If you use social media at all, the complaints may sound familiar: Someone buys a book, for example, and leaves a negative review online, only to have ads for that same book begin dotting their feed. It can sometimes feel like the internet is psychic, but when gaffes like this occur, that psychic might as well be wearing a bucket on their head. Welcome to the limits of artificial intelligence.
“You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place” (Voracious) gets into the minutiae of how AI works, and why it so often doesn’t, in a way that lay readers can easily follow. Author Janelle Shane uses line drawings both for comic effect and to illustrate the concepts she’s describing. (She has a TED talk online that serves as a good preview for the book if you’re curious.) If you’re concerned about artificial intelligence gaining sentience and enslaving humanity, the news is generally good; AI can only do what we ask it, and most of the time if problems arise, they come from its taking our instructions too literally, or embedding our biases into its coding.
Shane opens with a step by step description of her attempt to teach a neural network how to write a knock- knock joke. Unlike computer programming, where someone can create a program that follows the structure and then load it with potential answers that fit, neural networks are just given examples and instructions to replicate them. Beginning with garbled letters, it slowly evolves and figures out the joke structure, though very few of its jokes make sense (the closing couplet “Ireland who? Ireland you money, butt,” gets credit for a good try.) At one point in the process it wrote a joke that was just the word “Whock” multiple times; at another it was focused on a punchline that relied on “ooo” sounds and overused them to create a joke that was really just an extended howl. When it does create a joke with an actual funny punchline, Shane reminds us that the network doesn’t know it. It doesn’t know what words or letters are, much less knock-knock jokes, and is just trying to follow instructions to the letter.
That literal-mindedness sometimes leads an AI to perform in ways we think of as cheating, though again, it is just scrupulously following instructions. In one instance, a network was designed with a robot “body” (arms, legs, etc.) and directed to travel from point A to B, with the hope that it would walk. Instead, it stacked all its parts into a tower, stood at point A and fell like a tree, landing at point B. It’s instructive and also funny, but sometimes the results signal something more serious is at play. Many current applications of AI have been caught reinforcing systemic bias that was unintentionally fed to it as data. If the accepted resumes in a pool largely feature male names, guess who’s going to be rejected? When automatic paper towel dispensers are trained on white people’s hands, they may ignore anyone else’s as irrelevant. A diverse workforce can’t solve this problem if the data sets themselves aren’t reflecting that diversity. This can be corrected, of course, but engineers are still figuring out how to instruct neural networks without importing our problems to them.
Shane is a fine tour guide to this material, and as often as not a straight man to frequently hilarious lists of AI-generated material (the title is from a list of pick-up lines that also included, “You must be a tringle? Cause you’re the only thing here.”). She describes how it’s possible to piggyback on prior work with new data to save time, using the example of an AI trained to configure metal band names (Deathcrack, Inhuman Sand) that is then taught to create ice cream flavors (Cherry Chai, Lemon-Oreo). In between, though, you get Dirge of Fudge, Blood Pecan, and Spider and Sorbeast, among other very metal ice creams.
I picked up “You Look Like A Thing and I Love You” specifically for those AI-generated laughs, but came away with an appreciation for all the useful things it does, as well as where it needs to be made better than the people feeding it data. It’s highly entertaining, but also a generous introduction into something we’re often engaged with without realizing it. It will make you smarter and more aware, but you’ll also have a good time along the way.
Heather Seggel is a writer living in Northern California. Email heatherlseggel@gmail.com.
From The Progressive Populist, May 15, 2020
Blog | Current Issue | Back Issues | Essays | Links
About the Progressive Populist | How to Subscribe | How to Contact Us