We expect literature to be evocative, to tell a story or create a picture, and use words to invoke an emotional response in the reader.
We have looked at how AI-powered machines might approach creating music and paintings. In this article, we will explore how machines could create literature. Attempts to define what literature is end up trying to certify what is good literature, and what is not good enough to be literature. So we’ll simply observe that not all writing is literature. The user manual you get when you buy a phone is an instance of writing, but not of literature.
We expect literature to be evocative, to tell a story or create a picture, and use words to invoke an emotional response in the reader. If the story of Harry Potter were written as a drab news piece, for example, it may not elicit any emotional response from you; but when you experience a little boy’s wonder at encountering the world of magic, you begin to care about this character, and thus you also begin to experience literature.
How can we come up with AI capable of creating such writing? The high level answer is the same as last week: the AI program would need access to the existing instances of literature. It will need to scan the text in each of the works, identify turns of phrases, methods of describing characters, of mood and scene creation, and more. The AI will also need to understand character creation: what makes interesting characters? How are people different at various stages in life? How does that impact their motivations and behavior? Essentially, the AI-machine will need to understand humanity by ‘reading’ about it extensively.
But that alone will not be enough. Human novelists and poets typically need a lot of practice before they start writing good work. A machine will not be any different, and it will also need to practice writing stories and poems, and observe how its works are being received. The process is already underway in various forms. Early last year, a Japanese AI program authored a novel that competed successfully against other entries by human writers, and passed the first round of screening for a national literary prize. This is significant. The people reading the novel could not tell that a machine had authored it.
Similarly, AI has been successfully writing poetry for some time now. A website called Bot Poet, for example, presents you with different poems, some written by machines, some by humans, and you need to identify if the poem was written by a machine or a human. As you might expect, it is often tricky to tell the difference.
To be fair, though, it is somewhat gimmicky to focus only on being able to tell machine creations apart from human creations. Literature has many abstract, intangible qualities, but ‘must look as if created by a human’ is not one of them. The real question is: are machines writing literature that can invoke an emotional response? We are certainly far from that, and among all mediums of art, literature is probably the hardest for machines to tackle, especially compared to music.
To create a worthwhile work of literature, as we have seen, the first step for a machine is to understand what literature is, by ‘reading’ the thousands of books already in existence. but the second part is even more important: machines need to figure out what to write about. Human writers tend to write as a reaction to something. This may include personal experiences such as heartbreak or depression, or social phenomena such as terrorism, communalism, cultural and social changes, etc. When an author addresses such experiences in her writing, it appeals to a wider audience, since they have such experiences and concerns in common.
For a machine to write enduring, appealing literature, it needs to first work out what to write about. This might seem impossible, since a machine cannot have any experiences of its own, but ultimately, such experiences are simply data points. AI machines can scan newspapers and press releases, for example, identify that terrorism is a growing problem (this can simply be detected by counting the increasing number of mentions of terrorism in news), identity the various stakeholders and the roles they play, and thus pick what to write about.
You might feel like that is a reductionist or limited view of creativity, but all major Hollywood studios and most writers already use such market research to figure out what to write about and what will appeal to people. A machine will simply do it more efficiently.
*This article was originally published on May 1, 2017 in Telangana Today, and can be accessed at Can machines produce literature
Like most of you working in tech and startups, I absolutely loved reading Peter Thiel’s Zero to One. It’s a stunningly thought-provoking book, with a deep new insight pretty much on every page. It’s hard not to be impressed by the clarity of thought as well as the...
I’m sure we have all heard of that fraud where tourists walk into a restaurant, order some basic food item-dosa, idli kind of thing-and are presented with a bill for thousands at the end of it. And the swindlers at the restaurant claim that the tourists should have...
What are some things you are aspiring to do this year? Is traveling anywhere on your list? Almost everybody thinks they need to travel. Not convinced? Log in to Facebook, and see what your feed is showing. What friends of yours are vacationing currently? What pictures...