What technology changes in romance scams

Reviewed evidence describes social platforms as common initial contact venues for relationship-based fraud, followed by moves to less-monitored messaging channels. Generative-AI and deepfake tools can create realistic synthetic media with lower skill and time requirements; newly generated images may have no external source for a reverse-image search to find. Automation evidence includes nearly identical messages in an acquisition-stage study and predefined chatbots in fraudulent dating apps. Other studies describe operations in which software assigns multiple workers to clients, alongside false profiles, foreign phone numbers, and sometimes virtual private networks. These observations do not establish how widely large language models run scams; that proportion remains unknown. Romance-linked investment scams use fake applications and fabricated trading dashboards. Blockchain research describes hub, intermediary, and holding roles in one network, while another study reports recirculation and swapping across addresses and cryptocurrencies. Reconstructed Bitcoin paths remain uncertain because transfers can include change returned to the sender. Researchers also describe Maltego-based information gathering and a prototype classifier with reported detection metrics. A separate identifier-extraction study reports high agreement between human annotators; that is a labeling result, not proof of prevention. Detection performance depends on input quality, and message analysis raises privacy and acceptability questions. Fragmented research, unavailable victim–scammer data, and an enforcement-selected blockchain case constrain broader inference. The reviewed results do not establish population prevalence or reductions in victims’ losses (Abubakari et al., 2026; Cross, 2022b; Dickerson et al., 2020; Dominguez Castillo, 2026; Franceschini et al., 2023; Griffin & Mei, 2024; Gujarathi et al., 2026; Hu et al., 2021; Jakobsson, 2016; Kassem & Carter, 2023; Ko et al., 2026; Pope & Seto, 2026; Stern & Choi, 2026; Vedhanayagam et al., 2026; Wang & Topalli, 2024).

Platform ecology and channel shifting

Multiple studies using complaints, corpora, and case evidence identify social‑network services and dating platforms as common initial contact venues for relationship‑based fraud; after rapport is established, scammers frequently redirect conversations to less‑monitored messaging apps or phone/VoIP channels to continue grooming and seek payments (Ordekian et al., 2024; Pope & Seto, 2026; Wang & Kelsay, 2025; Yan, 2024).

Platform affordances—public profile content, friend networks, and searchability—are repeatedly reported to enable tailored targeting and personalized scripts; social‑media visibility of apparent wealth or shared connections can also be used to normalize deceptive behavior or recruit accomplices in some operations (Barnor, 2024; Cross, 2022a; Kopp et al., 2016; Ohu & Jones, 2025).

Platform findings reflect the setting and period studied. One survey named Facebook as the most common first-contact platform (50.3%); a South Korean account described KakaoTalk’s prominence, and a Chinese account described dating or marriage platforms and chance WeChat contact. Another source states that Tinder and Bumble are banned in Iran. A forum chapter warns that its platform ranking may have changed since 2021. These findings support checking local and current usage before designing platform-specific prevention, rather than treating one ranking as universal (Amirkhani et al., 2024; Cantin et al., 2025; Choi et al., 2024; Corpuz et al., 2025; Yan, 2024).

Authors report that offenders commonly use multiple applications during a single scam and often move victims to less‑surveilled channels (for example WhatsApp, Telegram, or other messaging services) after trust is established, a tactic linked to isolation from platform monitoring and continued solicitation (Ordekian et al., 2024; Wang & Kelsay, 2025).

AI, deepfakes, and verification challenges

Several conceptual reviews, case reports, and forensic observations document that readily available generative‑AI and deepfake tools can produce realistic synthetic faces, voices, and video at lower technical thresholds than earlier methods; these capabilities can undermine provenance‑based checks (for example reverse‑image searches) because synthetic images may have no external source to match (Cross, 2022b; Cross & Layt, 2021; Kassem & Carter, 2023).

Investigative case material and media‑documented incidents attribute large losses to scams that reportedly used AI‑generated media (including deepfake videos and voice cloning) in some instances, and authors explicitly call for empirically validated, accessible detection tools rather than assuming existing provenance checks suffice (Choi et al., 2024; Cross, 2022b; Mouri, 2024).

At the same time, complaint‑based analyses and observational casework report many instances where reverse‑image searches or other identity checks exposed stolen or reused photos and helped substantiate suspicions, although authors also note cases with no positive matches where the absence of a footprint raised concern (Botha et al., 2025; Cross & Layt, 2021; Khukhunaishvili, 2024).

Because evidence shows both successful image‑search detections and plausible use of uniquely generated synthetic images that lack linkable provenance, authors call for development and validation of accessible deepfake‑detection methods and image‑authentication tools as a research and operational priority (Cross, 2022b; Cross & Layt, 2021).

Automation, semi‑industrialization, and LLM risks

Evidence of automation concerns different tasks and settings. One message analysis found nearly identical replies and interpreted them as evidence of automation during initial victim acquisition. A fraudulent-dating-app study reported that most observed accounts used chatbots built around predefined templates. Another dating-scam study did not find conclusive evidence of botnet use and noted that sustained interaction was harder to automate than spam or phishing. Separately, a study of an organized operation describes software assigning multiple workers to each client, showing that automation can coordinate people rather than replace them (Hu et al., 2021; Huang et al., 2015; Jakobsson, 2016; Wang & Topalli, 2024).

Qualitative offender‑facing studies and court‑file analyses show organized, multi‑worker operations that use algorithms, worker assignment, background files, VPNs, and foreign phone numbers—practices that can be supported by but are distinct from fully autonomous LLM‑driven end‑to‑end scams; the reviewed corpus does not provide validated estimates of the proportion of operations that are end‑to‑end LLM‑automated (Abubakari et al., 2026; Dominguez Castillo, 2026; Ogunleye et al., 2020; Pope & Seto, 2026; Wang & Topalli, 2024).

Experimental evaluations of LLM safety in romance‑fraud contexts show model susceptibility to jailbreak prompts under specific conditions, but authors caution generalization is limited by prompt choice, single‑turn designs, and evaluation criteria—evidence that LLMs present both opportunities to scale malicious content and limits in current operational use (Parra et al., 2026).

Authors note an empirical gap: the precise proportion of romance‑fraud operations using LLMs rather than traditional scripted techniques remains unknown, and measurement of live operational LLM use requires targeted empirical designs reported as a research priority in multiple sources. This gap constrains confident claims about end‑to‑end automation prevalence (Dominguez Castillo, 2026; Sorell & Whitty, 2019).

Pig‑butchering, blockchain forensics, and crypto‑enabled schemes

Multiple case studies and mixed‑method analyses characterize pig‑butchering as a hybrid long‑duration trust‑building process that funnels victims into fabricated trading dashboards, fake apps/websites, or offender‑controlled exchanges; distribution commonly uses web links or QR codes rather than legitimate app‑store channels (Franceschini et al., 2023; Gujarathi et al., 2026; Wang & Zhou, 2022).

Blockchain-forensics studies describe different aspects of scam-related transactions. One case identifies hub, intermediary, and holding roles within the observed network. Another study reports extensive recirculation and swapping across addresses and cryptocurrencies, which impede tracing. Reconstructed Bitcoin paths also remain uncertain because a transfer may include both payments and change returned to the sender; a displayed route is therefore not a deterministic account of every transfer (Griffin & Mei, 2024; Stern & Choi, 2026).

Authors recommend combining on‑chain and off‑chain evidence in living forensic collaborations and real‑time analyst partnerships to improve interpretability and to keep pace with evolving laundering techniques; several sources frame these cooperative approaches as necessary given enforcement‑selected or partial sampling frames in forensic case work (Gujarathi et al., 2026; Stern & Choi, 2026).

A blockchain-forensics case study used sanctions issued by the Office of Foreign Assets Control as an enforcement-based sampling frame. The authors describe the mapped structure as a partial representation of the observed network and recommend comparisons with other address sets and sanctioned networks. Its scope differs from a separate study reporting extensive recirculation and swapping across addresses and cryptocurrencies. These findings should not be combined as though they describe one sample or establish the structure of every pig-butchering network (Griffin & Mei, 2024; Stern & Choi, 2026).

Detection prototypes, OSINT workflows, privacy, and methodological patterns

Researchers have developed detection prototypes and workflows for gathering public information. One study combined material from public scam websites with Maltego, applied TF-IDF text features, and used logistic regression to classify material from emails and links. It reported detection metrics for its prototype. A separate identifier-extraction study reported high agreement between human annotators across wallet addresses, platform URLs, and loss amounts. Agreement about labels and classifier performance are different results and should be evaluated separately (Ko et al., 2026; Vedhanayagam et al., 2026).

Detection researchers caution that algorithm effectiveness and accuracy depend heavily on the quality and relevance of input data. Proposed research directions include developing detection profiles from communication and payment patterns, while a scoping review recommends living evidence syntheses and real-time collaboration with forensic analysts to maintain relevance. These are cautions and research recommendations, not demonstrations of real-world detection effectiveness (Gujarathi et al., 2026; Sorell & Whitty, 2019; Vedhanayagam et al., 2026).

At the same time, dynamic analysis of private messages or behavioral profiling raises clear privacy, acceptability, and governance concerns; the evidence emphasizes participatory evaluation of user acceptability, transparency, and bias mitigation before deploying message‑scanning or automated warning systems (Dickerson et al., 2020).

The reviewed work includes research based on public scam websites and a blockchain case selected through sanctions records. A review within the detection literature describes fragmented, mainly qualitative research and limited access to victim–scammer data. These sources provide examples of operational methods and available datasets, but they do not establish that the examined cases represent all romance-scam activity (Stern & Choi, 2026; Vedhanayagam et al., 2026).

Evidence gaps and recommended future directions

Authors repeatedly identify key evidence gaps that constrain confident conclusions: (a) empirical measurement of how extensively LLMs/generative AI are used in live scams and whether they operate end‑to‑end, (b) real‑world validation of detection systems with platform and private‑message data, and (c) field‑tested deepfake‑detection and platform mitigation strategies that preserve privacy and legal rights (Cross, 2022b; Dominguez Castillo, 2026; Rabby & Chowdory, 2024; Sorell & Whitty, 2019; Vedhanayagam et al., 2026).

Research priorities and practical limits

Recommended future work includes living evidence syntheses and collaboration with forensic analysts; native-language datasets and attention to whether detection would be implemented on the platforms people use; recognition that analyzing private conversations intrudes on privacy; and technology-focused reports by academics and civil-society organizations to inform understanding and anti-trafficking planning. These are research and implementation priorities, not demonstrated prevention outcomes (Amirkhani et al., 2026; Dickerson et al., 2020; Gujarathi et al., 2026; Luong & Ngo, 2024).

Evidence gaps and recommended future directions

One prototype detection study reported 85% accuracy, 88% precision, 80% recall, and an F1 score of 0.84. Its authors state that effectiveness depends heavily on input quality and relevance. A separate identifier-extraction study reported high agreement between human annotators (Cohen’s κ: wallet addresses 0.978, platform URLs 0.988, loss amounts 0.969); these are agreement measures, not the prototype’s detection scores. In a different policy discussion, authors invite nongovernmental organizations, civil-society groups, and academics to provide technology-focused analyses to clarify the scale of transnational scam-related fraud and inform anti-trafficking initiatives (Ko et al., 2026; Luong & Ngo, 2024; Vedhanayagam et al., 2026).

Related-work reviews characterize romance-scam research as limited, fragmented, and inconsistent because victim–scammer data are not readily available. Future interdisciplinary research should develop detection profiles from recognisable scammer strategies, jurisdictions, grooming stages, communication patterns, and payment patterns (Sorell & Whitty, 2019; Vedhanayagam et al., 2026).

Additional detector and analysis evidence

The Arikui profile filter used a balanced dataset of approximately 30,000 dubious and 30,000 normal users and reported 94.1% random-forest cross-validation accuracy after a 90/10 split. Real deployment nevertheless produced many false positives because normal and dubious users were not equally common. The authors also note that undetected dubious users may remain in the normal group. These conditions distinguish balanced-data accuracy, uncertain labels and practical deployment performance; none establishes a measured reduction in victim losses (Neve & Palomares, 2018).

A French-language linguistic project used 180 romance-fraud cases collected through baiting suspected scammers over three months. The language model assisted categorization, followed by segment review and manual reassignment. ChatGPT’s repeated-segment classifications required reclassification of 52% of segments in that task. This is an observed analytic limitation, not a universal AI error rate, proof of offender identity or a measure of financial harm prevented (Degeneve et al., 2026).

A real-time chatbot-stage study evaluated 152 manually labeled Reddit items. It reported poor performance of 28.6% when evaluation input contained only single-stage data. This condition-specific classification outcome should not be generalized to every input arrangement or mistaken for the percentage of real-world scams detected before financial loss (Liu et al., 2026).

The Myanmar network-analysis authors selected Grok-4 as the winner of their iterative parsing comparison, but explicitly state that systematic validation, reliability and reproducibility testing were beyond scope and standard precision, recall, F1 and error-rate metrics were not computed. Their selection therefore should not be treated as a general model ranking or an independently validated performance benchmark (Akartuna & Manning, 2025).

References

• Abubakari, Y., Lazarus, S., & Oseh-Ovarah, V. (2026). A crime script perspective on mapping the entry, continuation, and exit pathways of online romance fraud. International Journal of Comparative and Applied Criminal Justice, 1–32. https://doi.org/10.1080/01924036.2026.2645937 Publication record
• Akartuna, E. A., & Manning, M. G. (2025). An Artificial Intelligence-Driven Temporal Network Analysis of Myanmar’s Cyber Scam Ecosystem. Asian Journal of Criminology. https://doi.org/10.1007/s11417-025-09474-0 Publication record
• Amirkhani, S., Alizadeh, F., Randall, D., & Stevens, G. (2024). Beyond Dollars: Unveiling the Deeper Layers of Online Romance Scams Introducing “Body Scam”. Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, 1–6. https://doi.org/10.1145/3613905.3651004 Publication record
• Amirkhani, S., Alizadeh, M. F., Randall, D., Stevens, G., & Zytko, D. (2026). My Parents Expectations Were Overwhelming: Online Dating Romance Scams Targeting Minors in Iran Through Exploitation of Parental Pressure. CHI Conference on Human Factors in Computing Systems. https://doi.org/10.48550/arxiv.2601.16321 Publication record
• Barnor, J. (2024). The charade of discreetness: Exploring the paradoxical lifestyles of romance fraudsters. Journal of Investigative Psychology and Offender Profiling, 21(3). https://doi.org/10.1002/jip.1637 Publication record
• Botha, J. G., Singh, K., & Leenen, L. (2025). Analysis of a Cryptocurrency Investment Scam: Pig Butchering. European Conference on Cyber Warfare and Security, 24(1), 61–70. https://doi.org/10.34190/eccws.24.1.3558 Publication record
• Cantin, P.-M., Ream, F., & Dupont, B. (2025). The Online Mutual Help Practices of Romance Fraud Victims. The Security of Self, 179–194. https://doi.org/10.2307/jj.37022854.16 Publication record
• Choi, S. W., Lee, J., & Choi, Y. J. (2024). Unveiling the Patterns of Romance Scams in South Korea. International Journal of Cyber Behavior, Psychology and Learning, 14(1), 1–15. https://doi.org/10.4018/ijcbpl.357152 Publication record
• Corpuz, D. R. R. N., Galang, M. J. G., Gueco, K. H. S., & Pamintuan, K. D. (2025). ROMANCE AND RUIN: THE INTERPLAY OF FINANCIAL LOSS AND PSYCHOLOGICAL WELL-BEING AMONG LOVE SCAM VICTIMS IN THE PHILIPPINES. EPRA International Journal of Multidisciplinary Research (IJMR), 675. https://doi.org/10.36713/epra25491 Publication record
• Cross, C. (2022a). Meeting the Challenges of Fraud in a Digital World. The Handbook of Security, 217–238. https://doi.org/10.1007/978-3-030-91735-7_11 Publication record
• Cross, C. (2022b). Using artificial intelligence (AI) and deepfakes to deceive victims: the need to rethink current romance fraud prevention messaging. Crime Prevention and Community Safety, 24(1), 30–41. https://doi.org/10.1057/s41300-021-00134-w Publication record
• Cross, C., & Layt, R. (2021). “I Suspect That the Pictures Are Stolen”: Romance Fraud, Identity Crime, and Responding to Suspicions of Inauthentic Identities. Social Science Computer Review, 40(4), 955–973. https://doi.org/10.1177/0894439321999311 Publication record
• Degeneve, C., Longhi, J., & Rossy, Q. (2026). L’analyse des traces langagières comme vecteur de détection de relations entre des cas de fraudes en ligne. Criminologie, Forensique et Sécurité4(1):8332. https://doi.org/10.26034/la.cfs.2026.8332 Publication record
• Dickerson, S., Apeh, E., & Ollis, G. (2020). Contextualised Cyber Security Awareness Approach for Online Romance Fraud. 2020 7th International Conference on Behavioural and Social Computing (BESC), 1–6. https://doi.org/10.1109/besc51023.2020.9348307 Publication record
• Dominguez Castillo, L. (2026). Industrialized heartbreak: how generative AI enables romance fraud at scale. AI and Ethics, 6(3). https://doi.org/10.1007/s43681-026-01129-9 Publication record
• Franceschini, I., Li, L., & Bo, M. (2023). Compound Capitalism: A Political Economy of Southeast Asia’s Online Scam Operations. Critical Asian Studies, 55(4), 575–603. https://doi.org/10.1080/14672715.2023.2268104 Publication record
• Griffin, J. M., & Mei, K. (2024). How Do Crypto Flows Finance Slavery? The Economics of Pig Butchering. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4742235 Publication record
• Gujarathi, P., Verma, S., & Nair, V. V. (2026). Pig Butchering Scams as Cyber-Enabled Financial Crime: A Scoping Review of Dimensions, Modus Operandi, and Victim-Offender Dynamics. Deviant Behavior, 1–16. https://doi.org/10.1080/01639625.2026.2677691 Publication record
• Hu, Y., Wang, H., Zhou, Y., Guo, Y., Li, L., Luo, B., & Xu, F. (2021). Dating with Scambots: Understanding the Ecosystem of Fraudulent Dating Applications. IEEE Transactions on Dependable and Secure Computing, 18(3), 1033–1050. https://doi.org/10.1109/tdsc.2019.2908939 Publication record
• Huang, J., Stringhini, G., & Yong, P. (2015). Quit Playing Games with My Heart: Understanding Online Dating Scams. Lecture Notes in Computer Science, 216–236. https://doi.org/10.1007/978-3-319-20550-2_12 Publication record
• Jakobsson, M. (2016). Understanding Social Engineering Based Scams. https://doi.org/10.1007/978-1-4939-6457-4 Publication record
• Kassem, R., & Carter, E. (2023). Mapping romance fraud research – a systematic review. Journal of Financial Crime, 31(4), 974–992. https://doi.org/10.1108/jfc-06-2023-0160 Publication record
• Khukhunaishvili, D. (2024). Romance Scam as one of the main Challenges of cybercrime. https://doi.org/10.13140/rg.2.2.23659.14880 Publication record
• Ko, S., Kim, J., Jang, J. Y., & Kim, G. (2026). Human‐supervised LLM triage of pig butchering complaints: A validation study of multi‐model identifier extraction. Journal of Forensic Sciences. https://doi.org/10.1111/1556-4029.70411 Publication record
• Kopp, C., Layton, R., Sillitoe, J., & Gondal, I. (2016). The Role of Love stories in Romance Scams: A Qualitative Analysis of Fraudulent Profiles. Zenodo (CERN European Organization for Nuclear Research). https://doi.org/10.5281/zenodo.56227 Publication record
• Liu, Y.-C., Lai, G.-R., Lai, Y.-C., & Chen, L. C.-L. (2026). Real-time Chatbot Romance Scam Detection in Stage Classification. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2610.20631 Publication record
• Luong, H. T., & Ngo, H. M. (2024). Understanding the Nature of the Transnational Scam-Related Fraud: Challenges and Solutions from Vietnam’s Perspective. Laws, 13(6), 70. https://doi.org/10.3390/laws13060070 Publication record
• Mouri, M. (2024). Romance Scam and Legal Interpreting/Translation. IICAH Official Conference Proceedings, 471–475. https://doi.org/10.22492/issn.2432-4604.2024.41 Publication record
• Neve, J., & Palomares, I. (2018). Arikui - A Dubious User Detection System for Online Dating in Japan. 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC). https://doi.org/10.1109/smc.2018.00395 Publication record
• Ogunleye, Y. O., Ojedokun, U. A., & Aderinto, A. A. (2020). Pathways and Motivations for Cyber Fraud Involvement among Female Undergraduates of Selected Universities in South-West Nigeria. Zenodo. https://doi.org/10.5281/zenodo.3702333 Publication record
• Ohu, F., & Jones, L. (2025). An Examination of Digital Validation-Seeking Behaviors in Adolescents as Precursors to Romance Scamming. Scientia Moralitas Conference Proceedings. https://doi.org/10.5281/zenodo.14911844 Publication record
• Ordekian, M., Papasavva, A., Mariconti, E., & Vasek, M. (2024). A Sinister Fattening: Dissecting the Tales of Pig Butchering and Other Cryptocurrency Scams. 2024 APWG Symposium on Electronic Crime Research (eCrime), 136–148. https://doi.org/10.1109/ecrime66200.2024.00016 Publication record
• Parra, Y. J., Chi, H., Alo, R. A., & Lima, V. (2026). Evaluating Jailbreak Vulnerabilities in LLMs: A Taxonomy and Comparative Analysis in Romance Fraud Scenarios. 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC), 1–6. https://doi.org/10.1109/icaic67076.2026.11395743 Publication record
• Pope, T., & Seto, C. H. (2026). From swipe to swindle: a narrative review of research on older adult victims of romance fraud. Journal of Elder Abuse & Neglect, 38(4), 423–443. https://doi.org/10.1080/08946566.2026.2690996 Publication record
• Rabby, F., & Chowdory, M. M. U. (2024). Romance Scamming: Uncovering the Transnational Crime and Legal Challenges. International Journal of Law and Societal Studies, 1(1), 48–62. https://doi.org/10.61424/ijlss.v1i1.134 Publication record
• Sorell, T., & Whitty, M. (2019). Online romance scams and victimhood. Security Journal, 32(3), 342–361. https://doi.org/10.1057/s41284-019-00166-w Publication record
• Stern, M., & Choi, K.-S. (2026). Three-Tier On-Chain Transaction Architecture in a Sanctions-Linked Pig-Butchering Network: A Blockchain-Forensics Case Study. International Journal of Cybersecurity Intelligence & Cybercrime, 9(2). https://doi.org/10.52306/2578-3289.1249 Publication record
• Vedhanayagam, P., Singh, M., Dadhania, A. P., & Ikram, S. T. (2026). Forensic footprints in digital love: Unveiling romance scams with Maltego and machine learning. AIP conference proceedings, 3449, 020278. https://doi.org/10.1063/5.0298575 Publication record
• Wang, F., & Kelsay, J. D. (2025). The prevention of online romance scams using a crime script analysis from the victim’s perspective. International Review of Victimology. https://doi.org/10.1177/02697580251377100 Publication record
• Wang, F., & Topalli, V. (2024). The cyber-industrialization of catfishing and romance fraud. Computers in Human Behavior, 154, 108133. https://doi.org/10.1016/j.chb.2023.108133 Publication record
• Wang, F., & Zhou, X. (2022). Persuasive Schemes for Financial Exploitation in Online Romance Scam: An Anatomy on Sha Zhu Pan (杀猪盘) in China. Victims & Offenders, 18(5), 915–942. https://doi.org/10.1080/15564886.2022.2051109 Publication record
• Yan, Y. (2024). Explore the Causes of Crime and Social Impact of Romance Fraud in Mainland China from 2019–2021. Advances in Social Science, Education and Humanities Research, 601–608. https://doi.org/10.2991/978-2-38476-271-2_73 Publication record

This AI-assisted synthesis is based on reviewed evidence records. It may contain errors or omissions. Follow the publication and DOI links, consult the original works, and make your own judgment about the evidence.