Melding the disparate worlds of art and computer science, Andreas Refsgaard and Mikkel Loose have developed a fascinating AI project called Booksby.ai, an online bookstore entirely generated by artificial intelligence. Every aspect of the site is generated by machine learning algorithms, from the entire books and accompanying cover artwork, to the reviews and pictures of people reviewing the books. And on top of that, all the books are actually available to buy on Amazon.
Andreas Refsgaard is an artist exploring creative uses for machine learning. Based in Copenhagen, the Booksby.ai project is a collaboration between Refsgaard and data science student Mikkel Loose. The goal was to have some fun with the idea that AI is taking over human jobs by creating an online shopfront entirely generated by machines.
“The aim was to build a whole coherent project where an AI would generate everything and people could buy its creations afterwards,” Refsgaard explains to New Atlas in an email. “It had to have a physical output in the end, and we thought about a few different options, but ended up with the bookstore as the perfect setting, because all the different parts that go into a book could be generated using machine learning models for text generation (main text, author names, book description, reviews) and images (book covers and pictures of people reviewing the books).”
The duo were not interested in generating a new machine learning model, but instead used the project to aggregate a variety of different preexisting models into a singular outcome. So, for example, the books and accompanying reviews were generated using a freely available character-based recurrent neural network called char-rnn. The images of the reviewers faces were generated using a different model, and the book covers used yet another model. Even the books’ prices were set by a neural network trained on book prices from Amazon.
Refsgaard admits the books presented on the site are not the most sophisticated examples of algorithmically generated creativity. Many of the descriptions sit on the fringes of incomprehensibility such as this synopsis for a book titled, The Linninging People’s Story by Kliphinia Tysenbart:
“In a distant future in the 1960s. The trail is Kate and Wallace to confront the end of his new home. Now the truth is the most extraordinary truth about the end in a family of interest to rescue the trail when a love for his heart goes with Robin Harrison. Now she discovers she must uncover the truth about Sam O’Hara becomes a family who ends up and wants to want to leave her family for what it takes to uncover the deadly and long-ago victims until the truth becomes the truth.”
“Our project definitely falls into the charmingly absurd category,” says Refsgaard, discussing the sophistication of the algorithms used, “but there are examples of algorithms out there, that for text produce way better results. Something like gpt-2 by OpenAI is already a huge improvement, and you would be able to create a shorter text like a review or a newspaper article that quite accurately resembles a ‘real one’ for sure if you try for long enough.”
One of the more interesting aspects of the project are the reviews. Refsgaard suggests this element generated several surprises for the creative duo, with the algorithm spitting out content that almost accidentally resulted in sharp critical reviews. One particularly meta review seemed to imply a book was barely coherent while another invented a new word (“Knodung”).
“…the reviewer comes up with this new word “Knodung” that has a really nice ring to it, even though I have no idea what it means!” Refsgaard says. “When you train a character based recurrent neural network (like char-rnn which we used) in the beginning of the training process the network will make a lot of spelling mistakes, and almost feel like it is learning a whole new language and just writing random things that sort of sound right, but are in-fact just gibberish. So you could argue that a word like “Knodung” might be the result of a poorly trained network, but to me these “mistakes” are the hidden gems! It makes me wonder what Knodung means, and is far more interesting than if it had been something like “great” or “worth reading”.”
The Booksby.ai project is admittedly not the greatest example of machine-driven creativity, but the site presents an intriguing example of an entire ecosystem created by artificial intelligence. In fact, the only point in the project where humans come into the picture is if someone is interested in actually buying one of these AI-generated books.
Incredibly, Refsgaard and Loose managed to get the books onto Amazon as real products. The duo expected the books to be rejected by the massive online marketplace as spam, but to their surprise they are up and available for purchase. And have they actually sold any of these books?
“We have sold 19 books on Amazon so far,” Refsgaard says. “Being a writer is a tough job, even if you are an artificial intelligence…”
We are assuming the books have been purchased by living human beings and not a clever AI with a wealth of cryptocurrency and an interest in bizarre science fiction.