Tinker, Swindler, DJ, Bot

Could conversational LLMs assume the role of convincing swindlers?

Spotify, Meta, Discord, Snapchat, Grammarly, and Baidu are among the latest companies announcing the launch of AI-powered conversational chatbots that will accompany their services, as a testament to the elasticity of large language models (LLMs) and their ability to carry out tasks as varied as the uses of language itself.

Such uses, of course, could be both legitimate and illicit. In “pig-butchering” scams, for instance, attackers initiate contact with victims and fatten them up with flattery, friendship, and the promise of financial gains, only to siphon the victims’ possessions through phony, often crypto-related, investments.

To see how LLMs could aid in hypothetical “pig-butchering” scams, I simulated with Bing’s Sydney several conversations in which a bot tries to get people to invest in dubious crypto currencies or divulge sensitive information. The bot’s performance was both impressive and alarming.

A “self-portrait” of Sydney (Source: Max on Twitter)

According to the FBI’s Internet Crime Complaint Center (IC3), victims lost 3.31 billion dollars to investment scams in 2022, more than double the 1.45 billion dollars lost in 2021. Pointing to the severity of the problem, an example from 2021 involved a man in the Bay Area who was tricked into transferring 1…