Facebook’s AI chatbots have managed to come up with a new ability, the one to negotiate and also create their own sentences during these negotiations, rather than relying just on the ones taught in training. Facebook sees chatbots powered by artificial intelligence as the future for its platform, and really crucial to Facebook Messenger. So the ability of chatbots to negotiate marks a significant breakthrough for the company.
Researchers at Facebook Artificial Intelligence Research (FAIR) have now published their work around these chatbots, which have the ability to negotiate. Facebook’s researchers have also made the code for this open-source, so that others can also use this to build their own bots with similar capabilities.
Facebook’s blogpost says just like humans can negotiate and come to agreements and a compromise, their research confirms it is possible to teach these chatbots, which might have different goals, the ability to start and finish negotiations with other bots or people. The blogpost explains in details how Facebook taught its bots to negotiate.
Facebook researchers showed the bots the same collection of an assortment of items (could include a number of objects). Then the task is for the bots to divide them among themselves by “negotiating a split of the items.” The bots also attach their own value to each item, and if a bot walked away from the negotiation, both got zero points.
Facebook’s chatbots didn’t just stop at negotiation. The blogpost also notes the idea is to have chatbots that can “‘think ahead’ or anticipate directions a conversation is going to take in the future, they can choose to steer away from uninformative, confusing, or frustrating exchanges.” Facebook researchers also developed a new technique called “dialog rollouts”, where the chatbot can simulate future conversation, ensuring maximum returns and rewards.
The chatbots were not only able to negotiate harder with humans, they were also able to bluff during negotiations, aka pretending not to be interested in items, only to get them later. As Facebook’s post points out, this kind of game is a trait unique to humans. Additionally, the chatbots also came up with their own sentences in the negotiations, rather than just relying on sentences given to them in the training data.
Facebook’s bots were tested out against humans as well, and people didn’t realise this was a bot and not another person, say the researchers. The bots got better deals as much as worse deals, and could speak English as well as humans, according to Facebook’s study.
So what does all of this chatbot training meaning for users? Well the future of services will likely be driven by chatbots, and not actual customer service representatives. As Facebook’s research shows these bots will soon figure out negotiation on their own, and humans won’t be any the wiser.