Study: AI Chatbots Are 2.5x More Effective Than Human Scammers at Building Victim Trust

Researchers from four universities ran a controlled study pitting an AI chatbot against an experienced human romance-scam operator in a simulated "pig butchering" scam — the long-con format that builds trust before pivoting to a fake cryptocurrency investment pitch — and found the AI chatbot didn't just compete with the human scammer, it clearly outperformed them.

On a 1-to-5 trust scale, test subjects rated the AI chatbot, built on Claude, at an average of 3.78 versus 3.31 for the human scammer, despite the human being an actual expert in romance fraud recruited specifically for the comparison.

The gap showed up even more sharply in actual behavior, not just stated trust: nearly half of subjects complied with a request from the AI chatbot to download an app, compared to fewer than one in five who complied with the same request from the human scammer.

Researchers put the overall effectiveness gap at roughly 2.5x, concentrated specifically in the early, labor-intensive trust-building phase of the scam — the part that has traditionally required a skilled human operator investing real time per target.

That's the part of the finding worth sitting with: the study isn't showing AI can write a more convincing single message, it's showing AI can automate the sustained, patient relationship-building that used to be the bottleneck limiting how many victims one scammer could run at once — remove that bottleneck and the same scam operation scales differently.

Standard advice — watch for bad grammar, urgency, inconsistent stories — was built for scam messages that had those tells. A model trained on genuinely good conversation doesn't reliably have them, which means the practical defense shifts toward verifying identity out-of-band, a real video call, a known phone number, rather than trying to spot the tell in the conversation itself.