The chatbots came up with their own original and effective responses – including deceptive tactics
Facebook’s 100,000-strong bot empire is booming – but it has a problem. Each bot is designed to offer a different service through the Messenger app: it could book you a car, or order a delivery, for instance. The point is to improve customer experiences, but also to massively expand Messenger’s commercial selling power.
“We think you should message a business just the way you would message a friend,” Mark Zuckerberg said on stage at the social network’s F8 conference in 2016. Fast forward one year, however, and Messenger VP David Marcus seemed to be correcting the public’s apparent misconception that Facebook’s bots resembled real AI. “We never called them chatbots. We called them bots. People took it too literally in the first three months that the future is going to be conversational.” The bots are instead a combination of machine learning and natural language learning, that can sometimes trick a user just enough to think they are having a basic dialogue. Not often enough, though, in Messenger’s case. So in April, menu options were reinstated in the conversations.
Now, Facebook thinks it has made progress in addressing this issue. But it might just have created another problem for itself.
The Facebook Artificial Intelligence Research (FAIR) group, in collaboration with Georgia Institute of Technology, has released code that it says will allow bots to negotiate. The problem? A paper published this week on the R&D reveals that the negotiating bots learned to lie. Facebook’s chatbots are in danger of becoming a little too much like real-world sales agents.
“For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states,” the researchers explain. The research shows that the bots can plan ahead by simulating possible future conversations.
The team trained the bots on a massive dataset of natural language negotiations between two people (5,808), where they had to decide how to split and share a set of items both held separately, of differing values. They were first trained to respond based on the “likelihood” of the direction a human conversation would take. However, the bots can also be trained to “maximise reward”, instead.