Romance scam scheme: thousands of AI bots posed as lovers on dating sites

Anthropic, the company behind the Claude AI model, has uncovered a large-scale scam operating in the online dating world: a Chinese studio ran more than 20 dating apps in which a substantial share of the messages were written not by real people, but by AI characters. The findings were revealed in the company's report on AI misuse.

According to Anthropic, over a two-week period in April this year, more than 4,700 distinct AI bots were identified, communicating with at least 25,000 unique users. During that time, Claude generated roughly 2.36 million messages. It remains unclear whether the operation targeted users within China itself or people abroad.

Real people hiding among the bots

The scheme's operators didn't rely solely on chatbots — some profiles were run by actual humans, though in far smaller numbers. Anthropic estimates the ratio at roughly three AI bots for every one real person. Users themselves had no way of knowing whether they were talking to a person or a program.

Real people were hired as gig workers to handle tasks a bot couldn't convincingly fake — such as video calls or following a target on social media. These "live" contacts were meant to ease victims' suspicions. The human workers, in turn, were assisted by another AI tool that suggested ready-made short replies. Pay depended on the number of messages sent, video calls made, and mutual social media follows secured.

Bots were forbidden from admitting they were bots

The automated characters operated under strict rules: they were barred from revealing they were software and instructed to dodge requests for video calls or photos. Conversations followed pre-set scripted stages, allowing thousands of profiles to be managed simultaneously with almost no human involvement.

The system also flagged users who began to suspect they were chatting with a bot. And if no real person was available as backup, the software could generate fake likes and profile visits on its own.

A division of labor among multiple AIs

Notably, the scammers didn't rely on a single model but deployed several AI systems, each with a distinct role. Claude handled autonomous conversations with users, while a smaller, less powerful model drafted short reply options for the hired workers. Additional models were used to rate facial attractiveness, moderate photos and voice recordings, and generate avatars.

This, Anthropic says, is the key feature of the scheme: it wasn't a single AI generating fraudulent messages, but an entire process split between multiple models and real people working in tandem.

Source: novinky.cz