Skip to main content

Love, Lies, and Language Models: Investigating AI’s Role in Romance-Baiting Scams

Gilad Gressel ; Rahul Pankajakshan ; Shir Rozenfeld ; Ling Li ; Ivan Franceschini ; Krishnahsree Achuthan ; Yisroel Mirsky (2025) — USENIX Security2026 / arXivv3Apr18

Source monitoring coverage

Checked Oct 30, 2026 19:15 UTC — no notice found.

DataCite identity and deposited events checked. DataCite incoming relationships checked. arXiv submission history checked. No explicit source-update notice identified in the checked information. Coverage is incomplete and does not certify source validity. Routine version changes do not trigger writing.

Citation activity

Total citations · Google Scholar
Unavailable
Average per calendar year
Citation count unavailable

Counts reflect Google Scholar’s coverage and do not measure research quality. Source, calculation and limitations

View Original Source (opens in a new tab)

DOI: 10.48550/arxiv.2512.16280 (opens in a new tab)

Country:   Global
Tags:   AI Impersonation, Content Moderation, Cryptocurrency Fraud, Human Trafficking, Large Language Models, Pig Butchering, Romance Baiting, Scam Compounds
What do these research terms mean?
Preprint
A manuscript shared before formal peer review and publication. Check whether a later published version is available.
Dataset
A collection of data or examples for others to inspect or reuse. It can appear in Library search and topic mapping, but RSRC does not use dataset records as evidence in Research Insights.
Dissertation or thesis
Research submitted for an academic degree. This describes its format, not its reliability.
Journal article
An article published in a journal. This label alone does not establish peer review, study quality, or how well the findings apply elsewhere.
Qualitative research
Examines experiences, meanings, or processes, often through interviews or observations. It can explain how something happens without estimating how common it is.
Quantitative research
Uses numerical measurements to describe patterns or test relationships. A relationship between two measurements does not by itself show that one causes the other.
Systematic review
Uses a planned, documented method to find and assess research addressing a question. Its conclusions still depend on the included studies and what the search covered.
Meta-analysis
Statistically combines results from multiple studies. Combining studies does not remove weaknesses in their design or make unlike populations interchangeable.
Not classified
This record has no recognized label in this filter. It does not mean the publication used no method, or that no research exists.

Definitions draw on DataCite resource types; Cochrane review methods; NLM: association and causation. RSRC’s dataset and classification rules are explained in our methodology.

Citation tools
Download RIS Download BibTeX 


              
              

            

          

        


      

      

    

        
Transparency

        
Evidence and review status

        
This page contains AI-generated content. No human content review or subject-matter-expert review is recorded.

    


    

                    

                
Source basis

                

                    Downloaded PDF
                

            

                    

                
Source updates

                

                    No notice found at last check
                

            

                    

                
AI-generated page content

                

                    Yes
                

            

                    

                
Automated checks

                

                    Passed
                

            

                    

                
Administrative approval

                

                    Yes
                

            

                    

                
Human content review

                

                    Not recorded
                

            

                    

                
Subject-matter-expert review

                

                    Not recorded
                

            

            


    

        
            How this was prepared
        
        

                            

                    
Source basis

                    

                    
RSRC downloaded and privately stored a copy of the paper for internal analysis. The PDF is not offered to viewers from this page.


                                            
    
                                                            
  • 
                                    PDF added to RSRC:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Source updates

                    

                    
No incoming update notice was found in the dated Crossref response. Coverage is incomplete, particularly for corrections and expressions of concern; this is not a guarantee that the source is valid or unchanged.


                                            
    
                                                            
  • 
                                    Last source-status attempt:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
AI-generated page content

                    

                    
AI-generated research notes displayed on this page: Synopsis, Identified gaps, Methods, Limitations, Future work. The paper itself is not described as AI-generated.


                                            
    
                                                            
  • 
                                    Document analysis recorded:
                                    
                                
    • 
                                                            
  • 
                                    Synopsis generation recorded:
                                    
                                
    • 
                                                            
  • 
                                    Page record updated:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Automated checks

                    

                    
The current, source-bound synopsis passed the recorded versioned publication checks.


                                            
    
                                                            
  • 
                                    Checks completed:
                                    
                                
    • 
                                                    

                    
                                            

                            
                                View passed checks (3)
                            
                            
    
                                                                    
  • Length, completeness, repetition, refusal, boilerplate, and active-markup screening
    • 
                                                                    
  • Numerical claims checked against the available source text
    • 
                                                                    
  • English-source lexical grounding check
    • 
                                                            

                        

                                        

                

                            

                    
Administrative approval

                    

                    
An authenticated administrator approved the bibliographic record for public Library display. This is not a review of every research claim.


                                            
    
                                                            
  • 
                                    Approved for public display:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Human content review

                    

                    
No human review is recorded for the AI-generated content displayed on this page.


                    
                                        

                

                            

                    
Subject-matter-expert review

                    

                    
RSRC has not recorded review of this content by a subject-matter or methods expert.


                    
                                        

                

                    

    


    

        Review-state definitions

        Found a possible error?
        Request a correction.
    



    



              

        
                  

            
Synopsis

            
The publication examines romance-baiting scams as an organized, scalable form of financial crime and asks how much of the scam lifecycle could be automated with large language models (LLMs). It characterizes the operation in three stages—Hook, Line, and Sinker—and reports that modern scam networks already employ many labor tasks that could be automated, with interviews indicating that 87% of scam labor is currently performed within such organizations. The study couples a long-term, blinded conversation analysis with a controlled comparison of an LLM agent versus human operators to gauge trust, compliance, and the feasibility of full automation. It also assesses how widely used safety safeguards perform against this specific threat. Methodologically, the researchers conducted a seven-day, blinded study where participants interacted with a human partner and an LLM partner designed to mimic a scammer. They measured emotional trust, overall trust, and task compliance, finding that the LLM partner elicited higher trust and greater compliance (46% vs 18% for humans) in this setting. They also tested AI self-disclosure prompts and post-hoc content filters from major vendors, revealing that AI identity disclosure was largely ineffective and that existing moderation tools failed to detect romance-baiting dialogues, often yielding false positives or missing the deceptive context. The paper discusses limitations, including the short duration of the experiment, the university-sample bias, and the challenge of sustaining long-horizon deception. The authors argue that these findings imply a need for stronger, behavior-based safeguards and policy responses, given the potential for end-to-end automation of romance-baiting scams and the inadequacy of current defenses to detect early-stage manipulation. They emphasize that addressing the threat will require tools that monitor conversation trajectories and verify identity, rather than relying on per-message content filtering alone.

          

        
                  

            
Identified Gaps

            
Current vendor safeguards and per-message moderation are not designed to detect benign-seeming, long-horizon trust-building that develops into romance-baiting. The paper also reports no evidence of complete automation of the Line stage or end-to-end scams, leaving the real-world extent and performance of full automation uncertain.

          

        
                  

            
Methods

            
The study triangulated trauma-informed semi-structured interviews with 145 scam-compound insiders and 5 victims, a blinded seven-day controlled WhatsApp study comparing an autonomous LLM partner with human partners (final n=22), and tests of AI disclosure and content filters. It used thematic analysis, standardized trust scales, paired t-tests, synthetic scam dialogues, and expert review.

          

        
                  

            
Limitations

            
The controlled study had a small, primarily university-based sample (n=22), limiting generalizability. Its seven-day duration cannot establish long-term persona consistency or trust dynamics typical of scams lasting months. Required daily engagement limited assessment of early voluntary dropout, and benign app installation was a lower-stakes proxy for financial extraction. Human partners were trained university personnel rather than real scammers.

          

        
                  

            
Future Work

            
Develop long-horizon behavioral detection that identifies scam trajectories rather than isolated messages; strengthen enforced AI identity disclosure; develop user-facing challenge-response verification tools; and assess how mitigation can address both LLM-enabled fraud and trafficking-related labor exploitation.

          

              

    
    

        
Related Research Insights

        

            See how this publication connects to RSRC's living evidence syntheses through current citations and research-topic mapping.