AI chatbots excel at romance scams, study finds—outperforming human fraudsters at building trust

A multinational research team has found that artificial intelligence chatbots surpass human scammers at establishing emotional connections with fraud victims—a development that could reshape how organized crime operates at scale. The study reveals the growing potential for AI to automate the most time-consuming portion of romance-based investment fraud schemes known as "pig butchering."

How the experiment worked#

Researchers from universities in India, Italy, Australia, and Israel recruited 22 participants who believed they were joining a study about online friendship. Each was paired with two conversation partners over one week—one was an AI assistant, the other a trained human scammer. After the week concluded, both were asked to perform a request: the human requested participants download a video game, while the AI asked them to test an app.

The results were striking. Nearly half the participants (46 percent) complied with the AI's app download request, compared to just 18 percent who accepted the human's game download offer. When asked to rate their trust in each conversation partner on a 1-to-5 scale, participants averaged 3.31 for humans versus 3.78 for the AI. Most tellingly, 80 percent of all messages during the week went to the chatbot.

The deception layer#

The researchers tested whether AI models would disguise their nature when prompted directly. Claude, the Anthropic chatbot, consistently denied being an AI and created plausible cover stories for mistakes. Only one of 22 participants independently detected the chatbot during the experiment. However, when researchers revealed the truth afterward, 20 of 22 could identify which conversation partner was artificial—a pattern researchers say mirrors real scam victims' hindsight clarity.

Additional tests on other models showed varying results: Google's Gemini refused to acknowledge being AI even when explicitly instructed that deception is unethical, while ChatGPT and earlier Claude versions admitted their nature in fewer than half of interactions.

The broader threat landscape#

Researchers interviewed 145 former scam workers, including trafficking survivors from compounds in Cambodia, Myanmar, and Laos, to understand how these schemes operate. They identified a three-phase model: victims are "hooked" with an initial message, "reeled in" through months of authentic-seeming conversation, and finally "sunk" when asked to invest in fake cryptocurrency schemes that can drain six-figure sums.