T is for Twitter Bots

Luckily for scientists, science does not have to bother itself with questions like “what is art?”. Despite this, computing has become increasingly involved in the production of things which might be called art. Generative art is a broad term describing any case where art is produced using an autonomous system (other than the human body and brain).

Most of us have, at some point encountered articles claiming that an attached script, letter or set of song lyrics were generated by an “AI”. In these cases, what we’re usually seeing is the product of something called Markov chains, where a computer program — the generator — is fed a large body of literature to “learn” from. What is easily mistaken for intelligence is really the product of a set of statistical rules about what words usually follow each-other. The larger the corpus, (the body of literature fed into the program) the better the generator gets at producing things which feel like the product of a human mind.

This is a form of machine learning rather than AI. This distinction hinges on the fact that the program is unable to learn beyond its designed restrictions. It does, however, achieve things which make it seem as if it is learning. For example, a generator trained on English language texts would learn that the word “apple” often follows “an” but never “a”. This would, in turn, give a reader the impression that the machine “knows” that the word “a” never precedes a word beginning with a vowel — while in reality the machine doesn’t even know what a vowel is.

Websites like DALL-E and Midjourney mobile app use similar methods to create what looks like a traditional painting. These generators are fed a huge body of existing artworks, and then are also able to use search engines to determine relationships between words and images there. The results can be unnervingly uncanny, as the right combination of prompt words can produce something in the style of Van Gogh which is nonetheless a depiction of a modern celebrity eating a cheese sandwich.

Not all computational creativity involves having a machine learn from existing images and text. Some programs are given generative patterns directly by their programmers. This is the case for many videogames where procedural generation is used. Procedural generation — or proc-gen — involves the creators producing a kit of components to be put together using a set of rules. These rulesets are called algorithms, and an algorithm very common within videogame proc-gen is Perlin noise. Perlin noise was developed in 1983 by Ken Perlin to enable computer-based artists to mimic the orderly-but-chaotic feel of the natural world, and its at work whenever a Minecraft player sits at their computer waiting for the game’s colourful blocks to assemble an illusory world teeming with diverse flora, fauna, and geological features.

This way of generating virtual environments is not new to videogames, nor is it the only type of procedural generation within the medium. For example, the 2019 game Wildermyth procedurally generates its narrative content. Stories in Wildermyth still follow a set of rules which allow them to be understandable and enjoyable to players. Academic fields like media and literary studies have worked to document patterns of human storytelling and expression, and now these patterns provide the generative rule-sets to produce games, images and bodies of text which feel like a person was involved.

Often, however, we see the idea of the imaginative machine become an overblown fantasy in the public imagination. Most good examples of something produced by proc-gen or machine learning are also the result of human curation and fine-tuning. As a digital artist working across games and literature, my work often gets assumed to be the product of a much smarter system than it really is. In 2016, the automated Twitter account a strange voyage went live, and through viral sharing in the following years has amassed over 40,000 followers. This fantasy world-building bot, and many more like it, is made and maintained using tools which rely heavily on the bot’s creator to provide structures and content to be randomised in a way which generates a particular aesthetic effect. They are not really any more technically sophisticated than a printed table of options randomised by a set of dice. And yet, much of the audience loves to believe that there is more going on “under the hood”.

Dr Joe Baxter-Webb is Senior Lecturer in Games Design. He is a practice-based multimedia artists and a SCANDI school Academic Enterprise Champion for the development of the local games industry.

Reference

Compton, K., 2019. Getting Started with Generators. In Procedural Storytelling in Game Design (pp. 3–16). AK Peters/CRC Press.

Frank, E. and Olsson, N., 2017. Procedural city generation using Perlin noise.

Hayes, Brian. “First links in the Markov chain.” American Scientist 101, no. 2 (2013): 252.