note the small cohort reflects the rarity of the expertise sought and the risk to officers, many of whom work covertly; they deliberately avoid listing participant details to reduce identification risk.
- Evident from the text: recruitment through professional networks and LinkedIn, voluntary participation without compensation, and reliance on practitioner accounts may introduce self-selection and perception-based limitations.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Valid · 3 Not valid · 16 Not checked · 104 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“This issue remains under-addressed in the literature.”
- Also stated in: § 1 Introduction · p. 2 · plain
“AI safety work has largely focussed on domains such as dis/misinformation, fraud, hate speech, and online abuse (e.g. Danry et al. 2025; Guo 2024; Lee et al. 2024), while CSEA remains comparatively underexplored.”
What the sources say
- Not checked — no full text of the sourcecites Guo 2024 source
- Not checked — no full text of the sourcecites Danry et al. 2025 source
- Could not verifycites Lee et al. 2024 sourceExperimental — accuracy not yet measured
The Exposure section (p. 16) discusses AI-generated nonconsensual pornography, but does not address child sexual exploitation and abuse or whether research on it is under-addressed. Searches for CSEA, child sexual abuse material, and a related literature gap found no passage making that point. (LLM, conf=0.73)
plain- 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.
- Also stated in: § 1 Introduction · p. 2 · plain
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“little information is publicly available on how law enforcement and its work have been impacted by this criminality and how to integrate the insights gained from their experiences into safety measures”
What the sources say
- Not matchedcites Europol 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Could not verifycites Lin et al. 2023 sourceExperimental — accuracy not yet measured
Our copy of 'Towards Healthy AI: Large Language Models Need Therapists Too' is a perspective piece on using psychotherapy-style agents (Chatbot, User, Therapist, Critic) to mitigate toxic LLM behaviour; no passage mentions law enforcement, police, criminality, or integrating officers' experiences into safety measures. The closest it comes is the Introduction (p. 1), which lists 'malicious uses' among ethical and social harms but credits that taxonomy to Weidinger et al., 2022 and says nothing about law-enforcement impact. Searches for law enforcement/police/crime/safety-measure integration returned only passages on RLHF, red teaming, Constitutional AI and the alignment problem, so the claim cannot be checked against this source. (LLM, conf=0.70)
- Not matchedcites Miró-Llinares et al. 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- 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.
- Rests on an assumption not shown here.
- Not matched
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Prior work has shown that CSAM harms are not limited to the moment of original abuse but continue through circulation, viewing, downloading, recontextualisation and re-exposure (Page et al. 2025; Salter et al. 2025; Schmidt et al. 2024).”
What the sources say
- Not checked yetcites Page et al. 2025 source
Checked before — its earlier verdict is held back until a new check runs.
- Not checked — no full text of the sourcecites Salter et al. 2025 source
- Partly supportedcites Schmidt et al. 2024 sourceExperimental — accuracy not yet measured
The Discussion subsection 'Mental Health Impact of NCSSI' explicitly imports the CSAM literature to make the ongoing-harm point: the 'global accessibility, permeance, and availability of the internet' as the most difficult element of CSAM (circulating images) and 'the risk for re-traumatization … due to the knowledge that many people have seen their image' (viewing/re-exposure); the Introduction similarly ties circulation to fear, loss of control and lasting distress. So the claim's core — harm extending beyond the original abuse through circulation, viewing and re-exposure — is in our copy, reported by this review from Quayle et al. 2012, Whittle et al. 2013 and Martin & Slane 2015. However, two of the mechanisms the claim enumerates, 'downloading' and 'recontextualisation', appear nowhere in our copy (searched: downloading, recontextualisation, repurposing of images, harm beyond the original abuse), and the review's own scope is nonconsensual sharing of youth-produced sexual images, with the CSAM material treated as a parallel body of literature rather than as its primary subject. (LLM, conf=0.75)
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.
- Not checked yet
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Across the interviews, participants raised numerous CSEA offences involving technology and AI.”
plain- 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.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“law enforcement practitioners express fears about the ways AI might be mutating criminality”
hedged- 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.
34 more claims
Create a free account to see all 132 claims and to verify the 11 references still open.
- It may state two separate claims as one. 49
- Probably not a substantial claim. 71
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C. Ethical oversight and controlled disclosure · p. 16
“Figure 1's analytic board is intentionally blurred, with an unredacted version held under restricted access, to preserve methodological transparency without disclosing operationally sensitive material.”
§ D. Interview board · p. 16
“Figure 2 illustrates the flow chart of Interview board (blurred for confidentiality and safety)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ D. Interview board · p. 17
“Data availability No datasets were generated or analysed during the current study.”
§ 3.2.1 Data analysis · p. 5
“Therefore, the analysis involved multiple readings of the transcripts by the authors, notes and annotations within the interview documents, and group data sessions where emerging interpretations were discussed and refined.”
Citations
Cites
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01042 · Cited by 20,302
Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., & Neubig, G. (2023) · ACM Computing Surveys · doi:10.1145/3560815 · Cited by 4,361
Abductive Analysis: Theorizing Qualitative Research
Tavory, I., & Timmermans, S. (2014) · doi:10.7208/chicago/9780226180458.001.0001 · Cited by 1,726
Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts
Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B., & Yang, Q. (2023) · Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3544548.3581388 · Cited by 1,274
Wu, T., Terry, M., & Cai, C. J. (2022) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3491102.3517582 · Cited by 483
A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications
Sahoo, P., Singh, A. K., Saha, S., Jain, V., Mondal, S., & Chadha, A. (2024) · arXiv · doi:10.48550/arxiv.2402.07927 · Cited by 288
child pornography and the internet: perpetuating a cycle of abuse
Quayle, E., & Taylor, M. (2002) · Deviant Behavior · doi:10.1080/01639620290086413 · Cited by 282
A review of online grooming: Characteristics and concerns
Whittle, H., Hamilton-Giachritsis, C., Beech, A., & Collings, G. (2013) · Aggression and Violent Behavior · doi:10.1016/j.avb.2012.09.003 · Cited by 242
Online Child Sexual Exploitation: Prevalence, Process, and Offender Characteristics
Kloess, J. A., Beech, A. R., & Harkins, L. (2014) · Trauma, Violence, & Abuse · doi:10.1177/1524838013511543 · Cited by 182
Zeng, Y., Lin, H., Zhang, J., Yang, D., Jia, R., & Shi, W. (2024) · arXiv · doi:10.48550/arxiv.2401.06373 · Cited by 174
Research works printed but not linked: 3
- Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, Neelakantan A, Shyam P, Sastry G, Askell A et al (2020) Language models are few-shot learners. Adv Neural Inf Process Syst 33:1877–1901
- Franks MA, Waldman AE (2018) Sex, lies, and videotape: deep fakes and free speech delusions. Md Law Rev 78:892
- ECPAT International (2025) Terminology guidelines for the protection of children from sexual exploitation and sexual abuse (Luxembourg guidelines), 2nd edn. https://ecpat.org/wp-content/uploads/2025/04/Second-Edition-Terminology-Guidelines-final.pdf. Accessed 7 May 2026
Metadata
- Authors
- Demeocq, C., Taylor, A., Ross, B., & McFeeters, A.
- Affiliations
- University of Edinburgh (GB)
- Corresponding author
- Demeocq, Cyndie
- Taylor, Alex
- Ross, Björn
- Venue
- AI & SOCIETY
- Published
- 2026-07-01
- Volume
- 41
- Issue
- 10
- Pages
- 9919-9937
- ISSN
- 0951-5666 (ISSN-L)
- 1435-5655
- OpenAlex
- W7166838882
- Semantic Scholar
- d73f73434086b4a0063a41953709596cb534a16c
- Cited by
- 1
- References
- 44
- Citations by year
- 2026: 1
- FWCI
- 5.64
- Type
- journal article
- Language
- en
- 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)
- Edinburgh Research Explorer (University of Edinburgh) (publishedVersion, cc-by)
- Funders
- UKRI AI Centre for Doctoral Training in Responsible and Trustworthy in-the-world Natural Language Processing (EP/Y030656/1)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Law › Digitalization, Law, and Regulation
- Social Sciences › Law › Cybercrime and Law Enforcement Studies
- Keywords
- generative AI, law enforcement perspectives, AI safety evaluation, image generation, AI regulation, responsible AI
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: Generative AI in child sexual exploitation and abuse: views from UK law enforcement. https://app.cochat.ai/papers/10.1007/s00146-026-03176-6
- The paper
- Demeocq, C., Taylor, A., Ross, B., & McFeeters, A. (2026). Generative AI in child sexual exploitation and abuse: views from UK law enforcement. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03176-6
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.

