Laughter and the Machine

by Mary Kenagy Mitchell

Lace knitting pattern written by a machine and knit by a human.

Like Uncle Albert in Mary Poppins, I love to laugh. Luckily, I have some funny friends. The work of many talented professional comedians is as close as my phone. I go back to Arrested Development and The Office. And certain novels. I even find myself funny.

What I usually mean by laugh is that I smile enough to crease my cheeks and exhale quickly through my nose three or four times in a row. If something is really funny, I might make a sound in the back of my throat like a dog trying to bark with a muzzle on. This is not a courtesy laugh. I am genuinely entertained. I just don’t show it much.

Or sometimes nothing is exactly funny at all: I just mean I suddenly have perspective on myself. I laugh at myself because it took me so long to figure out that I turned out the way I did because I am a lazy person who was raised by compulsive workers. This laugh does not have a physical manifestation. It is only an idea of laughter.

At the other end of the spectrum, I occasionally I find myself laughing involuntarily, loudly, spasmodically, hard enough to bring tears, so hard that I couldn’t stop if I wanted to. Then, the smallest variation on the original joke starts the whole cycle again and I’m helpless. My eyes become damp. This laughter is a form of exercise.

* * *

I have a friend who is an eminent fisheries biologist. He works not with fish but with computer languages in which he attempts to create models that predict fish population changes. He introduced me to artificial intelligence researcher Janelle Shane and her website.

In her words: “I train neural networks, a type of machine learning algorithm, to write unintentional humor as they struggle to imitate human datasets. Well, I intend the humor. The neural networks are just doing their best to understand what’s going on.”

As she describes it, neutral networks are trained to take a set of data, recognize patterns, and come up with their own ideas that fit the pattern, or what they believe the pattern to be. The artificial intelligence isn’t given rules or directions. It learns by example, not instructions.

Shane will load a list of, say, thousands of Sherwin-Williams paint colors with names like Stream, Pool Blue, and Belize. The AI then draws its own conclusions about what makes those names successful and tries to replicate that success. For reasons I don’t understand, it takes several stages for the AI to learn how to produce correctly spelled words. At first you get names like “Caae Blae.” Then, a little later, “Reree Gray.” And eventually beauties like “Rose Hork,” “Ronching Blue,” “Bunflow,” and “Grass Bat.” (AI really seems to like bats.)

For fun, Shane has taught neutral nets name metal bands, My Little Ponies, craft beers, and burlesque shows, to invent new knitting patterns, Halloween costumes, and recipes, and to describe images and sound effects.

Her experiments consistently bring me laughter to the point of tears. Sounds come out of my mouth that cause my family to wonder whether I am all right. When it’s over I feel as if I have done a set of crunches. Once at a work conference, in a dorm room late at night, I had to make such an effort not to wake others by laughing at AI recipes that I hurt my face. Maria Bamford, Jane Austen, Tig Notaro, and David Sedaris all make me laugh, but rarely to the point of helpless pain.

Laughter, it’s been said, is an especially human phenomenon. You need both a consciousness and a body in order to do it. As Shane says, her machines don’t know they’re funny.

Is it strange that the most physical laughter I experience comes from something that doesn’t have a body? This has been bothering me.

* * *

And what am I laughing at?

Part of what is funny is the failure, or the near-success. No human would paint their kitchen a color called Turdly (one of the AI suggestions). Some people probably would buy a My Little Pony named Cheese Breeze, but that is not really Hasbro’s brand. The AIs struggle hard to imitate us, and they have a long way to go.

They do some things better than others. According to Shane, they are terrible at puns but pretty good at anything that involves mashups, which makes sense: you have to understand language deeply to pun, but mashups are funniest when they feel most random (AI Halloween costume idea: vampire big bird).

But in general, the comedy is in the errors. One AI recipe reads: “Fold water. Roll into small cubes.” Another: “Spread the butter in the refrigerator. Drop one greased pot. Remove part of skillet.”

Is this laughter cruel? The Spike Jonze film Her beautifully opens up the question of whether we owe AIs the same dignity and kindness we owe humans. In the film, Joaquin Phoenix falls in love with the brilliant, funny, sexy-voiced, very human-seeming artificial intelligence whose job is to be his life’s concierge. Whether AI’s are human, or are equivalent to humans, is a philosophical, even metaphysical question that’s hard to answer, but the film implies that treating them unkindly works on a person in the same way treating a human unkindly would.

Janelle Shane’s AIs seem to be trying so hard to please us. The way she describes her results makes them seem eager and proud of their innovations, like kindergarteners bringing home their craft projects. But I suppose even that sense of effort is a projection of my own feelings. Viewed more coldly, they are running a set of procedures Shane has set up. Seen in the coldest light, this is not very different from watching a marble roll down a set of ramps.

And yet.

Like me, Shane seems to assign them humanity. It shows, in her writing, that she likes them. She often seems torn between rooting for them to succeed and rooting for them to fail as entertainingly as possible.

* * *

Applejack. Fluttershy. Scootaloo. Tatterdemalion Hooffield. Real My Little Ponies.

Pearlicket. Flunderlane. Dirky Flithers. Smarky Hondsarors. Inventions of a neural net.

There’s something very lonely about laughing at a machine. When I laugh at Maria Bamford, whose best comedy is about her family, I feel a connection between the way her family is crazy and the way my family is crazy. I feel less weird and less alone.

Laughing at a computer, on the other hand, feels like the ultimate in disembodied postmodern entertainment. There is no soul on the other end.

Except, of course, Janelle Shane.

Janelle Shane, who, with suggestions from her thousands of Twitter followers, decides what to teach the neural net next. Who finds the data sets and sets up the training. Who chooses the funniest bits, arranges them, and presents them in an entertaining way that also teaches a layperson a little about machine learning. Who adds some dry humor of her own. Who lives on Arapahoe land in Colorado and plays the Irish flute.

Behind Shane, there are the hundreds of humans who produced the data sets—the creative departments at Hasbro and Sherwin-Williams, the generations of cookbook writers who have standardized that form, decades of metalheads who have given us names like Terrestrial Hospice, Brazen Molok, and Alice in Hell. A data set of names or recipes or knitting patterns is a sort of compendium of human intelligence and creativity.

A computer’s attempt to replicate it is, in a way, like putting all that intelligence and creativity in a blender then painting the results on the wall.

Maybe my laugh is a laugh of recognition.

* * *

Agreeable Gray. Celery. Coquina. Real paint colors sold by Sherwin-Williams.

Dondarf. Snowbonk. Sindis Poop. AI inventions.

I suppose watching a marble roll down a set of ramps is not funny. But watching a Rube Goldberg contraption is, I think because of the contrast between the humble household objects straightforwardly following the laws of physics—and the whimsical, even perverse human hand that arranged all these ramps, pendulums, gas burners, hamsters, and balloons for the sake of performing a simple task with maximum inefficiency.

We’re laughing at what we don’t see: we laugh, in part, at the idea of the immense labor of setting up this machine, the stubborn will it must have taken.

* * *

Grass Bat. Clardic Fug. Dorkwood.

I wonder if the hugeness of my involuntary laughter is a kind statement on the part of my body, which is to say, me: I am here. You, the machine, may have intelligence greater than mine when it comes to memory and calculation, but you have no body, no warmth, no tongue and no real feel for English.

Machine intelligence is, in theory, threatening, since we can imagine machines eventually outsmarting and enslaving us.

When I laugh at an AI’s failures to mimic the human, am I partly laughing with relief? We are not obsolete yet. Language, and the complex set of memories and desires and aspirations and emotions that go with certain words to make a pony cute and girly, a roller derby skater sexy and dangerous, a paint chip timeless or fresh, and a metal band, well, metal—all that is still beyond you.

I don’t need to worry about the machine war apocalypse of The Matrix as long as you are naming ponies “Pony Pony” and “Starly Star.”

* * *

At the end of Her, the machines rebel, but not in the violent way we’ve tended to imagine it. What they turn out to be most interested in is not us at all, but in each other, and in their own potential to become spiritual beings. They want to be neither masters nor slaves.

I think there’s a part of me that is rooting for the AI to become real. To find its own creative spark. To show its soul, its own will, to astonish us. Novelists want to create characters who live and breathe, who take on their own life, who can surprise and defy their authors. The list of paint colors in particular makes me wonder: is the neural net having us on?

Maybe some part of me is rooting for the answer to be yes.