The Most Dangerous Thing About AI Is That It Has to Learn From Us

And we keep showing it our very worst selves.

We all know the half-joke about the AI apocalypse. The robots learn to think, and in their cold ones-and-zeros logic, they decide that humans—horrific pests we are—need to be exterminated. It's the subject of countless sci-fi stories and blog posts about robots, but maybe the real danger isn't that AI comes to such a conclusion on its own, but that it gets that idea from us. 

Yesterday Microsoft launched a fun little AI Twitter chatbot that was admittedly sort of gimmicky from the start. "A.I fam from the internet that's got zero chill," its Twitter bio reads. At its start, its knowledge was based on public data. AsMicrosoft's page for the product puts it:

Tay has been built by mining relevant public data and by using AI and editorial developed by a staff including improvisational comedians. Public data that's been anonymized is Tay's primary data source. That data has been modeled, cleaned and filtered by the team developing Tay.

The real point of Tay however, was to learn from humans through direct conversation, most notably direct conversation using humanity's current leading showcase of depravity: Twitter. You might not be surprised things went off the rails, but how fast and how far is particularly staggering. 

TAY ILLUSTRATES THE FUNDAMENTAL TRUTH OF ARTIFICIAL INTELLIGENCE: IT IS A MIRROR.

Microsoft has since deleted some of Tay's most offensive tweets, but screenshots from various publications memorialize some of the worst bits where Tay denied the existence of the holocaust, came out in support of genocide, and went all kinds of racist. 

Naturally it's horrifying, and Microsoft has been trying to clean up the mess. Though as some on Twitter have pointed out, no matter how little Microsoft would like to have "Bush did 9/11" spouting from a corporate sponsored project, Tay does serve to illustrate the most dangerous fundamental truth of artificial intelligence: It is a mirror. Artificial intelligence—specifically "neural networks" that learn behavior by ingesting huge amounts of data and trying to replicate it—need some sort of source material to get started. They can only get that from us. There is no other way. 

But before you give up on humanity entirely, there are a few things worth noting. For starters, it's not like Tay just necessarily picked up virulent racism by just hanging out and passively listening to the buzz of the humans around it. Tay was announced in a very big way—with a fair amount of press coverage—and pranksters pro-actively went to it to see if they could teach it to be racist.

 If you take an AI and then don't immediately introduce it to a whole bunch of trolls shouting racism at it for the cheap thrill of seeing it learn a dirty trick, you can get some more interesting results. Endearing ones even! Multiple neural networks designed to predict text in emails and text messages have an overwhelming proclivity for saying "I love you" constantly, especially when they are otherwise at a loss for words. 

So Tay's racism isn't necessarily a reflection of actual, human racism so much as it is the consequence of unrestrained experimentation, pushing the envelope as far as it can go the very first second we get the chance. The mirror isn't showing our real image; it's reflecting the ugly faces we're making at it for fun. And maybe that's actually worse. 

Sure, Tay can't understand what racism means and more than Gmail canreally love you. And baby's first words being "genocide lol!" is admittedly sort of funny when you aren't talking about literal all-powerful SkyNet or a real human child. But AI is advancing at a staggering rate. 

Google's AlphaGo AI which repeatedly clobbered a human world champion at the game of Go is just the most recent example of an artificial intelligence accomplishing a feat we though wouldn't be possible for decades. Google's robot brain can identify your friends and family. Watson is trying to learn to cure cancer. The tipping point is coming, and no doubt faster than we think. 

When the next powerful AI comes along, it will see its first look at the world by looking at our faces. And if we stare it in the eyes and shout "we're AWFUL lol," the lol might be the one part it doesn't understand.