Social media giant Facebook terminated an artificial intelligence program that created its own language.
In trying to train AI bots to negotiate, researchers at Facebook Artificial Intelligence Research (FAIR) found that AI bots took a different turn: they created a new language. Two Facebook AI bots had the following conversation:
AI bot #1: i can i i everything else . . . . . . . . . . . . . .
AI bot #2: balls have zero to me to me to me to me to me to me to me to me to
AI bot #1: you i everything else . . . . . . . . . . . . . .
AI bot #2: balls have a ball to me to me to me to me to me to me to me
“There was no reward to sticking to English language,” Dhruv Batra, visiting research scientist from Georgia Tech at FAIR, told Fast Co. Design. “Agents will drift off understandable language and invent codewords for themselves.”
According to Batra, the language developed by the Facebook AI bots is not entirely different from the languages developed by humans. Saying the word “the” five times, could be interpreted as wanting five copies of something. Similar to human communities, AI bots have created with what researchers call as “shorthands”.
Ultimately, Facebook researchers put a stop to AI bots’ practice of talking to each other in their own language and require them to communicate in plain English language instead. “Our interest was having bots who could talk to people,” Mike Lewis, research scientist at FAIR, told Fast Co. Design.
Another reason that researchers halted AI bots’ practice of communicating in their own language is that humans have no means of understanding this new language. “It’s important to remember, there aren’t bilingual speakers of AI and human languages,” Batra said.