Two of Facebook’s AI bots recently made the news for apparently developing a new language that their human overseers couldn’t understand. The bots had been instructed to trade with each other to swap bats, balls and books, using their conversational exchanges to improve their negotiation skills. However the researchers running the experiment forgot to specify that the exchanges had to take place in English, and the bots’ conversations quickly drifted into what looks like gobbledegook as they competed to get the best deal:
Bob: “I can can I I everything else.”
Alice: “Balls have zero to me to me to me to me to me to me to me to me to.”
Despite the bizarre appearance of these exchanges, they continued to make sense to the bots, who successfully completed some negotiations in this mode. Researchers think they may have effectively created a form of shorthand.
“[AI] Agents will drift off understandable language and invent codewords for themselves,” says Facebook researcher Dhruv Batra. According to other researchers at OpenAI, the phenomenon has been observed before in other multi-agent environments, where it is more efficient for bots to speak to each other in a kind of Morse code.
Facebook’s bots were eventually redirected to speak in English, as the company’s researchers are more interested in having bots that can talk to people. However the episode has shed interesting light on the more niche research area of leading machines to develop their own languages.
Allowing AIs to speak to one another in their own codes and languages could potentially bring benefits, such as improving the interoperability of software, apps and services. However the trade-off would be losing our human understanding of inter-machine communication. As Batra says, “There aren’t any bilingual speakers of AI and human languages.” Would the loss of transparency be worth it?