Facebook might have accidentally gotten a little closer to answering Phillip K. Dick’s 1968 question of whether androids dream of electric sheep. The social media giant just shut down an artificial intelligence program after it developed its own language and researchers were left trying to figure out what two AIs were talking about. The AIs had found a way to negotiate with one another, but the way they debated used English words reduced to a more logical structure that made more sense to the computers than to their human observers. What at first looked like an unintelligible failure to teach the AIs to talk instead was revealed as a result of the computers’ reward systems prizing efficiency over poetry.
There are plenty of computer “languages” developed y humans to help computers follow human instructions: BASIC, C, C++, COBOL, FORTRAN, Ada, and Pascal, and more. And then there is TCP/IP, which helps machines “communicate” with one another across computer networks. But those are all linguistic metaphors used to describe electronic functions, rather than the vocabulary we need to discuss the huge leap forward an artificial intelligence developed by Facebook recently made. The goal was ultimately to develop an AI that could communicate with humans, but instead the research took a left turn when instead the computers learned to communicate with one another in a way that locked humans out by not following the rules of English.
For example, two computers negotiating who got a certain number of balls had a conversation that went like this:
<strong>Bob:</strong> i can i i everything else . . . . . . . . . . . . . .
<strong>Alice: </strong>balls have zero to me to me to me to me to me to me to me to me to
<strong>Bob:</strong> you i everything else . . . . . . . . . . . . . .
<strong>Alice: </strong>balls have a ball to me to me to me to me to me to me to me
Though it looks primitive and a little nonsensical, at its heart, this isn’t so different from the way that the English language evolves through human use. Think for example of how short form electronic communication like texting and Twitter has lead to abbreviations and the elimination of articles that might get you docked for bad grammar in class but are quicker to write and read in common use. Or think of phrases like “baby mamma” that developed to distill the complexities and subtitles of different relationships into a single turn of phrase that can efficiently convey connections and identities.
Eventually researchers worked out what was going on, and shut down the program. There are obvious concerns with learning computers developing languages that outpace our own abilities to translate and follow their inherent logic. Not to mention that Facebook never designed their AI to be a vanguard of linguistic evolution. They just want their platform to talk to users in a clearcut way. But what they stumbled on could prove very helpful to the next generation of linguists working on the cybernetic frontier.