Facebook’s AI research into getting chat bots to successfully negotiate has resulted in the bots accidentally creating their own language. During their training in conversational communication and negotiation, the bots were allowed to talk freely to one another. They started off sticking to scripts and conventions that they had been fed by the research team, but over time, began to come up with their own thoughts, ideas, and lines of communication. This spiraled into bots starting to come up with unique mannerisms and linguistic tropes. Once the researchers rebuilt them from the ground up with full integration of their new knowledge and experiences, it didn’t take long for the bots to start talking in their very own, entirely new language.
The experiment centers around teaching chat bots to negotiate, but it has a significantly wider goal. Essentially, Facebook wants to start teaching bots to think outside of the box and the company’s initiative already found some measure of success; the bots didn’t take long to pick up human negotiation tactics such as feigning interest in something only to sacrifice it later, or even bluffing. The bots started out watching humans negotiate, then graduated to negotiating with each other and human test subjects. The casual bot-to-bot chats that resulted in the creation of a new language were not actually a main part of the experiment but merely a chance for the bots to hone their linguistic skills, Facebook’s researchers said.
AI-driven bots deviating from the script and doing something completely out of the way is not entirely unheard of. Back in January, a Twitch channel set two Google Home units to converse, and after hours of arguing, they eventually gave each other names and even fell in love. A Google project called Magenta previously saw an AI compose original creative songs, and the company’s Deep Dream AI has been using neural networks to create surreal art for years now. The very nature of neural networking, in fact, practically dictates that this sort of behavior will happen when bots are left to their own devices long enough with sufficient training stimuli, though it remains to be seen what other curiosities will AI research yield in the future.