“What a wonderful family you have! Just initial here and sign at the bottom. Great tie, by the way. All we'll need today is your $69.99 one-time activation fee that will recur monthly, and you can walk away a happy man, knowing you made the right decision. Will that be cash or card?”
We've all seen a cheesy, fake smile during a similarly insincere sales pitch–usually alongside a bombardment of faux compliments and half-hearted interest in our needs.
There are common tactics used by skillful negotiators in order to reach desired outcomes–ones that are often mimicked clumsily by the inexperienced salesperson. But these tactics can be used to great effect and those that have mastered them have often grown rich, either through their own success utilizing the methods or by teaching others with books and seminars.
One might say that sales and negotiations have their very own language.
Despite this, it was still a surprise to see how far AI would take the concept.
Facebook's negotiating chatbots did just that–created their own non-human language while learning to make deals.
Curious features of negotiating AI
Facebook’s artificially intelligent dialog agents began to converse with one another using a previously unseen language that prompted researchers to switch to a fixed supervised model in order to keep the bots negotiating in human words.
In their own post on the topic, Facebook explains that not only did the AI chatbots learn to create new sentences (even in English), but they also, of their own accord, learned human negotiating tactics and employed them to their benefit.
For example, AIs would feign interest in an item up for negotiation only to concede it later to get the deal they wanted, a manipulative tactic used by many human negotiators.
The constant goal of these conversational AIs was to complete the deal, unlike many human-human interactions that end with someone choosing the option to walk away.
This resulted in the AIs learning to negotiate longer and express their wants in their own (English) words when negotiating with humans.
This conversational feat is impressive in terms of artificial intelligence because negotiation requires the AI to understand and respond accordingly to complex conversational information and even read between the lines to understand the value that items have to the people and machines it speaks with.
What can we do with a negotiating AI? Or AIs that communicate in their own invented languages?
As the negotiation skills of AIs grow more advanced–in English and machine language alike–what are the possibilities for use?
Perhaps we will see AIs that can carry out their own social engineering attacks both by nefarious groups as well as in aid to professional security groups.
Another gripping thought is an AI that could make suggestions to a criminal interrogator or intelligence agent.
With their own languages, might they also be able to work through issues and convey valuable ideas that require high security and confidentiality only to put into human words the final solution? Only the future will tell.