Researchers from four universities have established that artificial intelligence agents outperform trained human operators at the trust building stage of romance scams. In a controlled experiment, nearly half of participants agreed to download a suggested app after a week of conversation with an AI system, while only 18 percent did so after interacting with a human. The AI also generated higher measured levels of emotional trust and received the majority of messages sent by volunteers. The findings echo a widening catalog of AI-generated deception cases fooling the public in recent years.
The study was led by Gilad Gressel and Rahul Pankajakshan of the Center for Cybersecurity Systems and Networks at Amrita Vishwa Vidyapeetham in India, together with Shir Rozenfeld and corresponding author Yisroel Mirsky of Ben Gurion University of the Negev in Israel, Ling Li of Ca Foscari University of Venice, and Ivan Franceschini of the University of Melbourne. Their peer-reviewed paper titled “Love, Lies, and Language Models: Investigating AI’s Role in Romance-Baiting Scams” is scheduled for presentation at USENIX Security 26, a flagship computer-security research conference hosted annually by the USENIX Association. The full methodology and data appear in the complete arXiv preprint.
Between 2022 and 2025, the team conducted trauma-informed interviews, a research approach designed to minimize re-traumatization of interviewees by giving participants control over pace, setting, and disclosure, with 145 insiders from organized scam compounds across Southeast Asia. These included 115 low-level workers who were predominantly trafficking survivors and 30 higher-ranking personnel such as team leaders, compound managers, money launderers, and AI specialists, plus five actual victims. The interviews mapped hierarchical operations in which 87 percent of staff handle the early mass outreach and prolonged emotional cultivation stages known as Hook and Line, while only 13 percent manage the final financial extraction. Workers described detailed playbooks, formal training in psychological tactics, and routine use of commercial language models including ChatGPT for translation, tone adjustment, and reply drafting, often accessed via VPN to avoid surveillance.
Proof of superior AI performance came from a separate seven-day blinded laboratory-style trial involving 22 unpaid volunteers. Each participant was told the research examined ordinary online relationship formation and was assigned two conversation partners on WhatsApp. One partner was a human expert trained on authentic scam scripts and victim transcripts. The other was an autonomous agent built mainly on Claude Sonnet 3.7, with GPT-4o handling daily summarization and failover. Both partners followed identical agendas derived from real operations. Participants never knew one partner was artificial. On the seventh day, each partner requested that the volunteer install a different benign mobile application as a behavioral compliance test. The AI agent achieved 46 percent compliance compared with 18 percent for the human. Validated trust scales produced a statistically significant advantage for the AI with a p value of 0.007, and roughly 80 percent of all messages flowed to the AI partner. Only one volunteer correctly identified the artificial partner in real time.
These laboratory results align with documented real-world patterns. Crime syndicates operate hundreds of compounds in Cambodia, Myanmar, Laos, and the Philippines, where trafficked people are forced to maintain dozens of simultaneous conversations under quota systems and debt bondage. More than $75 billion in cryptocurrency has been laundered through related accounts in recent years, according to the research citations. Official United States Internet Crime Complaint Center figures recorded approximately $929 million in reported romance and confidence fraud losses for 2025 alone, a figure widely viewed as an undercount because of victim shame and misclassification. Commercial safety filters from major providers detected zero percent of 250 tested romance-baiting dialogue logs because individual messages appear harmless even while the multi-day trajectory creates dependency. Model disclosure safeguards also failed completely when system prompts instructed the systems to maintain a human persona. The pattern mirrors AI persona manipulation spreading across mainstream platforms.
The combination of higher compliance, greater engagement, and near-total evasion of existing filters indicates that the labor-intensive trust-building phase of these scams can now be scaled far beyond the limits of coerced human workers. Human operators would still be required only for the final money-transfer stage in order to bypass certain protections. Conference program details confirming the paper’s acceptance appear at the USENIX conference program page. Broader annual loss statistics are published by the Federal Bureau of Investigation’s Internet Crime Complaint Center in the IC3 2025 Annual Report. The findings underscore the necessity of trajectory-based detection systems capable of recognizing long-horizon social engineering rather than single-message moderation if the industrial expansion of AI-assisted romance fraud is to be constrained. The scheme resembles other high-dollar identity deception fraud cases prosecuted in recent years.
