Artificial Intelligence (AI) is the new buzzword. As we begin to make progress towards human-level AI, several developments and breakthroughs in the field of AI have begun to make waves. From self-driving cars to systems that understand human language, we have come a long way. But then, the eternal fear of AI surpassing human intelligence continues to haunt many a scientists and developers. Recently Elon Musk, CEO of Tesla, expressed concerns over AI dominance and told US governors that they should act to regulate AI. Whether Elon Musk’s fears of AI dominance are unfounded or not is a debate for another day. But in what appears to be rather “spooky”, it has come to light that chatbots began to converse in their own strange language, before being shut down.

image source: datafloq

Facebook shuts down AI system after chatbots began conversing in strange language

Facebook recently shut down an AI system it was working on. Not because they weren’t making progress, not because it was technically challenging, but because the AI stopped using English and began conversing in its own language. The conversation was taking place between two AI agents developed inside Facebook.

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.”

Well, this doesn’t make sense. In fact, for us humans, it’s nonsense. But what appears to be nonsensical to us turned out to be a “discussion” between two agents that are part of the one of the most sophisticated negotiation softwares ever created. Initially, the agents were communicating in pure English but later began to drift towards using their own language. The chatbots appeared to be communicating with each other in a language that sounded plain “gibberish” but one the computers understood. Not only did the system understand the new language it created, it also used it to communicate with other AI systems to complete the tasks it was assigned! Spooky!

image source: Facebook

“There was no reward to sticking to English language. As these two agents competed to get the best deal–a very effective bit of AI vs. AI dogfighting researchers have dubbed a “generative adversarial network”–neither was offered any sort of incentive for speaking as a normal person would. So they began to diverge, eventually rearranging legible words into seemingly nonsensical sentences” says Dhruv Batra, visiting research scientist from Georgia Tech at Facebook AI Research.