Technology and AI: Evidence Synthesis on Digital Platforms, Automation, Deepfakes, and Emerging Technologies in Relationship-Based Fraud

Across the reviewed evidence, digital platforms and mainstream social-media and dating services are consistently documented as the primary initiation channels for relationship-based fraud; offenders exploit platform affordances (easy profile fabrication, public personal data, and rapid migration to private messengers) to establish and maintain deceptive relationships and then move communications to less‑traceable channels. Generative AI, deepfakes, and related automation are repeatedly identified in the corpus as plausible enablers that lower the skill and time needed to create convincing persona media (synthetic faces, voice clones, cloned or altered video) and to scale interactions; however, empirical measurement of how much AI/LLMs increase victimization, conversion rates, or aggregate losses at population scale is limited. A well-documented hybrid modality—often called pig‑butchering or romance-initiated investment fraud—combines grooming with fraudulent trading dashboards, on‑chain transfers, mixers/swaps, and recirculation; reproducible blockchain tracing and case analyses show mixers and swaps are common and associated with higher individual victim loss in traced samples. Proposed technical countermeasures include linguistic and interaction-pattern classifiers, OSINT/Maltego link-analysis workflows, and blockchain forensics; prototype evaluations report promising internal metrics in constrained datasets but face privacy, validation, language, and deployment limits. The corpus is heterogeneous: stronger contributions provide offender-conviction files, reproducible blockchain-tracing studies, and large complaint corpora, while many contributions are conceptual, small-case, or regionally limited. Identified evidence gaps that appear across the reviewed materials include rigorous, privacy-preserving real-time detection at scale; reproducible multi‑chain blockchain tracing linked to communication patterns; and empirical field tests of synthetic-media effects on victim decision‑making and detection tools. ^{[1] [2] [3] [4] [5] [6] [7] [8] [9]}

Channels, platform affordances, and movement to private messaging

Multiple empirical studies and complaint-based analyses identify mainstream social media, dating sites, and messaging services as the primary first-contact channels used by romance scammers, and researchers repeatedly document an early, deliberate migration from public-facing platforms to private messaging channels to sustain grooming and reduce public scrutiny. ^{[1] [2] [3]}

Authors show that platform affordances—easy profile fabrication, public personal data, and algorithmic visibility—are exploited to craft tailored ‘love stories’ and to plan and edit fabricated personal information; complaint‑analysis studies also report that scammers commonly redirect victims across multiple apps (often up to three) and use phone/VoIP, text, or one‑way webcams to intensify contact. ^{[10] [11] [5] [12]}

Researchers emphasize that public profile information (posts, apparent common friends, family details) is used to tailor narratives and that offenders deliberately build false infrastructure—multiple profiles, foreign phone numbers, VPNs, and email accounts—to obscure provenance and frustrate tracing. ^{[10] [13] [14]}

Grooming stages, language, and isolation tactics

Qualitative and discourse‑analytic work supports stage‑based grooming models in which repeated instant messaging, edited photographs, curated profiles, and paralinguistic cues (for example, emojis) are used to build emotional rapport and idealized impressions that lower victims’ skepticism and support stepwise extraction. ^{[12] [15] [11] [16]}

Comparative corpus studies document linguistic distinctions between scammer profiles and general-user profiles and note that scammer language is often recycled, fragmented, translated or partly machine‑generated; authors caution that general-user comparison samples may include undetected scammers, which reduces certainty about contrasts. ^{[8] [17]}

Research also reports tactics that isolate targets—encouraging moves from public platforms to private messaging, using frequent near‑real‑time text/voice contact, and combining chat with one‑way webcams or fabricated narratives that increase intimacy while limiting verification opportunities; these practices complicate independent verification and support progressive extraction. ^{[18] [17]}

Generative AI, deepfakes, and measurement limits

Several conceptual syntheses, practitioner reports, and empirical cases document that generative-AI and deepfake tools can produce photorealistic faces, cloned voices, and altered video, lowering the skill, time, and cost of producing convincing persona media and enabling synthetic profiles that need not trace to a real image source. ^{[4] [19]}

Practitioner and media-documented case evidence attributes convincing audiovisual artifacts (including at least one reported deepfake-supported incident with large reported losses) to deepfake technologies, and authors repeatedly warn that AI-synthesized images can undermine traditional provenance checks such as reverse‑image search. ^{[4] [20] [21]}

Although scholars and commentators argue that AI and LLMs plausibly enable scaling—such as supporting many simultaneous interactions or more convincing personas—there is limited empirical quantification in the reviewed corpus of the incremental, population‑level effect of AI or LLM use on victimization rates or aggregate monetary losses. ^{[22] [23] [9] [4]}

Pig‑butchering and on‑chain laundering practices

Hybrid romance‑initiated investment fraud (often described as pig‑butchering) combines grooming with fraudulent dashboards or apps that simulate trading, then moves victim funds on‑chain; multiple forensic and mixed‑method studies report use of mixers, swaps, peel chains, and recirculation to obscure flows and impede recovery. ^{[24] [4] [25] [26] [27]}

Blockchain‑forensic analyses that assembled and validated reported scammer addresses found high rates of mixer use (for example, Tornado Cash appeared in a traced sample) and reported associations between mixer use and larger individual victim losses in analysed datasets, while also noting attribution limits from incomplete seeds and reporting imprecision. ^{[6] [27]}

Authors recommend expanding reproducible multi‑chain tracing and linking transaction patterns to communication traces, and they flag that decentralization and crypto anonymity complicate cross‑border investigations and judicial evidence gathering and that some wallet‑drain malware has been documented in case reports. ^{[27] [26] [28]}

Detection prototypes, OSINT workflows, and practical constraints

Proposed detection and investigative countermeasures in the corpus include linguistic and interaction‑pattern classifiers, Naive Bayes and logistic‑regression models, OSINT/Maltego link‑analysis workflows, and combined network + NLP approaches developed to identify suspicious profiles, phishing sites, or anomalous transaction patterns. ^{[14] [7]}

One implemented OSINT + Maltego workflow reported internal experimental metrics (accuracy 85%, precision 88%, recall 80%, F1 0.84) for a logistic‑regression suspiciousness score on a constrained dataset, but authors highlight that performance depends heavily on input-data quality and relevance and recommended expanding datasets and model complexity in future work. ^[7]

Authors and practitioners consistently flag practical and ethical constraints: dynamic analysis of private conversation content intrudes on privacy, platform‑partnered evaluations are required to assess real‑world performance, and linguistic and cultural diversity reduce model accuracy unless datasets are broadened and validated. ^{[29] [14] [30]}

Methodological patterns, limitations, and priority evidence gaps

The reviewed literature is methodologically heterogeneous: contributions include qualitative crime‑script work, comparative profile‑corpus analyses, prototype detection recommendations, and reproducible blockchain‑tracing studies; these different strands inform complementary aspects of how technology enables and conceals relationship-based fraud. ^{[18] [8] [27]}

Common, explicitly reported methodological limitations across the literature include reliance on self‑reported and complaint datasets, language and platform sampling biases, imprecise or inactive victim‑reported wallet addresses in blockchain work, and constrained ability to draw population‑level inferences from case-based or convenience samples. ^{[31] [15] [27]}

Authors explicitly identify priority evidence gaps: rigorous empirical tests of synthetic‑media effects on victim decision‑making in real‑world settings; reproducible, multi‑chain blockchain tracing linked to communication patterns; privacy‑preserving, platform‑partnered detection evaluations across languages and cultural contexts; and longitudinal or population‑level quantitative studies of AI adoption by offenders. ^{[4] [27] [26] [9] [23]}

Practical recommendations and research directions stated in the corpus

Across the corpus, authors recommend interdisciplinary, platform‑partnered research that combines linguistic, behavioral, and transaction features and that evaluates detection tools in situ with attention to privacy, false positives, and cultural/language coverage; several sources also advocate for living evidence syntheses and real‑time collaboration with forensic analysts to keep pace with evolving tactics. ^{[32] [25] [33]}

Because generative tools can change offender affordances but prevalence and population‑level impacts remain uncertain, several authors call for empirical measurement of AI/LLM uptake among offenders and for evaluation studies that test detection models under realistic, adversarial conditions while protecting user privacy; the literature documents the uncertainty about how many operations currently use LLMs and related automation. ^{[4] [23] [33]}

Specific operational proposals appearing in the reviewed works include platform‑side machine‑learning detection and internal flagging systems, strengthened app-store and transaction screening, reverse-image or image-detection recommendations, and expanded reproducible blockchain-forensic capacity to assist cross‑border investigations; many authors present these ideas as recommended next steps rather than as fully validated fielded solutions. ^{[34] [35] [36] [18] [27]}

References

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.