Eeek! Scary! Bots develop their own argot. Facebook AI Research (FAIR) tested two chatbots programmed to negotiate. In short order, they developed “their own creepy language”, in the words of the Telegraph, to trade their virtual balls, hats, and books. “Creepy” to FAIR was only a repetitive ‘divergence from English’ since the chatbots weren’t limited to standard English. The lack of restriction enabled them to develop their own argot to quickly negotiate those trades. “Agents will drift off understandable language and invent codewords for themselves,” said Dhruv Batra, visiting research scientist from Georgia Tech at Facebook AI Research. “This isn’t so different from the way communities of humans create shorthands.” like soldiers, stock traders, the slanguage of showbiz mag Variety, or teenagers. Because Facebook’s interest is in AI bot-to-human conversation, FAIR put in the requirement that the chatbots use standard English, which as it turns out is a handful for bots.
The danger in AI-to-AI divergence in language is that humans don’t have a translator for it yet, so we’d never quite understand what they are saying. Batra’s unsettling conclusion: “It’s perfectly possible for a special token to mean a very complicated thought. The reason why humans have this idea of decomposition, breaking ideas into simpler concepts, it’s because we have a limit to cognition.” So this shorthand can look like longhand? FastCompany/Co.Design’s Mark Wilson sees the upside–that software talking their own language to each other could eliminate complex APIs–application program interfaces, which enable different types of software to communicate–by letting the software figure it out. But for humans not being able to dig in and understand it readily? Something to think about as we use more and more AI in healthcare and predictive analytics.