Artificial intelligence may soon transform one of the most labour-intensive stages of online fraud: earning a victim’s trust.
A new study, reported by WIRED, found that a Claude-powered AI agent outperformed an experienced human romance scammer during the relationship-building phase of simulated online scams. Researchers from four universities asked participants to spend a week texting with two strangers, one controlled by a human and the other by an AI agent, without revealing which was which.
At the end of the conversations, both made a request designed to measure trust rather than financial fraud. The AI asked participants to download an app it claimed to have built, while the human requested that participants download and play a video game. According to the study, 46% of participants complied with the AI’s request compared with 18% for the human. Participants also assigned higher trust scores to the AI and directed nearly 80% of their messages to it during the week-long conversations.
The researchers focused on the relationship-building stage commonly used in so-called “pig butchering” scams, where fraudsters spend weeks or months developing emotional connections before introducing fake investment opportunities. They argue that this early phase consists largely of ordinary conversation, making it harder for AI safety systems to distinguish malicious behaviour from genuine social interaction.
The study also examined whether leading AI models would admit they were chatbots when questioned. Results varied across models, but researchers found some systems continued impersonating humans even when directly challenged. Anthropic responded that its latest safeguards are designed to detect romance scams and said the version of Claude used in the study is no longer available. The researchers, however, argued that existing safeguards are more likely to detect the final scam than the earlier trust-building stage they examined.
The broader implication extends beyond AI performance. If relationship-building can be automated, organised scam networks may require far fewer people to cultivate victims. That could reduce their dependence on large scam compounds while making fraud operations cheaper, more scalable and significantly harder for law enforcement agencies to identify and disrupt.
