AI Chatted With Scammers 2.5 Million Times and Pulled Out 250,000 Leads

Scam calls are so common in China that even iPhone owners switch on system-level filtering of unknown numbers and accept that couriers sometimes get cut off along with the spam. A project covered by Jiqizhixin came at the problem from another side: an AI answers the call. According to the outlet, the system held about 2.5 million conversations and pulled 250,000 pieces of intelligence on fraud schemes out of them. That is one lead per ten conversations.

Why blocking loses

The asymmetry is obvious. Changing a number costs a scammer almost nothing: block one today, and they call from another tomorrow. The user has to react by hand every time, or rely on a filter that errs in both directions. The author of the piece describes exactly this experience: you block a number, and a minute later a call comes in from a new one.

The project changes the underlying logic. The call is not discarded; it becomes a conversation that yields something useful. While the scammer spends time on the AI, he is not working a real person. The byproduct of that wasted time is those 250,000 pieces of information.

What is known and what is not

The available excerpt of the publication contains no technical details. It is unclear what model powers the conversational agent, how the voice layer works, or who operates the system. It also does not say what exactly counts as "intelligence." It could be phone numbers, conversation scripts, platform names, payment details or links, but without the full text that is only a guess.

The public data boils down to this:

• about 2.5 million conversations with scammers;
• about 250,000 collected pieces of information on fraud;
• average yield: roughly 0.1 pieces per conversation;
• format: the AI picks up incoming calls and keeps the conversation going instead of dropping the call.

How to read these numbers

Ten percent looks modest, but from an intelligence standpoint it is a lot. No conversation costs any human hours, and the scale reveals patterns that no individual victim can see: which scenarios are in use right now, which lines the calls come from, how the cover stories change.

The figure has weak spots. A "piece of information" is not defined, and the 250,000 almost certainly include duplicates. It is unknown how the data is verified, or whether it is passed to telecom operators, banks or law enforcement. The legal status of recording the conversations is also an open question. Finally, scammers adapt. If they learn to spot a bot quickly, for example by its pauses or unusual replies, the yield will drop, and the conversational agent will have to become more convincing.

What to do about it

For users, the practical takeaway is simple: you still should not answer unknown numbers, and filters and number reports remain the basic defense. If conversational traps start working at scale, they will give banks and operators the data to shut down schemes before those reach people.

Teams building anti-fraud systems should borrow the approach itself and measure three things: how long the AI keeps a scammer on the line, what share of the collected data is unique, and how well it is confirmed by independent sources. Those metrics will show whether the system works as intelligence gathering or just generates a large volume of noise.

For now, the public record has two numbers and the idea itself. If the authors disclose their methodology, it will become clear whether the approach can scale beyond China.