I Love Twitter Bots and I Cannot Lie

“I love bots” feels like a controversial statement now, given the rise of troll farms deploying them and the proliferation of egg accounts spreading bad information or engaging in harassment. Even if you set aside, say, their use by state actors to engage in digital political warfare, some bots are just annoying, spamming links to shady ecommerce web shops underneath viral posts. Who could love bots?

Me, I do. I love bots.

Consider this my hill to die on: Bots are not just agents of harm, and in our fight against their more malicious uses, we should avoid throwing out the metaphorical baby with the bathwater.

Think about a bot—what immediately comes to mind? Bots can be so many things beyond just the regular misinfo social media bots that are covered again and again in the news. The Barbican Centre’s 2019 AI: More Than Human show referenced “golems” as some of the earliest forms of AI. (Golems, according to Jewish folklore, are artificially created humans.) Following this through line, we can consider the Mechanical Turk—a chess automaton performed by a hidden human, designed to trick other humans into believing it to be real, moving, and sentient—as a bot also.

So what is a bot? It’s a bit of a philosophical question, because finding the answer entails figuring out the context in which you are inquiring about the bot. Are you talking about an entity, or a kind of behavior? A “bot” or something “bot-like” could describe both. A gaming bot can be a preprogrammed AI or an NPC (non-playable character) that populates the backgrounds of video games, engaging in random conversations and simulated tasks. A social media bot can be a Twitter account tweeting out nonsensical messages from a dataset, or planned programmatic responses in response to a word, theme, or even specific date.

So what do these examples have in common? Most simply, a bot is “an automated piece of software that performs predefined assignments, usually over a network.”

And yet bots aren’t one thing but many things. Bots are pluralistic, strange, and pure computation often mixed with simple “stupidity.” Bots can do only the things they are programmed to do, so in a way, they are extremely literal, and thus simple and “dumb.” Of the narrow “things” bots can do, they can still do a lot—and some of the actions are incredibly creative, even if other bots still tweet out spam links.

There are  joke bots and art bots and Slack bots that send emojis and dancing parrots and can help you order lunch. My favorite bots play precisely into this simplicity with beautiful and hilarious results.

Some of my favorite bots are programmed by Darius Kazemi, an engineer and artist based out of the Pacific Northwest in the United States. Kazemi’s bots are silly, often referencing memes or internet jokes. There’s a structure to some standard jokes, like knock-knocks, for example, or “How many X does it take to screw in a lightbulb?”, and some memes have that structure too, with images and text combined to produce something humorous, snarky, or the internet’s “joke of the week.” As a result, jokes work really well as a space for bots to subvert and explore.

Similar to how we diagram and break apart sentences to see their structure, knowing that there is a fixed structure to a joke allows for bots to fill in the blanks. It’s not that hard to technically do. There are different kinds of code libraries and word libraries a bot creator can use to map related words together, such as word2vec. For example: As an entity, weather has related associations with words and descriptors like cold, hot, snow, flooding, a cold front, etc. With words already associated and mapped together, a joke bot just has to fill the word associations into the joke structure.

Roof Slapping Bot takes the roof-slapping meme (one of my favorites) and turns it into an automatic meme-joke generator. The meme’s structure lends itself perfectly for a joke bot to fill in the most unexpected and hilarious word associations within the meme’s framework, akin to a bot-like Mad Libs, automating and generating jokes every four hours.

The bot uses so many different combinations that I would never think to try. The joy I get out of this bot is similar to the joy of watching a child learn jokes, because the bot’s outputs are so nonsensical and unexpected.

The unexpected outcome is the main reason for my love of bots. A few years ago, I got to play around with a project by Tate London called Recognition. The project is somewhere online, but currently lost in a sea of link rot, redirects, and unarchived web pages; its lostness sweetens my memory of the interactions I had with it. The Recognition project uses an AI system to find visual similarities between images from two open source databases; one database consists of artworks from the Tate’s collection and the other of photojournalism images from Reuters. I select an image and Recognition displays a related but different output.

So many of the image pairings I saw didn’t quite make sense, even for machine learning. A silver orb was paired with Elon Musk, which today holds only some comedy, but some image pairings didn’t even have similar colors or shapes. They were incredibly random, but that randomness made the more cohesive pairings feel that much more special.

When the system was right, it felt incredibly right, and almost magical. One of the pairings I selected was a painting by Dennis Creffield from 1987 titled “Canterbury Cathedral.” In response, Recognition suggested a photograph taken on November 15, 2016 by Thomas Mukoya, a photojournalist. The photograph showed over 520 illicit firearms that were collected in Kenya near Nairobi.

At the time, the experience of using the tool was both under- and overwhelming, but I still think about that pairing of church and firearms. It never would have occurred to me, an artist with a background in photography, to put those two images together; but once I saw the pairing with its visual similarity, I couldn’t unsee it. It was too perfect, too much in conversation with each other. Recognition gave me an unexpected curatorial tool, creating diptychs a human never would have made otherwise, and that was the delicious surprise that made Recognition fun to engage with. Recognition could “see” potential within these datasets that I could not see.

Recognition didn’t succeed as an industrial machine learning-based tool because so many of the outputs were wrong, very wrong. But reframing Recognition as a bot-like project makes it feel like a wild success. “Bot-like” is smaller and hacker-y, in my mind, than a large-scale AI project. Recognition was bot-like in autonomously making similar pairings as quickly and randomly as possible. We can place Recognition on a scale between the “small” social media bots like Roof Slapping Bot and the sophistication of Dall-E.

I know there’s tension in my love of bots because I can’t help but disclose the bad while highlighting the good. I want an internet with bots more than I want an internet without them. Some of the solutions thrown around by social networks to clamp down on harmful bots is to have bots registered, or to change API calls to make all bots harder to exist—I think the solution could be a mixture of those two, along with more human oversight.

I worry that in a rush to quickly fix a problem, we are losing a lot of what makes the internet so delightful. I want more weird and snarky bots, a bot that responds with images of watering wells whenever someone tweets “well, actually” into a conversation thread. I want more tools that use AI and bot-like behaviors to extend our creativity as humans, like Recognition and GPT3. I want every meme to have a bot, and I still want to stumble on wonderful corners of the internet, because what makes the internet great is the weird internet. Who doesn’t want to stumble on an esoteric joke bot when doom-scrolling?