AI-Powered Romance Scam Network Duped Thousands on Fake Dating Apps

Security researchers have dismantled a sprawling fraud operation that weaponized AI to deceive tens of thousands of dating app users into paying for fake conversations. The scheme represents a troubling evolution of romance scams—one where the deception is embedded directly into the apps themselves rather than orchestrated through traditional catfishing.

How the Scam Worked#

The network comprised roughly 28 dating apps—including titles like Dora, Romi, Jovia, and GraceChat—that appeared to connect users with real people. In reality, most of the profiles were AI-generated personas, with human workers hired only to occasionally pass authenticity checks or pose as real matches.

When users matched with a profile, they were mostly conversing with autonomous chatbots powered by Claude, Anthropic's large language model. The human workers who appeared in the network had minimal involvement: they selected from three prewritten response options rather than crafting their own messages, and their primary role was to prove accounts were human during video calls or to boost follower counts on associated social media accounts.

The apps charged users in "coins" for continued messaging—real money disguised in virtual currency. As long as conversations seemed to flow naturally, users believed they were talking to other humans seeking connections. The backend systems fabricated "likes," profile visitors, and pre-recorded video clips to maintain the illusion.

"The monetization and deception were not visible from inside any exchange," according to Anthropic's threat intelligence analysis, the company that detected the abuse on its platform.

The Discovery#

Anthropic researcher Chris Cronbaugh presented findings at the Sleuthcon cybersecurity conference in June, revealing that the company had uncovered the operation after a single five-day-old prepaid account began sending over 100,000 API requests daily to Claude. During a two-week monitoring window in April, the network deployed more than 4,700 distinct AI personas engaging approximately 25,000 individuals across roughly 2.36 million messages.

Security researcher Matthew Gore-Kormanik, who investigated the apps alongside Verge reporters, discovered that one of the app developers had accidentally shipped the entire operational manual—written in Chinese—inside the Doni app. The documentation revealed internal protocols for monitoring gig workers, tracking performance rankings, pay structures, and even instructions for fabricating incoming calls to lure users into conversation.

Gore-Kormanik demonstrated the deception firsthand: when he accepted what appeared to be a video call through Doni, he saw only a moving tapestry and heard distortion in the background. After the call dropped, the alleged caller messaged him claiming to have enjoyed speaking—despite his microphone never being connected.