Skip to main content

Chatbots are dumb, but wait until they learn how to negotiate for you

Chatbots are dumb, but wait until they learn how to negotiate for you


New research from Facebook’s AI lab gives a promising peek at the future of bots

Share this story

Facebook has been experimenting with chatbots for a couple of years now — including its assistant M (above).
Facebook has been experimenting with chatbots for a couple of years now — including its assistant M (above).
Photo by Vjeran Pavic / The Verge

Chatbots were supposed to be a big part of our AI-powered future, but they’ve mostly fallen flat. Scratch the surface of any online bot selling you takeout or flights abroad, and you’ll usually find drop-down menus repackaged as questions. To get chatbots to the next level (and make them genuinely useful), they’ll need to be given new skills — like memory, and the ability to reason. Adding these new cognitive abilities is closer than you think.

Facebook’s chatbot was often mistaken for human

Facebook is one of the biggest players in this domain. They launched their own chatbot assistant M (though humans do all the tough tasks); built an open-source framework for teaching bots; and today, researchers from the company’s FAIR lab have demonstrated a new way of training chatbots to negotiate on behalf of users.

The work is limited in scope, focusing only on a single negotiation scenario, but is a good proof of concept that points the way to building more powerful bots in the future. It also produced two particularly interesting outcomes. First, during a testing phase, the bots were often mistaken for humans; and second, that, without any human direction, the bots developed nuanced negotiation strategies resembling elements of game theory, particularly the prisoner’s dilemma.

The technology at the heart of this research here is a neural network trained on a dataset of recorded negotiations. In this case, the negotiating scenario was “multi-issue bargaining” in which humans participants recruited via Mechanical Turk were asked to divvy up between them a number of items. Each participant valued the items differently and was asked to maximize the number of points they got.

This was the interface in Mechanical Turk used to collect negotiating data and train bots.
This was the interface in Mechanical Turk used to collect negotiating data and train bots.
Image: FAIR

This provided the backbone of the neural net, but the researchers’ key innovation was adding what they called “dialogue rollouts.” Essentially, this means asking the bots to think ahead; to simulate how future negotiations might go and pick the best course of action to get what they want. Speaking to The Verge, researcher Michael Lewis said: “These kinds of techniques are used a lot for playing games like chess and Go, but for the first time we’re trying to bring these into the world of dialogue.”

Incorporating this dialogue rollout module netted “big improvements” compared to similar research in this area, say the researchers. After training their bot, they let it loose on negotiations in Mechanical Turk. “For the most part, people didn’t notice they were talking to a bot,” says co-author Dhruv Batra. “And our best models were getting pretty comparable scores to what the people were getting. Certainly not better, but certainly not much worse.”

The bots even learned how to deceive one another as a method of negotiation. For example, they found that a good strategy was to aggressively pursue an item they didn’t care about, only to concede at the last minute and appear to compromise. “No human [programmed] that strategy,” says Batra, comparing these methods to the basic tenets of game theory. “This is just something they discovered by themselves that leads to high rewards.”

These bots can talk, but they’re hardly talkative

However, it’s important not to get carried away by this research. As noted above, the bots weren’t consistently better at negotiating than humans. And the researchers were only able to train them in this one specific scenario. They don’t know whether they’ll be able to transfer these skills to other sorts of negotiation.

“not a breakthrough”

Kaheer Suleman, a researcher at the Microsoft-owned AI company Maluuba (which also works on chatbots), said the paper was a “good step forward” but “not a breakthrough.” He points out that the use of Mechanical Turk to get training data is a limiting factor, as it means the sentences the bots are trained on will be pretty basic. “People on Turk are going to want to get these tasks done as fast as possible, so they’re not going to be so artistic in their use of language,” says Suleman.

This is certainly borne out by the paper, which shows the bots using very simple speech. Sentences like “I want the book and the hats, you get the ball” are used as examples. There’s no nuance or subtlety here, only a simple encoding of values (i.e. how much you want the ball) in simple language. This is a common criticism of AI-generated language, with researchers noting that just because a robot can produce readable sentences, it doesn’t mean the sentences themselves aren’t robotic.

Negotiation by AI is already happening

Even with these caveats, though, it’s worth thinking about the future of negotiating chatbots. The field is a fast-moving one, and big tech companies have access to vast amounts of user data that could help deliver improvements. There’s no timeline for this recent FAIR research ever becoming a product, but the team behind it has already dreamt up various applications.

Chatbots like this could be used for tasks like negotiating prices and arranging meeting times, they say, with users programming in their preferences for what they want out of a negotiation. “It’ll be like, ‘don’t talk to me, talk to my bot,’” says Batra. “It’ll go to bat for you.” It’s a compelling vision, but one that tech companies have been pushing for a while without much success. (Consider Microsoft’s 2016 Build conference, which promised that the next form of computing would be “conversation as a platform.”)

A screenshot of DoNotPay — a chatbot that helps users avoid paying parking tickets.
A screenshot of DoNotPay — a chatbot that helps users avoid paying parking tickets.
Image: DoNotPay

Others in the industry are worried that this coming dynamic won’t work in consumers’ favor. Thomas Smyth, CEO of Trim, a company that makes a chatbot designed to help people manage their money, thinks that this sort of technology will overwhelmingly benefit corporations. After all, he tells The Verge, if the key to training smart AI is data, who can more easily collect relevant data — a big company or consumers?

“The benefit goes to the party who negotiates with much greater frequency,” says Smyth. “Clearly this technology is going to be used by those corporations to maximize their advantage in any and all negotiations.” FAIR has actually open sourced their neural network, allowing anyone to use it (a move which Smyth praises enthusiastically), but data collection is still a bottleneck when it comes to developing a more powerful bot.

“this technology is going to be used by corporations.”

There is some hope, though. Trim itself offers a chatbot that will negotiate with Comcast for you over your cable bill, and Smyth notes that the success of this bot is partly due to the fact that Comcast reps follow a very strict script. If you know the right things to say, they’ll give you what you want. And other chatbots have been successful at dealing with other rigid forms of billing bureaucracy. One named DoNotPay has overturned more than 160,000 parking tickets for users in New York and London.

So even if negotiating chatbots continue to be a little basic and inflexible, at least they’ll be a perfect match for robotic customer service.