If on a Winter’s Night a Computer

It’s 1979 and Italo Calvino’s If on a Winter’s Night a Traveler is published. It is revelatory! The novel is lauded for its “post-modernist” narrative, “labyrinthine” structure, and ability to play with the notions of being “a reader.” It’s all very new and cool and futuristic. What draws less attention is the novel’s vision of AI authors, their potential role in creating apocryphal human art, and Calvino’s use of an algorithm to formulate the plot.

It’s 2016, and the art world is graced with its first piece of AI apocrypha when developers create a “new” Rembrandt portrait, except it isn’t by Rembrandt. The Next Rembrandt, as it is known, is a computer-generated 3D print-painting created using a facial recognition algorithm informed by data from 346 paintings by the man himself. It is comprised of 148 million pixels and based on 168,263 fragments of works by Rembrandt and, by referencing all his other portraits, makes a computer’s best guess at what he might have painted next. It is—I am disappointed to say—fairly convincing. But using an algorithm to create a picture is a very different process than writing a book.

If on a Winter’s Night is punctuated by the handiwork of one of its main characters, a fiendish translator called Ermes Marana, who puts the literary world into chaos by swapping translations and inserting apochrypha—his intent being to show fiction’s “pretense, misunderstanding, falsehood.” He is keen to get his hands on the latest (incomplete) manuscript by the world-famous author Silas Flannery, who is sitting on a mountaintop in Switzerland watering his zinnias and suffering from a paralyzing bout of writer’s block—much to the annoyance of Flannery’s agents, publishers, and advertisers. A commercially successful, if somewhat formulaic writer of thrillers, Flannery is an ideal candidate for the organization Marana works for: the Organization for the Electronic Production of Homogenized Literary Works (OEPHLW). Marana succeeds in stealing Flannery’s incomplete manuscript and feeds it into a computer; using this data, the computer is able to complete the manuscript “with perfect fidelity to the stylistic and conceptual modes of the author,” and thereby, has the capacity to continue to produce Flannery apocrypha.

When If on a Winter’s Night was first published, the idea of an AI author was but a twinkle in the human eye. William Gibson had not yet uttered the term “cyber-space.” The idea was speculative science fiction. But as with a lot of science fiction, it has ended up coming true: We are now witnessing the glittering inception of the AI author, and computers are taking their first clumsy steps towards literature, meaning that writers are facing the exact reality imagined by Calvino for Silas Flannery.

George R.R. Martin, writer of the Game of Thrones series, is only human, and because he is only human, he is taking a long time to deliver the next GoT installment, The Winds of Winter. Too long, some feel. Having waited since 2011, fans are so eager for his sixth book that one—Zack Thoutt, a software engineer—took fan-fiction to a whole new level: He designed a program to create it. To do this, Thoutt used a type of AI now typical for the creation of literature and computers’ comprehension of natural languages called “recurrent neural networks” or “RNNs.” These are machine-learning algorithms that mirror the neural pathways of the human brain, meaning that sequences are linked and have the ability to loop back to previous information, and therefore, inform the next sequence. The result is “obviously not perfect,” Thoutt tells Motherboard, but having input 5,376 pages of the previous five books, his program has not only successfully strung a lot of sentences together (five chapters’ worth), it mimics Martin’s style and lexicon, and makes plot predictions that have already been circulated by hardcore GoT fans. It also spouts some absolute nonsense:

“I feared Master Sansa, Ser,” Ser Jaime reminded her. “She Baratheon is one of the crossing. The second sons of your onion concubine.”

Imperfect as the results may be, Thoutt’s algorithm—like Marana’s—is mining content from an existing database, which means it had a head start. Creating original text is a more difficult matter, as Angela Fan, of Facebook AI Research, tells New Scientist: “[AI programs] write in a very simplistic way, deciding word by word what to say next…staying on topic is quite difficult for neural model networks because they have no explicit memory.” Fan’s team trains their algorithms to stay on topic by using writing prompts, such as “Aliens start abducting humans,” while Mark Reidl of the Georgia Technological Institute uses a different approach: focusing on climaxes (like a marriage, or a death of a character). Both Fan and Reidl use recurrent neural network programs and both have been successful in creating original short stories that remain on topic. But in January of last year, Fan and Riedl’s short story benchmark was surpassed with the novella The Day a Computer Writes a Novel, an original text written by AI that proved to be a convincing entrant in a Japanese literary competition. The novella’s characters may have “needed developing,” but the judges didn’t suspect that a computer had written it.

In If on a Winter’s Night, the aim of conflicted Silas Flannery is to “capture in the book the illegible world, without centre, without ego, without I.” This, supposedly, is something that AI can accomplish without too much effort. It is also something that Calvino attempted himself—probably with considerable effort. Not satisfied with merely theorizing about algorithms able to write novels, Calvino wrote If on a Winter’s Night using one: a semiotic square. Calvino was a member of the Oulipo group—Ouvroir de Littérature Potentielle, or, Workshop of Potential Literature—a selection of mathematician-writers and writer-mathematicians who relished “the seeking of new structures and patterns which may be used by writers in any way they enjoy.” In La Bibliothèque Oulipienne Volume II, Calvino explains that he used a code of his own, employing the model of the semiotic square to create a sequence to structure the plot of If on a Winter’s Night… .

The coding used to create the kind of sequencing capable of producing works of fiction is infinitely more complicated than Calvino’s semiotic square. Nonetheless, AI is everywhere. It is free to those who can afford it, and free to those who can’t. It is our chat bots. It is our social media feeds, our maps, our banking, our advertising. It is in our inboxes. It is our news. (Associated Press, Yahoo, and Comcast all use the “natural language generation platform” Wordsmith.) And, on the whole AI, in partnership with data, has been successfully harnessed to improve business life, to make everything quicker and better and faster and shinier. But…can data touch the human soul? Ron Augustus, director of SMB Markets at Microsoft, and part of the Next Rembrandt team posits this as a possibility. AI has already passed a Turing Test-of-sorts by duping a large proportion of unsuspecting public into thinking that its poetry was written by a human; and poetry, surely, is the most secure line to the soul.
I am a coal-truck

by a broken heart

I have no sound

the sound of my heart

I am not
This is an example of one of the better poems written by the AI poet developed by Microsoft’s Kyoto researchers. It certainly looks poem-y, the choice of vocabulary is interesting, and it has that Thomas Chatterton melancholia we tend to associate with poetry. But does it actually make you feel anything? The reason literature is currently an area that “progress” has been unable to infiltrate is not that it is “skilled labor,” but that it is human. Something that is not human could make your heart beat faster by writing a perfect thriller plot. It could possibly turn you on by writing a few racy sex-bot scenes. But could it touch our sublime instrument? I’m not convinced. It could have an effect on copyright, however.

The impact AI authors will have on writers in regard to copyright and intellectual property was discussed in articles in the summer edition of The Author Magazine and in last year’s WIPO magazine. In the U.S., to be subject to copyright an original work must have been created by a human being. But as Andres Guadamuz, senior lecturer in intellectual property law at the University of Sussex says, “Things are likely to become more complex as the machines get better at producing creative works, further blurring the distinction between artwork made by a human and that made by a computer. When you give a machine the capacity to learn styles from a dataset of content, it will become ever faster at mimicking humans.”

This “dataset of content” is important. In October, The New York Times profiled Robin Sloan, an author using AI to his advantage. With his latest novel, Sloane is assisted by a machine-learning algorithm: he takes notes such as “The bison are gathered around the canyon,” hits a button, and the computer concludes the sentence with “by the bare sky.” Which Sloan thinks is, “kind of fantastic.”

Sloane’s machine was initially given a dataset of science fiction stories from the ’50s and ’60s to draw from; now it is also informed by the likes of Steinbeck, Didion, and Johnny Cash’s poems—ostensibly a great selection. But when it comes to machine-learning algorithms, who chooses the dataset—who chooses the “good books” that inform the AI? And what about the bad books? The books we read and hate or are bored by still influence us in positive ways. By eliminating these lesser books, the selective process becomes like the echo chamber we are so familiar with in social media—built by algorithms. Furthermore, is a computer influenced by someone’s work different to a human being influenced by someone’s work? It is, of course, more calculated. Rather than being influenced or inspired, it steals—yes, “like all great artists.” The author Nick Harraway says, “We are not at risk from the rise of the robots. We are at risk of exploitation by companies and individuals who unthinkingly regard the complete, finished text as a found object and resent the idea that they should have to share the vast proceeds of its digital exploitation.”

The Copyright Office in the United States will “register an original work of authorship providing it was created by a human being.” But Hong Kong (SAR), India, Ireland, and the U.K. grant copyright to the person who “made the operation of artificial intelligence possible.” So as with Microsoft Word, it is the person who uses the program, rather than the programmer who retains copyright. But as Guadamuz points out, with the use of this particular type of AI, the person who uses the program is often doing little more than pressing a button. This could have a huge impact on the livelihoods of authors and their translators, as programs are fed enough information from existing literature to generate works that are original enough to evade copyright laws.

In If on a Winter’s Night, Calvino also introduces the concept of “reading machines,” which can not only read a book but also judge its merit. We can assume that this will also become a reality eventually, a depressing or exhilarating thought depending on where you stand. But for those of us who wish to remain autonomous, all is not lost; hard work got us so far, but it was not efficiency that got us this far. It was in the moments between productivity that genius was inspired. In If on a Winter’s Night’s closing pages there is a scene in a library where one reader says to the others, “If a book truly interests me, I cannot follow it for more than a few lines before my mind, having seized on a thought that the text suggests to it, or a feeling, or a question, or an image, goes off on a tangent and springs from thought to thought, from image to image…The stimulus of reading is indispensable to me…even if of every book I manage to read no more than a few pages. But those few pages already enclose for me whole universes, which I can never exhaust.” Our avenues of idle thought have led us to some of our greatest achievements and our most profound sentiments. We are “infinite in faculty,” as Shakespeare said. The AI author might be around for a while, and things might get pretty hairy for human writers, translators, and journalists for a while, but eventually, this too shall pass.