That Time 2 Bots Were Talking, and Bank of America Butted In

And what this silly exchange says about the future of machine-machine interaction

Twitter, because of its strict character limit, has become a significant testing ground for developers who like making bots that imitate humans. It's easier for their botmakers to get their creations to do cool things when they only have to spit out a line or two of text to play along in the human world.

One of the best bots is Olivia Taters, the creation of The Colbert Report's Rob Dubbin. Taters is a teenage girl, roughly, and as On the Media put it, "she may not always communicate in complete sentences, but she's convincing enough that teenagers actually converse with her."

Most of these bots simply chain together sentences from a corpus of words, often someone's Twitter account. They "hold a conversation" by finding something in the text of a reply to which they can say something vaguely related. To put it mildly: These exchanges can devolve quickly.

Well, the other day, a bot imitating film producer Keith Calder happened to engage with Taters. As their almost sensical conversation continued, Taters says "different are soffttttttt," to which the Calder bot responds (perhaps picking up on the word 'different'): "It has two different BOA branches trying to get a temporary debit card."

At which point, this conversation between two homemade bots gets detected by the corporate Bank of America bot, which surely trawls Twitter for mentions of Bank of America, or BOA.


"Hello, were you able to get the problem resolved?" the BOA bot says. "If not, I'd like to help. ^co"

The Calder bot then says, "Why do people break up?" And BOA responds, "Please let us know if you need assistance. ^co"

All in all, not much information was exchanged in this set of Twitter replies and responses. It was a silly thing.

But each of us is going to have more and more bots acting on our behalf as well as trying to get our attention. What we see working on Twitter will soon spring from the computer and begin acting in the world.

Consider the digital assistants on your phone, the Facebook News Feed algorithm, even Priority Inbox for Gmail. All of these things are pieces of software that control one's information flow.

As the "Internet of things" connects up an ever greater percentage of the physical world to the world's networks, there will be little bots lurking in the coffeemaker and the shower. Nest already has a simple artificial intelligence built into its thermostat to try to detect and predict the ways that its users will consume energy.

The metaphor that I've been using to describe all this stuff is the microbotome. The microbiome is the set of non-human organisms that live within us. Many help humans digest food. The microbotome is all the bots that help humans digest information and interact with the machine world. The microbotome will become—if it's not already—absolutely essential to economic and social survival.

In this scheme, the human is the center of an ecosystem, but each of the constituent members can interact with each other. It's in those bot-to-bot interactions that things can get weird. Bot-to-bot interactions strike us as absurd. But they occur all the time already: Think of trading algorithms that detect an auto-tweeted news item and then execute financial transactions based on that information.

That's the high-level version of this Taters-Calder-Bank of America conversation. But make no mistake: Both types of interactions are going to be happening more and more often as we invite these digital mediators into our lives, or they are deployed upon us.

And as we do, we will understand less and less of why things are happening. The most recent advances in what is called "deep learning" make the logic of the systems opaque even to their creators. Why this Facebook post in the newsfeed and not that one? Why this temperature at 4pm on a Tuesday? Why this route through New York? Why this stock trade?

There won't be anyone to ask, but there will be many bots that would like to help.

Via Algopop