Halloween approaches. Night falls early and the wind howls menacingly. Leaves swirl and crackle along the ground. Thoughts turn to Dracula and Frankenstein. Meanwhile, novice writers huddle in their homes and gear up for NaNoWriMo. All of which conjures up the question: Could a software program write a novel? Yikes! Open those pod bay doors, Hal, and let me out! Any time we consider having an automaton perform an acitivity that we have always regarded as distinctly human, it sends shivers down our spines. But let’s close our eyes for a moment, take a deep, calming breath and explore this a bit.
First, what would it take for a computer program to create a believable work of fiction? Certainly it would need some state of the art linguistic processing capability. Not just the ability to create syntactically and grammatically correct prose, but also an understanding of semantics. And that’s just the foundation. On top of that Robo Writer would need the ability to conceive interesting stories, create memorable characters and animate them with credible dialogue and behaviors. Could the same kind of fallible logic that frustrates our use of the word processor and spreadsheet possibly do any of this?
Well – not yet. But the technologies required are continually evolving. An example of linguistic processing improvement is WhiteSmoke. WhiteSmoke software analyses text on-the-fly or at the user’s request and suggests grammatical improvements, amends spelling and enriches text through suggesting alternative or additional wording. What is interesting about WhiteSmoke’s approach is that it is supported by an online database that constantly crawls Internet sites for common usage of English. It uses that knowledge to edit prose based on the type of English style selected, for instance commercial, legal, medical, casual, creative, executive and even dating.
Beyond these baby steps, one can imagine databases of characters, built up from standard physical and psychological profiles, similar to those used in law enforcement. Drawing on the work of Joseph Campbell and Carl Jung, other databases of archetypal story themes could provide templates into which scanning software could plug items plucked from news sites, blogs and other social media to generate stories.
The ultimate way to decide whether a computer had arrived at literary competence would be something like the Turing test. Alan Turing, one of the father of the modern stored program computer considered the question of how we could know whether a computer had achieved human scale intelligence. He devised a simple test. A human sits in a room and can communicate with another entity on the other side of a wall by exchanging written messages through a slot in the wall. The human participant can ask the entity any kind of question to determine whether it is human or machine. If, at the end of the questioning, the human can’t tell whether the entity is human or machine, then effectively the computer program has passed the Turing test and achieved the level of human intelligence. In fact, there is a contest, held each year, to determine whether any software has met this criteria. The winner of this annual contest is the best entry relative to other entries for that year, regardless of how good it is in an absolute sense. (No one is claiming absolute victory yet.)
We envision two Turing literary tests for a computer generated work of fiction.
- Soft test – Human readers can’t tell it’s not human generated.
- Hard test – Human readers not only can’t tell it’s not human generated, but they’ll actually purchase it.
For the moment, let us suspend our judgment about whether we want computers writing books and see where such a technology might take us. Here are some speculations.
- Publishers could eliminate those pesky authors. (Well, maybe not entirely. Authors, in addition to their writing contribution, are also the most important part of the marketing equation for a book. Perhaps Disney could chip in some animatrons.)
- Authors could use software programs as writing assistants. Few authors are expert at every aspect of writing: story, scene structure, character, dialogue. Let the software do some of the heavy lifting e.g. fleshing out the minor characters or adding polish to the dialogue. The good news is, software programs won’t ask for title credit.
- Reverse the process. If a program can write a novel, couldn’t it also read a novel. Goodbye slush pile! Publishers could employ armies of “robot readers” who would funnel the good stuff up the food chain. They could also take a manuscript that had been accepted for publication and rework it into a salable work.
- Mass customization. Who says a book has to be the same for every audience. Let the software add subtle cultural nuances for different audiences in different countries. Kind of like the way McDonald’s tweaks it burgers for different tastes in different countries.
There is precedent. We’ve already seen technology invading the film and music industries. In each case, the human element isn’t replaced, but empowered. So Robo Writer may just turn out to be a friendly Frankenstein. Maybe some wealthy, aging patron, seeking a bit of immortality, could offer a substantial prize for the first work of fiction written entirely by a computer program, to pass the hard literary Turing test. Who knows, maybe one of this year’s NaNoWriMo contestants may turn out to be disembodied bit of logic. On the Internet, nobody knows you’re not human.
- The Quiet Revolution in Book Writing
- A Second Life for your Blog Novel
- NaNoPreMo – a pecha kucha warmup for nanowrimo