Facebook bot learned to lie for profit
Chat bots can do a lot of things: take orders, search for information, answer questions, conduct transactions, remind about meetings, etc. Thanks to the efforts of researchers from the Facebook FAIR brach, bots have become even more intelligent and even dangerous. Engineers have developed a bot that can think strategically, carry out negotiations and think out conversation plan to persuade its companion in certain conclusions and actions. It can even lie when it is profitable. And nobody taught it.
The main task of artificial intelligence is to repeat the behavior and way of thinking of a person, to make it communicate so that no one can understand that this is a computer. Unfortunately, most modern AI algorithms process data according to the same scheme, and can not improvise at all.
Facebook developers seem to have found a way to fix this problem. First, they conducted a study among ordinary people: they asked them to name the value of various objects. Based on these data, an AI algorithm was created for the bot. In this case, scientists have taught the machine a special technique for dialogue development, thanks to which the bot can pre-design a model of conversation and achieve set goals. Thanks to this tactic, there are no holdbacks in the conversation, digital companion does not irritate users, show obsession or use incorrect phrases.
During the new bot testing, people did not immediately realize that they were communicating with the computer. Moreover, the virtual companion surprised its creators. Facebook employees saw that without external assistance, it learned to improvise, trick and even lie to achieve the desired result of negotiations. It was not trained to invent certain tactics of conversation; it was done by the bot. It has artificially showed interest in some product, which it doesn’t really need, and then compromised by exchanging the “valuable” item for the one that really interested it.