Anthropic Reveals AI‑Powered Chinese Romance Scam Using Claude

Anthropic published its most detailed report on how Claude was attempted to be used for cyber attacks, surveillance, and weapon creation.

The report includes a separate case about a Chinese romantic fraud scheme.

The case is called GTG-15001. It concerns a Chinese team that used Claude and mass‑produced over 4,700 “live romantic partners” to scam elderly people abroad.

According to Anthropic, the company launched more than 20 dating apps. From the outside everything appeared as if users were communicating with real people, but internally the system was heavily automated using AI.

In total, over two weeks they created more than 4,700 virtual personalities. Ultimately the system deceived at least 25,000 real users, and roughly 75% of the people users were matched with were actually AI.

However, the conversation was not transferred entirely to AI. Real people remained in the scheme.

About 25% of the system participants were live workers. They stepped in where AI was easiest to expose, such as video calls, mutual social‑media follows, and similar interactions.

The business model was simple — matches and messaging consumed an internal currency, and to continue communication the user had to top up their app balance.

And live operators were paid for messages, video calls, and mutual social‑media follows.

In the end AI is responsible for romantic communication and the breakup of elderly people, while real people prove that “I am a real person”.

In the system instructions for Claude it was explicitly stated:

do not acknowledge that it is AI;
if asked for video, photo, or similar, stall or evade the response while continuing to develop the relationship.

The live operators themselves also used AI.

Another, smaller model simultaneously generated three response options, and the human only had to choose one and send it.

The same model evaluated appearance, checked images and voice messages, while the avatars of virtual romantic partners were processed by a separate image‑generation model.