AI Now Automates 'Grooming' Phase of Romance Scams, Researchers Find
EU-funded security researchers reveal that LLMs are bypassing safety filters to build emotional trust with victims at scale.
Organized crime groups are now using Large Language Models (LLMs) to automate the initial stages of romance fraud, significantly increasing the scale and effectiveness of these attacks. Findings presented by EU-funded security researchers at the 35th USENIX Security Symposium reveal that AI is now handling the critical emotional labor required to lure victims.
According to the research paper "Love, Lies, and Language Models: Investigating AI's Role in Romance-Baiting Scams," criminal enterprises have integrated AI into a three-stage process: the "hook" (initial contact), the "line" (building trust and grooming), and the "sinker" (the final financial robbery). The study indicates that 87% of the labor in scam compounds—specifically the hook and line stages—consists of tasks susceptible to LLM automation. While the final "sinker" stage still requires human managers to execute the theft, the initial grooming is becoming highly automated.
The Failure of AI Safety Filters
The shift toward AI-driven fraud is facilitated by a systemic failure in current AI safety guardrails. During researcher testing, safety filters from major providers, including OpenAI, Google, and Meta, detected 0.0% of romance-baiting dialogues in a sample of 250 conversation logs. Because the conversations appear as benign, empathetic human interaction, they bypass filters designed to catch explicit malice or hate speech.
Yisroel Mirsky, head of the Offensive AI Research Lab at Ben-Gurion University, noted that while LLMs can perform these tasks alone, some engineering is required to build "agent harnesses" so the output feels more human. A Europol spokesman added that the same qualities making AI helpful—accessibility, adaptability, and sophistication—also make it a powerful tool for criminal networks.
Escalating Financial Impact
This automation arrives as financial losses from romance scams reach record highs. In the United States, the FBI reported that romance scam losses in 2025 totaled approximately $929 million across 23,159 cases, a sharp increase from the $672 million reported in 2024.
Historically, these operations have relied on physical scam compounds utilizing trafficked labor in Africa (Nigeria, Ghana) and Southeast Asia (Myanmar, Cambodia, Laos). However, the transition to AI reduces the reliance on these physical hubs, making criminal operations more resilient to law enforcement raids and easier to scale globally.
What's Next
The research suggests a growing gap between the capabilities of offensive AI and the defensive filters provided by major tech companies. As criminal groups refine their "agent harnesses," the ability to distinguish between a genuine romantic interest and an automated bot will diminish. Industry observers are now watching whether AI providers will implement more nuanced behavioral analysis to detect the patterns of "love bombing" and grooming that current keyword-based filters miss.