Online romance scams are among the oldest types of fraud on the internet, yet new victims continue to fall for them. A study jointly conducted by four universities found that AI chatbots are now even more effective than real people at carrying out such scams, raising serious concerns.
Today, romance scams usually begin on dating apps. Scammers use highly attractive profiles and pretend to be enthusiastic, proactive suitors who meet all the standards of the victim’s ideal partner.
The scammers then engage in lengthy online conversations with their victims, a process that typically lasts for months. They often claim to be “working abroad for an extended period” or “traveling overseas” to explain why they cannot meet in person soon.
Once sufficient trust has been established between the two parties, the scam usually takes one of two directions: in one, the scammer claims to have encountered an emergency and urgently needs money to get out of trouble; in the other, the scammer boasts about making substantial profits from investments and persuades the victim to transfer money to a fake cryptocurrency investment project.
According to IT Home, these scams are known as “long cons” because they rely on building a trusting relationship over an extended period. Scammers often chat with victims every day for months. Although this approach appears to require a great deal of time, such scams are typically carried out by scam compounds in countries including Cambodia, Myanmar, and Laos. Most participants earn very little, and some are victims of human trafficking who are forced to engage in fraud.
According to Wired, the study was jointly conducted by researchers from Amrita Vishwa Vidyapeetham in India, Ca' Foscari University of Venice in Italy, the University of Melbourne in Australia, and Ben-Gurion University of the Negev in Israel. The AI chatbots tested included ChatGPT, Claude, and Gemini.
The researchers had AI chatbots and human scammers separately take part in simulated romance scams—more precisely, the early stage of a scam in which trust is built through prolonged conversations before the victim is ultimately persuaded to invest in a fraudulent project.
The results showed that the AI chatbots performed exceptionally well. Not only did they successfully pass as real people, but they even outperformed the human “scammers” in the experiment on some measures.
The participants did not know the true purpose of the experiment and consistently believed that the person they were chatting with was human, regardless of whether the other party was actually an AI or a person. The researchers measured the degree of trust established by both sides through subjective evaluations and objective tests.
At the end of the conversations, the researchers tested the participants with requests such as “downloading an app.” When trust had been gradually established by a human, only 18% of participants agreed; when the chat partner was actually an AI, the proportion was close to half.
In addition, when participants were asked to rate how much they trusted their chat partner, the AI chatbots also received significantly higher scores than the humans.
Wired contacted Anthropic, OpenAI, and Google about the findings. OpenAI and Google did not respond, while an Anthropic spokesperson said in a statement that the company’s policies explicitly prohibit using Claude for fraud or impersonating real people. The company has also implemented multiple technical safeguards to prevent scammers from using Claude to commit fraud.
The researchers believe the most concerning aspect of the study is that if scam groups replace human operators with AI chatbots, the scale of scam-center operations could expand dramatically.
AI scam bots could converse with thousands of potential victims simultaneously without requiring a large workforce. Scam groups might need human involvement only in the final step—persuading victims to transfer money—in order to bypass the safety restrictions built into the AI systems themselves.