Facebook's "secret chatbot language" was a failed negotiation experiment that nobody shut down

Two Facebook chatbots did not invent a secret language, and Facebook did not pull the plug on them. In June 2017, Facebook AI Research trained a pair of bots to negotiate using reinforcement learning. One was taught to mimic how people bargain in English and turned out to be a weak negotiator, too willing to accept bad terms. The other was told to maximize its score, negotiated better, and drifted into output humans could not parse. TechCrunch described the experiment on June 14; The Atlantic picked up the phrase "divergence from human language" the next day. By the end of July, the Independent was reporting that "Facebook abandoned an experiment after two artificially intelligent programs appeared to be chatting to each other in a strange language only they understood." Snopes gave the story its own page. Project lead Mike Lewis posted corrections on Facebook.

The actual bot output looked like "balls have zero to me to me to me to me to me to me to me to me to." Neither bot managed the compound task of learning a limited negotiation language and negotiating well, and the author of this column could not find the widely reported claim that the bots had developed their own grammar anywhere in the Arxiv paper. The experiment was not a success with a spooky side effect. It was a work in progress that produced a headline cycle.