A Survey of Generative AI for Detecting Pedophilia Crimes
Abstract
1. Introduction
- Automated CSAM Detection—AI algorithms can analyze images, videos, and text to flag both known and previously unseen instances of child sexual abuse material, including within encrypted environments.
- Online Predator Identification—By analyzing communication trends, linguistic markers, and digital behavior patterns, AI systems can help identify individuals engaged in grooming or exploitation attempts.
- Proactive Risk Assessment—Although still highly speculative and surrounded by profound ethical concerns, AI models could theoretically be used to identify behavioral markers associated with an increased risk of offending. However, any such application would require extensive ethical safeguards and governance to ensure that preventive actions are both responsible and justified.
- Identifying and analyzing the specific limitations of each individual work surveyed, presenting these findings systematically in our comparative analysis table.
- Providing a dedicated discussion of the collective technological and ethical risks associated with deploying Generative AI in such a sensitive context, including privacy concerns, algorithmic bias, and the potential for misuse.
- Proposing concrete and structured directions for future research, which are directly derived from the identified gaps and limitations in the current body of work.
- RQ: How can generative AI be effectively used to detect and prevent pedophilia-related crimes in digital environments?
2. Background
2.1. Online Grooming
- Target Selection: Predators use online platforms and social media to identify vulnerable children;
- Relationship Building: Predators build trust with the child, often pretending to be a peer or offering emotional support;
- Risk Assessment: Predators assess the child’s vulnerability and the level of parental supervision;
- Exclusivity: Predators attempt to isolate the child from their friends and family, fostering a sense of dependency;
- Sexualization: Predators introduce sexual topics into the conversation, gradually desensitizing the child to inappropriate content;
- Exploitation: The final stage involves attempts to meet the child in person or to coerce the child into producing explicit material.
2.2. From Traditional Machine Learning to Generative AI
2.3. Large Language Models
- Understand and generate human-like text: LLMs can participate in conversations, answer questions, summarize text, and even mimic different writing styles;
- Identify patterns and connections in language: This enables them to detect topics, sentiments, and potential warning signs in online exchanges;
- Learn and adapt: LLMs can be fine-tuned to specific datasets, such as conversations related to online grooming, to improve their accuracy in detecting suspicious behavior.
- While platforms like Perplexity AI, Google Gemini, and ChatGPT are primarily engineered for general-purpose information retrieval and dialogue generation, their underlying LLM architectures present valuable opportunities for specialized applications such as crime detection and child protection. Google Gemini, with its extensive context window and multimodal capabilities, enables the analysis of lengthy, nuanced conversations and supports the interpretation of diverse content formats—including text, images, and video—making it particularly effective in the detection of long-term grooming behavior and CSAM [6]. ChatGPT, known for its adaptability via fine-tuning, is well-suited for domain-specific implementations, such as detecting linguistic cues indicative of grooming or facilitating interactive simulation environments for proactive threat detection. Perplexity AI, leveraging real-time search integration and transparent source attribution, offers substantial potential for monitoring live online forums and chat spaces, where grooming often manifests. Collectively, these tools—when used with awareness of their distinct capabilities and limitations, can inform the development of more targeted and effective generative AI-driven safety systems.
- Regarding data retrieval, ChatGPT depends solely on its static training corpus, whereas both Gemini and Perplexity AI utilize dynamic search-based retrieval mechanisms. Gemini stands out with a significantly larger context window than the other two, allowing it to process and analyze complex documents in greater depth. Additionally, Gemini’s multimodal processing sets it apart from the primarily text-based functionalities of ChatGPT and Perplexity. Cost and accessibility also vary: ChatGPT offers a freemium model, Perplexity remains free, and Gemini’s pricing is tiered based on model version and access features.
2.4. Comparative Technical Analysis
2.4.1. Architectural and Performance Differences
- ChatGPT is built upon the GPT architecture, which has evolved through several versions [12]. Its foundation consists of a series of transformer encoder layers, each utilizing a multi-head self-attention mechanism and a Feedforward Neural Network (FNN) [13]. The model is further aligned with human preferences using Reinforcement Learning from Human Feedback (RLHF) to enhance safety and produce more helpful responses [12,13].
- Gemini is also based on a decoder-only Transformer architecture [12]. Its design includes specific modifications for efficient training and inference on Tensor Processing Units (TPUs) and employs multi-query attention [13]. A key architectural feature is the integration of Retrieval-Augmented Generation (RAG), which grounds its responses in retrieved information to improve factual accuracy [12]. Gemini is designed as a natively multimodal system, capable of processing a combination of text, images, audio, and video [12]. In terms of performance, it is noted for prioritizing computational efficiency, potentially outperforming ChatGPT in speed and energy consumption [12].
- LLaMA, released by Meta AI, also uses the Transformer architecture but with several technical differentiators [13]. It employs Root Mean Square Layer Normalization (RMSQLN) instead of traditional layer normalization, and it uses the Swish-Gated Linear Unit (SwiGLU) activation function [13]. For positional information, it utilizes a Rotary Position Embedding (RoPE) scheme [13]. Regarding its context window, the original LLaMA was trained with a 2 K token context length, which was extended to 4 K tokens for LLaMA2 [13]. The practical implications of its computational requirements are significant; while compact models like LLaMA 7B can be run on local machines, more extensive versions demand impractical processing times, in the order of minutes per answer [13].
2.4.2. Implementation and Fine-Tuning Challenges
3. Methodology
3.1. Data Source
3.2. Search Query
3.3. Inclusion Criteria
- Relevance to the topic: Only studies that directly addressed the use of AI to detect pedophilic crimes or analyze online conversations related to child sexual exploitation were included;
- Full-text availability: Studies were only included if their full text was available through Google Scholar or open-access repositories.
3.4. Exclusion Criteria
- Publication type: Materials such as books, book chapters, internal reports, theses/dissertations, citations, presentations, abstracts, and appendices were excluded from consideration;
- Language: Only studies published in English were included in the review.
3.5. Characterization of Selected Papers
4. Literature Review: AI Approaches for Detection
4.1. Identifying Grooming Behaviors and Risk Factors
4.2. Machine Learning Models
4.3. Deep Learning and Large Language Models
4.4. Comparative Analysis
5. Benefits of Generative AI in Detecting Pedophilia
5.1. Automating CSAM Detection
5.2. Spotting Online Predators
5.3. Predicting Potential Offenders
5.4. Comparative Efficiency of AI Methods Versus Traditional Techniques
6. Limitations and Ethical Risks of Using GenAI
6.1. Privacy Concerns
6.2. Algorithmic Bias
6.3. Erroneous Identifications
6.4. Misuse by Malicious Parties
6.5. Hallucination
6.6. Legal and Workflow Integration Challenges
7. Synthesis of Findings
7.1. Potential of Generative AI
7.2. Effective Detection and Prevention
7.3. Ethical Challenges and Mitigation Strategies
8. Gaps and Future Research on GenAI for Child Safety
8.1. Development of Dynamic and Adaptive Models
8.2. Cross-Platform Threat Intelligence
8.3. Enhance Contextual Understanding with Advanced NLP
8.4. Robust, Realistic, and Bias-Mitigated Datasets
8.5. Hybrid Human-AI Systems for Intervention
8.6. Advancing Explainable AI for Generative Models
9. Conclusions
Author Contributions
Funding
Data Availability Statement
Conflicts of Interest
Abbreviations
AI | Artificial Intelligence |
BoW | Bag-of-Words |
CSA | Child Sexual Abuse |
CSAM | Child Sexual Abuse Material |
FL | Federated Learning |
GenAI | General Artificial Intelligence |
GPT | Generative Pre-trained Transformer |
LLM | Large Language Model |
LSTM-RNN | Long Short-Term Memory Recurrent Neural Networks |
ML | Machine Learning |
NLP | Natural Language Processing |
OCR | Optical Character Recognition |
PCI | Predatory Conversation Identification |
SGS-V | Sexual Grooming Scale Victim Version |
SimCSE | Simple Contrastive Sentence Embedding |
SPI | Sexual Predator Identification |
SVM | Support Vector Machine |
VPD | Victim-Predator Differentiation |
References
- Marvasti, J.A. (Ed.) Psychiatric Treatment of Sexual Offenders: Treating the Past Traumas in Traumatizers. A Bio-Psycho-Social Perspective; Charles C Thomas Publisher: Springfield, IL, USA, 2004. [Google Scholar]
- Cook, D.; Zilka, M.; DeSandre, H.; Giles, S.; Weller, A.; Maskell, S. Can We Automate the Analysis of Online Child Sexual Exploitation Discourse? arXiv 2022, arXiv:2209.12320. [Google Scholar]
- Wolbers, H.; Cubitt, T.; Cahill, M.J. Artificial intelligence and child sexual abuse: A rapid evidence assessment. Trends Issues Crime Crim. Justice 2025, 711, 1–18. [Google Scholar] [CrossRef]
- Levy, I.; Robinson, C. Thoughts on child safety on commodity platforms. arXiv 2022, arXiv:2207.09506. [Google Scholar]
- UNICRI—United Nations Interregional Crime and Justice Research Institute. New! How AI Is Leading the Fight Against Online Child Abuse. 2023. Available online: https://unicri.org/News/AI-for-Safer-Children-%20article-Emerging-Europe (accessed on 6 May 2025).
- Puentes, J.; Castillo, A.; Osejo, W.; Calderón, Y.; Quintero, V.; Saldarriaga, L. Guarding the Guardians: Automated Analysis of Online Child Sexual Abuse. In Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), Paris, France, 2–6 October 2023; pp. 3730–3734. [Google Scholar] [CrossRef]
- Borj, P.R.; Raja, K.; Bours, P. Online grooming detection: A comprehensive survey of child exploitation in chat logs. Knowl.-Based Syst. 2023, 259, 110039. [Google Scholar] [CrossRef]
- Mou, J.; Duan, P.; Gao, L.; Liu, X.; Li, J. An effective hybrid collaborative algorithm for energy-efficient distributed permutation flow-shop inverse scheduling. Future Gener. Comput. Syst. 2022, 128, 521–537. [Google Scholar] [CrossRef]
- Jeglic, E.L.; Winters, G.M.; Johnson, B.N. Identification of red flag child sexual grooming behaviors. Child Abus. Negl. 2023, 136, 105998. [Google Scholar] [CrossRef] [PubMed]
- Rani, G.; Singh, J.; Khanna, A. Comparative Analysis of Generative AI Models. In Proceedings of the 2023 International Conference on Advances in Computation, Communication and Information Technology (ICAICCIT), Faridabad, India, 23–24 November 2023; pp. 760–765. [Google Scholar] [CrossRef]
- Prosser, E.; Edwards, M. Helpful or Harmful? Exploring the Efficacy of Large Language Models for Online Grooming Prevention. In Proceedings of the EICC 2024: European Interdisciplinary Cybersecurity Conference, Xanthi, Greece, 5–6 June 2024; pp. 1–10. [Google Scholar] [CrossRef]
- Rane, N.; Choudhary, S.; Rane, J. Gemini versus ChatGPT: Applications, performance, architecture, capabilities, and implementation. J. Appl. Artif. Intell. 2024, 5, 69–93. [Google Scholar] [CrossRef]
- Buscemi, A.; Proverbio, D. ChatGPT vs Gemini vs LLaMA on Multilingual Sentiment Analysis. arXiv 2024, arXiv:2402.01715. [Google Scholar]
- Pasca, P.; Signore, F.; Tralci, C.; Longo, M.; Preite, G.; Ciavolino, E. Detecting online grooming at its earliest stages: Development and validation of the Online Grooming Risk Scale. Mediterr. J. Clin. Psychol. 2022, 10, 1–24. [Google Scholar] [CrossRef]
- McGhee, I.; Bayzick, J.; Kontostathis, A.; Edwards, L.; McBride, A.; Jakubowski, E. Learning to Identify Internet Sexual Predation. Int. J. Electron. Commer. 2011, 15, 103–122. [Google Scholar] [CrossRef]
- Upadhyay, A.; Chaudhari, A.; Arunesh; Ghale, S.; Pawar, S.S. Detection and prevention measures for cyberbullying and online grooming. In Proceedings of the 2017 International Conference on Inventive Systems and Control (ICISC), Coimbatore, India, 19–20 January 2017. [Google Scholar] [CrossRef]
- Keyvanpour, M.; Nayebi, N.G.; Ebrahimi, M.; Ormandjieva, O.; Suen, C.Y. Automated identification of child abuse in chat rooms by using data mining. In Data Mining Trends and Applications in Criminal Science and Investigations; IGI Global: Hershey, PA, USA, 2016; pp. 245–274. [Google Scholar] [CrossRef]
- Ngejane, C.H.; Mabuza-Hocquet, G.; Eloff, J.H.P.; Lefophane, S. Mitigating Online Sexual Grooming Cyber-crime on Social Media Using Machine Learning: A Desktop Survey. In Proceedings of the 2018 International Conference on Advances in Big Data, Computing and Data Communication Systems (icABCD), Durban, South Africa, 6–7 August 2018. [Google Scholar] [CrossRef]
- Anderson, P.; Zuo, Z.; Yang, L.; Qu, Y. An Intelligent Online Grooming Detection System Using AI Technologies. In Proceedings of the 2019 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), New Orleans, LA, USA, 23–26 June 2019. [Google Scholar] [CrossRef]
- Pranoto, H.; Gunawan, F.E.; Soewito, B. Logistic Models for Classifying Online Grooming Conversation. Procedia Comput. Sci. 2015, 59, 357–365. [Google Scholar] [CrossRef]
- Fauzi, M.A.; Wolthusen, S.; Yang, B.; Bours, P.; Yeng, P. Identifying Sexual Predators in Chats Using SVM and Feature Ensemble. In Proceedings of the 2023 International Conference on Emerging Trends in Networks and Computer Communications (ETNCC), Windhoek, Namibia, 16–18 August 2023; pp. 70–75. [Google Scholar] [CrossRef]
- Ebrahimi, M.; Suen, C.Y.; Ormandjieva, O. Detecting predatory conversations in social media by deep Convolutional Neural Networks. Digit. Investig. 2016, 18, 33–49. [Google Scholar] [CrossRef]
- Nasir, L.H.M.; Saaya, Z.; Baharon, M.R. Identifying Online Sexual Grooming Content in Social Media Using Classification Technique. J. Adv. Comput. Technol. Appl. 2022, 4, 33–42. [Google Scholar]
- Mila, K.C.; Montreal, H.; Caporossi, G.; Rabbany, R.; De Cock, M.; Mila, G.F. Early Detection of Sexual Predators with Federated Learning. May 2023. Available online: https://openreview.net/pdf?id=M84OnT0ZvDq (accessed on 6 May 2025).
- Eilifsen, T.N.; Shrestha, B.; Bours, P. Early Detection of Cyber Grooming in Online Conversations: A Dynamic Trust Model and Sliding Window Approach. In Proceedings of the 2023 21st International Conference on Emerging eLearning Technologies and Applications (ICETA), Stary Smokovec, Slovakia, 26–27 October 2023; pp. 129–134. [Google Scholar] [CrossRef]
- Liu, D.; Suen, C.Y.; Ormandjieva, O. A Novel Way of Identifying Cyber Predators. arXiv 2017, arXiv:1712.03903. [Google Scholar]
- Vogt, M.; Leser, U.; Akbik, A. Early detection of sexual predators in chats. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, Bangkok, Thailand, 1–6 August 2021; pp. 4985–4999. [Google Scholar] [CrossRef]
- Simen, M.A. Fine Tuning BERT for Detecting Cyber Grooming in Online Chats. Master’s Thesis, Norwegian University of Science and Technology, Trondheim, Norway, 2023. Available online: https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3088473 (accessed on 6 May 2025).
- Borj, P.R.; Raja, K.; Bours, P. Detecting Online Grooming by Simple Contrastive Chat Embeddings. In Proceedings of the Thirteenth ACM Conference on Data and Application Security and Privacy, Charlotte, NC, USA, 26 April 2023; pp. 57–65. [Google Scholar] [CrossRef]
- Nguyen, T.T.; Wilson, C.; Dalins, J. Fine-Tuning Llama 2 Large Language Models for Detecting Online Sexual Predatory Chats and Abusive Texts. In Proceedings of the European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium, 9–11 October 2024; pp. 613–618. [Google Scholar] [CrossRef]
- Kim, S.; Lee, B.; Maqsood, M.; Moon, J.; Rho, S. Deep Learning-Based Natural Language Processing Model and Optical Character Recognition for Detection of Online Grooming on Social Networking Services. Comput. Model. Eng. Sci. 2025, 143, 2079–2108. [Google Scholar] [CrossRef]
- Kim, D.; Kim, T.; Yang, J. Early Detection of Online Grooming with Language Models. In Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing, Catania International Airport, Catania, Italy, 31 March–4 April 2025; pp. 963–970. [Google Scholar] [CrossRef]
- Hamm, L. Advancing Grooming Detection in Chat Logs: Comparing Traditional Machine Learning and Large Language Models with a Focus on Predator Tone. Master’s Thesis, Uppsala University, Uppsala, Sweden, 2025. Available online: https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-550352 (accessed on 6 May 2025).
- Pereira, M.; Dodhia, R.; Anderson, H.; Brown, R. Metadata-Based Detection of Child Sexual Abuse Material. IEEE Trans. Dependable Secur. Comput. 2023, 21, 3153–3164. [Google Scholar] [CrossRef]
- Struckman, K. Wilson Center. Combatting AI-Generated CSAM. February 2023. Available online: https://www.wilsoncenter.org/article/combatting-ai-generated-csam (accessed on 6 May 2025).
- European Parliament. Children and Generative AI. February 2025. Available online: https://www.europarl.europa.eu/thinktank/en/document/EPRS_ATA(2025)769494 (accessed on 6 May 2025).
- Hua, Y.; Namavari, A.; Cheng, K.; Naaman, M.; Ristenpart, T. Increasing Adversarial Uncertainty to Scale Private Similarity Testing. In Proceedings of the 31st USENIX Security Symposium, Boston, MA, USA, 10–12 August 2022; pp. 1777–1794. Available online: https://www.usenix.org/system/files/sec22summer_hua.pdf (accessed on 6 May 2025).
- Liang, P.P.; Wu, C.; Morency, L.P.; Salakhutdinov, R. Towards Understanding and Mitigating Social Biases in Language Models. In Proceedings of the 38th International Conference on Machine Learning, Virtual, 18–24 July 2021; Volume 139, pp. 6565–6576. Available online: https://arxiv.org/pdf/2106.13219 (accessed on 6 May 2025).
- Monash University. Digital Child Abuse: Deepfakes and the Rising Danger of AI-Generated Exploitation. February 2025. Available online: https://lens.monash.edu/@politics-society/2025/02/25/1387341/digital-child-abuse-deepfakes-and-the-rising-danger-of-ai-generated-exploitation (accessed on 6 May 2025).
- GOV.UK. Innovating to Detect Deepfakes and Protect the Public. February 2023. Available online: https://www.gov.uk/government/case-studies/innovating-to-detect-deepfakes-and-protect-the-public (accessed on 6 May 2025).
- Bang, Y.; Ji, Z.; Schelten, A.; Hartshorn, A.; Fowler, T.; Zhang, C.; Cancedda, N.; Fung, P. HalluLens: LLM Hallucination Benchmark. arXiv 2025, arXiv:2504.17550. Available online: https://arxiv.org/pdf/2504.17550 (accessed on 6 May 2025).
- EUCPN—European Crime Prevention Network. Artificial Intelligence and Predictive Policing: Risks and Challenges. June 2022. Available online: https://www.eucpn.org/document/recommendation-paper-artificial-intelligence-and-predictive-policing-risks-and-challenges (accessed on 6 May 2025).
- GlobeNewswire. Multimodal AI at a Crossroads: Report Reveals CSEM Risks. May 2025. Available online: https://www.globenewswire.com/news-release/2025/05/08/3077301/0/en/Multimodal-AI-at-a-Crossroads-Report-Reveals-CSEM-Risks.html (accessed on 6 May 2025).
- Coelho, T.; Ribeiro, L.S.F.; Macedo, J.; Santos, J.A.D.; Avila, S. Minimizing Risk Through Minimizing Model-Data Interaction: A Protocol For Relying on Proxy Tasks When Designing Child Sexual Abuse Imagery Detection Models. arXiv 2025, arXiv:2505.06621. Available online: https://arxiv.org/pdf/2505.06621v1 (accessed on 6 May 2025).
- Bilal, A.; Ebert, D.; Lin, B. LLMs for Explainable AI: A Comprehensive Survey. arXiv 2025, arXiv:2504.00125. [Google Scholar]
- European Parliament. Artificial Intelligence Act: MEPs Adopt Landmark Law. Press Release, 13 March 2024. Available online: https://www.europarl.europa.eu/news/en/press-room/20240308IPR19015/artificial-intelligence-act-meps-adopt-landmark-law (accessed on 6 May 2025).
Feature | ChatGPT (OpenAI) | Gemini (Google AI) | LLaMA (Meta AI) |
---|---|---|---|
Core Architecture | GPT with Reinforcement Learning from Human Feedback (RLHF) [12,13]. | Transformer-based with Retrieval-Augmented Generation (RAG) and modifications for TPU efficiency [12,13]. | Transformer-based with Root Mean Square Layer Normalization (RMSQLN) and Swish-Gated Linear Unit (SwiGLU) [13]. |
Multimodality | Primarily text-focused, with some multimodal capabilities noted as a relative weakness [12]. | Natively multimodal, designed to process text, images, audio, and video [12]. | Primarily text-focused, trained on text and code [13]. |
Key Strengths | Excels in conversational flow, creativity, and following instructions [12]. | Prioritizes factual accuracy [12] and high computational efficiency (speed and energy) [12]. | Shows high consistency in performance across different languages, suggesting easier transferability [13]. |
Key Weaknesses | Susceptible to “jailbreaking”, logical inaccuracies, and high computational costs [12]. | Exhibits inconsistent safety/censorship filters [13]. Being a more recent model, it has less real-world data exposure [12]. | Exhibits a strong “optimistic bias”. Larger models have impractical hardware requirements for local use [13]. |
Context Window | Able to handle prolonged interactions and maintain context [12]. | Able to handle extended conversations and decipher intricate prompts [12]. | Trained with a 2 K (LLaMA) or 4 K (LLaMA2) token context length [13]. |
Study | AI Method | Dataset Used | Accuracy (%) | Key Strength | Limitation |
---|---|---|---|---|---|
Jeglic et al. (2022) [9] | SGS-V Scale | CSA Victim Reports | 88 | Early grooming indicators | Recall bias |
Pasca et al. (2022) [14] | OGR Scale | Survey-Based | 89 | Focus on risk factors | Limited generalizability |
Kontostathis et al. (2011) [15] | Rule-based system ChatCoder, decision trees, instance-based learning | Perverted Justice | 68 | ChatCoder 2’s classifications were more accurate than the hand-coded truth set | ML algorithms did not significantly outperform the rule-based ChatCoder system |
Upadhyay et al. (2017) [16] | Machine learning, image/content filtering, NLP | Bad Words Dataset, Sensitive Words Dataset | Not provided | Promising potential to protect youth from online grooming and cyberbullying | Complexity of distinguishing lawful content from illicit material |
Keyvanpour et al. (2018) [17] | Text Mining and SVM | PAN2012 | 85 | Network behavior analysis | Imbalanced datasets |
Ngejane et al. (2018) [18] | SVMs, k-NNs, CNNs, semi-supervised anomaly detection | Perverted Justice, PAN2012 and MovieStarPlanet | 98 | Models demonstrated promising detection capabilities | Scarcity of labeled datasets, and the difficulty of generalizing results to real scenarios |
Borj et al. (2019) [7] | Neural Networks | Perverted Justice and PAN2012 | 90 | Diverse algorithm use | Privacy concerns |
Zuo et al. (2019) [19] | Fuzzy-rough feature selection and fuzzy twin SVMs | Perverted Justice, PAN13-Author-Profiling | 61 | AI technologies integration promise for overcoming the limitations of traditional detection methods | The ever-evolving nature of grooming language and chat formats; Need for a larger representative dataset |
Pranoto et al. (2020) [20] | Logistic Regression | Perverted Justice | 95 | High detection accuracy | Lack of child-specific language |
Fauzi et al. (2020) [21] | SVM with BoW | PAN2012 | 98 | High accuracy in VPD | Noisy data impact |
Ebrahimi et al. (2021) [22] | One-Class SVM | PAN2012 | 75 | Works with less labeled data | Lower F1-score |
Nasir et al. (2022) [23] | Text mining, text classification, neural network, n-gram and sequence models | YouTube, Instagram, and Twitter | 82 | Enhanced precision with n-gram and sequence models | Sequence model accuracy lower than n-gram model |
Chebouni et al. (2022) [24] | Logistic Regression | PANC | 79 | Privacy-preserving approach, addresses non-IID data, early detection focus | Limitations of PANC, biases in BERT, computational cost of FL |
Eilifsen et al. (2023) [25] | Naive Bayes, Tree-Based Models, Neural Networks, and Dynamic Trust Model | PAN2012 | 88 | Multinomial Naive Bayes model with Term Frequency (mult-nb-tf) demonstrated superior accuracy | Modest sample size, rigidity of fixed sliding window dimensions |
Liu et al. (2017) [26] | LSTM-RNN for sentence vectors and conversation classification; Fast text for sentiment analysis | PAN2012 dataset; IMDB movie review dataset (for sentence vector evaluation) | 99.43 (chat); 98.35 (Predator); 83.2 (IMDB) | Effectively captures dependencies in conversations; Sentence vectors reduce input dimensionality; Sentiment analysis enhances predator identity | Performance variability on IMDB dataset; Complexity in processing noisy and varied online chat data |
Puentes et al. (2021) [6] | BERT-based LLM | Te Protejo Dataset | 93 | Handles sensitive data | Data sensitivity issues |
Vogt et al. (2021) [27] | BERT Transformer | PAN2012 and ChatCoder2 | 92 | Early-stage detection | Limited real-world data |
Aarnseth et al. (2023) [28] | BERT | PAN2012 and AIBA AS | Uses F1-score; F1-score of 86 | Model is highly robust, proving effective at detecting cyber grooming even when conversations contain a large amount of informal language and slang | Data imbalance |
Borj et al. (2023) [29] | RoBERTa/BERT, SVM | PAN2012 | 99 | Robust sentence-based feature extraction for high true-positive predatory conversation detection. | Accurately interpreting nuanced online conversations is challenging due to the context-dependent nature of language. |
Nguyen et al. (2023) [30] | Llama 2 LLM | PAN2012 and Urdu | 96 | Multilingual capabilities | Data imbalance in non-English sets |
Prosser and Edwards (2024) [11] | ChatGPT3.5, 4, PaLM2, Claude2, LLaMA2, Mistral | ChatCoder2 and Perverted Justice | Not provided | Prompt Design Impact Analysis | No LLM reliable; Consistency/Safety Issues |
Rho et al. (2025) [31] | KcELECTRA | A combination of Korean datasets | 95 | The framework outperforms existing transformer-based models in accuracy and is specifically optimized for the nuances of colloquial Korean language found in SNS conversations | Data scarcity and semantic similarities between hate speech and sexually explicit content pose challenges |
Yang et al. (2025) [32] | BERT and LLM models | PAN2012 dataset translated into Korean augmented with Korean SNS | 81 | Proposes a new metric, Human-to-Model Ratio (HMR), for a more nuanced evaluation of detection speed | Models struggle to interpret figurative language used in grooming conversations |
Hamm et al., (2025) [33] | LLaMA 3.2 1B | PAN2012 | 99 | The LLaMA 3.2 1B model outperforms both traditional machine learning models and previous, larger LLMs in grooming detection | The research relies on the PAN12 dataset, which is over a decade old and features law enforcement officers posing as victims, not real children |
Study | AI Method | Dataset Used | Dataset Domain | Dataset Language | Size |
---|---|---|---|---|---|
Jeglic et al. (2022) [9] | SGS-V Scale | CSA Victim Reports | Victim survey reports | English | 913 participants |
Pasca et al. (2022) [14] | OGR Scale | Survey-Based | Adolescent survey | Not Specified | 316 adolescents |
Kontostathis et al. (2011) [15] | Rule-based system ChatCoder, decision trees, instance-based learning | Perverted Justice | Online chat logs | English | 50 transcripts selected (33 used) |
Upadhyay et al. (2017) [16] | Machine learning, image/content filtering, NLP | Bad Words Dataset, Sensitive Words Dataset | Social media content | Not Specified | Not Specified |
Keyvanpour et al. (2018) [17] | Text Mining and SVM | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Ngejane et al. (2018) [18] | SVMs, k-NNs, CNNs, semi-supervised anomaly detection | Perverted Justice, PAN2012 and MovieStarPlanet | Online chat logs, social gaming | English | 288,142 lines (PAN2012 training set) |
Borj et al. (2019) [7] | Neural Networks | Perverted Justice and PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Zuo et al. (2019) [19] | Fuzzy-rough feature selection and fuzzy twin SVMs | Perverted Justice, PAN13-Author-Profiling | Online chat logs | English | 1200 documents |
Pranoto et al. (2020) [20] | Logistic Regression | Perverted Justice | Online chat/story logs | English | 160 transcripts |
Fauzi et al. (2020) [21] | SVM with BoW | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Ebrahimi et al. (2021) [22] | One-Class SVM | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Nasir et al. (2022) [23] | Text mining, text classification, neural network, n-gram and sequence models | YouTube, Instagram, and Twitter | Social media comments | Not Specified | 1000 comments |
Chebouni et al. (2022) [24] | Logistic Regression | PANC | Online chat logs | English | 32,510 segments |
Eilifsen et al. (2023) [25] | Naive Bayes, Tree-Based Models, Neural Networks, and Dynamic Trust Model | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Liu et al. (2017) [26] | LSTM-RNN for sentence vectors and conversation classification; Fast text for sentiment analysis | PAN2012 dataset; IMDB movie review dataset (for sentence vector evaluation) | Online chat logs, movie reviews | English | PAN2012 (288,142 lines); IMDB (50,000 reviews) |
Puentes et al. (2021) [6] | BERT-based LLM | Te Protejo Dataset | Child abuse hotline reports | Spanish | 1196 reports |
Vogt et al. (2021) [27] | BERT Transformer | PAN2012 and ChatCoder2 | Online chat logs | English | 32,510 segments |
Aarnseth et al. (2023) [28] | BERT | PAN2012 and AIBA AS | Online chat logs | English | PAN12 (288,142 lines); AIBA AS (4429 messages) |
Borj et al. (2023) [29] | RoBERTa/BERT, SVM | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Nguyen et al. (2023) [30] | Llama 2 LLM | PAN2012 and Urdu | Online chat/abusive texts | English, Urdu | PAN2012 (288,142 lines); Roman Urdu (20,000 samples); Urdu (15,000 samples) |
Prosser and Edwards (2024) [11] | ChatGPT3.5, 4, PaLM2, Claude2, LLaMA2, Mistral | ChatCoder2 and Perverted Justice | Online chat logs | English | 100 prompts |
Rho et al. (2025) [31] | KcELECTRA | A combination of Korean datasets | Social media chat images | Korean | ~1.7 million sentences + 9400 comments |
Yang et al. (2025) [32] | BERT and LLM models | PAN2012 dataset translated into Korean augmented with Korean SNS | Online chat logs | Korean | 57 training + 147 test convos (+888 SNS convos) |
Hamm et al., (2025) [33] | LLaMA 3.2 1B | PAN2012 | Online chat logs | English | 288,142 lines (PAN2012 training set) |
Category | Limitation | Studies Citing This Limitation |
---|---|---|
Data-Related | Data Scarcity, Quality and Imbalance: Lack of large, high-quality, labeled, and balanced datasets. | Keyvanpour et al. (2018) [17], Ngejane et al. [18], Zuo et al. [19], Aarnseth et al. [28], Nguyen et al. [30], Rho et al. [31] |
Outdated Datasets: Reliance on older datasets (e.g., PAN2012) that may not reflect current online behaviors. | Hamm et al. [33] | |
Lack of Realism: Datasets using volunteers instead of real victims can lack authenticity. | Vogt et al. [27], Hamm et al. [33] | |
Noisy and Unreliable Data: User-generated content is often unstructured, contains misspellings, and slang. | Liu et al. [26], Fauzi et al. [21] | |
Technical/Methodological | Model Reliability and Consistency: LLMs can provide inconsistent or overly cautious responses and are prone to “hallucination”. | Prosser and Edwards [11] |
Contextual Understanding: Difficulty in interpreting nuanced, figurative, or evolving language. | Borj et al. [7], Borj et al. [29], Yang et al. [32] | |
Difficulty of the Task: The inherent complexity of distinguishing legitimate content from harmful content. | Upadhyay et al. [16], Liu et al. [26] | |
Recall Bias and Methodological Rigidity: Limitations in non-AI studies, such as recall bias in surveys or rigid window sizes. | Jeglic et al. [9], Eilifsen et al. [25] | |
Ethical and Privacy | Privacy Concerns: Monitoring online communications raises significant privacy issues. | Borj et al. [7] |
Algorithmic Bias: Models can have inherent biases (e.g., from BERT) that need mitigation. | Chehbouni et al. [24] | |
Data Sensitivity: Handling real reports from abuse hotlines requires extreme care and restricts data availability. | Puentes et al. [6] | |
Implementation and Generalization | Limited Generalizability: Difficulty in applying models trained on one dataset to different, real-world scenarios. | Pasca et al. [14], Ngejane et al. [18] |
Computational Cost: The cost of training and deploying large or federated models can be a significant barrier. | Chehbouni et al. [24] |
Intervention Timeline | Description | Approach/Key Feature | Studies |
---|---|---|---|
Foundational/Risk Factor Identification | Research focused on identifying grooming behaviors and risk factors without creating an automated detection system. This work is foundational for later AI models. | Development and validation of scales (SGS-V, OGR) through surveys and statistical analysis. | Jeglic et al. [9], Pasca et al. [14] |
Retrospective Analysis | Analysis of static, historical datasets of conversations or content. Useful for investigations, pattern discovery, and benchmarking models. | Classification and analysis of established datasets like PAN2012, Perverted Justice, or social media data. | Keyvanpour et al. [17], Fauzi et al. [21], Ebrahimi et al. [22], Liu et al. [26], Borj et al. [29], Hamm et al. [33] |
Real-Time/Early Detection | Monitoring and analyzing conversations as they occur to flag suspicious activity immediately, enabling intervention before harm escalates. | Use of sliding window techniques, federated learning on user devices, or memory-based context retrieval. | Chehbouni et al. [24], Eilifsen et al. [25], Vogt et al. [27], Yang et al. [32] |
Multilingual and Multimodal Detection | Systems designed specifically to handle different languages or content types beyond text (e.g., images). | Fine-tuning language-specific models (e.g., Korean) or using OCR to extract text from images. | Nguyen et al. [30], Rho et al. [31] |
LLM Capability Assessment | Studies focused not on building a detector, but on evaluating the inherent capabilities and safety of existing, general-purpose LLMs for this task. | Prompt-based testing of commercial LLMs like ChatGPT, PaLM2, etc. | Prosser and Edwards [11] |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Silva, F.; Silva, R.R.; Bernardino, J. A Survey of Generative AI for Detecting Pedophilia Crimes. Appl. Sci. 2025, 15, 7105. https://doi.org/10.3390/app15137105
Silva F, Silva RR, Bernardino J. A Survey of Generative AI for Detecting Pedophilia Crimes. Applied Sciences. 2025; 15(13):7105. https://doi.org/10.3390/app15137105
Chicago/Turabian StyleSilva, Filipe, Rodrigo Rocha Silva, and Jorge Bernardino. 2025. "A Survey of Generative AI for Detecting Pedophilia Crimes" Applied Sciences 15, no. 13: 7105. https://doi.org/10.3390/app15137105
APA StyleSilva, F., Silva, R. R., & Bernardino, J. (2025). A Survey of Generative AI for Detecting Pedophilia Crimes. Applied Sciences, 15(13), 7105. https://doi.org/10.3390/app15137105