New technology from Facebook could let bots handle more dreary tasks.
Anyone with an Amazon Alexa or Google Home device (or the right smartphone app) can order takeout food or hail an Uber directly via these technologies. But what if the software bots, or chatbots, running in those devices could not only follow our commands, but also negotiate pricing or perks for us?
The researchers developed technology to enable a bot to build a model of a dialogue between it and another party, with whom it is engaging in a transaction. The goal was to enable the bot to anticipate the other party’s possible reactions to a given offer, and plan out its responses in each case.
The difference between today’s bots and future generations is that the latter should be able to “think ahead” or anticipate directions a conversation could take in order to optimize the possible outcome.
Bots are fine for simple processes: you ask a bot to perform a task that other party (Uber, Dominos) wants to perform for a fee and it orders your ride or your pizza. All parties are aligned.
But in other cases there will be unknown factors. Carnegie Mellon University’s contest pitted its AI against human poker champions is one example. In poker or other games, players purposely withhold information or seek to mislead the other contestants, including the bot, by bluffing. That is a longer-term transaction that requires multiple steps, and longer term thinking on the part of the bot.
In this case, Facebook thinks future bots can do better at negotiating tasks that people do all the time and use natural language processing to do that.
“Negotiations are super common in everyday life,” Lewis tells Fortune. “Two people have different goals and need to come to agreement on things like which TV channel to watch, meeting times, what restaurant to go to. People tend to do this in natural language—so far there is very little AI involvement.”
Facebook hopes to change that not only with the new research paper, but by open-sourcing software it created to facilitate bot-driven negotiation to anyone who wants to work with it.
Batra said there is a spectrum of human-bot interactions: Today you tell a bot what to do and if possible, it does it. Orders a pizza, a Lyft, books a plane ticket. It’s very binary.
But when a bot has to facilitate a transaction in a semi-adversarial environment, conditions get much more complex. It knows what you want, but not necessarily what the other party is willing to provide or at what price. Say you want to book a plane ticket, but you’re only willing to pay X dollars. What if a bot could bargain with the airline or travel service, shielding what it knows from the other side, to get the best deal?
Get Data Sheet, Fortune’s technology newsletter.
While the Facebook researchers focus on the underlying technology, they do see practical applications down the road. Lewis, like many others, would love a bot to schedule meetings.
“Booking meeting times is super annoying, there is so much back-and-forth” he said. “A bot could do this for me. I would give my preference that I can definitely do these times, I’d prefer not to do it at 7 a.m., but will if I have to, and the bot would respect my preferences and get me an outcome.” In that case, the bot might propose optimal meetings, and when they are shot time, come back with other options—but hopefully not the 7 a.m. time Lewis balks at.
Nearly every tech giant—including Microsoft msft , Google goog , and IBM ibm —is investing heavily in AI technologies, such as natural language processing and image recognition. They all see huge opportunities in making their software more capable of human-like thought and able to take on more types of work.
At the same time, there are growing worries that as software gets smarter, it will kill jobs for more classes of workers.