The Intervention of AI in the Writing Industry

Originally published in Bound Vol. 1

While we may laugh at attempts by Artificial Intelligence (AI) to write Shakespeare, or Olive Garden commercials, the emerging technology is far from stagnant. Machines lack imagination or self-awareness, but their aptitude for narrowly defined and predictable tasks means that their taking jobs in editing, publishing and even writing only requires a set of rules by which they can operate.

Machine writers exist today. Commonly used for journalism and short fact-focused pieces, they come up short when put to the task of constructing narrative fiction. Their limitations are a result of their inorganic inhumanity, their inflexibility, unimaginativeness and the intrinsic specificity of their digital and physical constructions. However it should be considered that these limitations come not from a failure of machines, but from our own misunderstandings of the nature of writing and editing, and even dreaming.

Removed from its context and considered in isolation, any given novel is a wonder of innovation. When you consider all novels ever written, there are undeniable patterns—grammar, of course, but character, plot, dialogue, theme and most importantly, audience expectations—that span the majority of works in a given subset, such as genre, or as a whole. Today, AI is used elsewhere in the publishing process—in editing and aspects of publishing particularly for market analysis. But the skills demanded here—pattern recognition, understanding of spelling, grammar, syntax, story conventions and plot structures—can be transferred into the process of writing itself.

A machine can learn such patterns to generate an outline, into which a human can intervene, or an entire text. A machine may not be able to write something as original as A Void, or as complex as Catch 22, but they are supremely capable of churning out formulaic material, considering that notable literary theorists have long established overarching plot structures (The Seven Basic Plots by Christopher Booker) present in much of fiction and appealing to mass audience sensibilities. Machines are not gifted narrative writers, but they are learning, and the skills required in other roles lay a foundation upon which we might one day see true machine writing.

Section One: What the Machines Can Do

Brynjolfsson and McAfee, researchers in digital business at MIT, see computers ‘doing for mental power—the ability to use our brains to understand and shape our environment—what the steam engine and its descendants did for muscle power’ ¹ . The modern era progresses based on three principles, which they outline as exponentiality, the rate of advancement in terms of memory, speed and global reach, the digitisation of almost everything from cameras to cartography, and recombination, the ways in which technologies, old and new, can be combined to produce new functions. At least as we understand them, AI lacks flexibility—the limitation of a machine’s intelligence is in its inability to do anything it has not been programmed to do. ² The most important question in considering the ability of AI to overtake humanity in any given task, is whether that task can be reduced to a series of direct commands. For some positions, such automation as seen in customer service and clerical work is easily realised from repetition. Anything that is repeated can be programmed into a machine—the issue with writing is that very little is repeated.

As Brynjolfsson and McAfee say in The Second Machine Age, ‘computers and robots remain lousy at doing anything outside the frame of their programming’ ³ . While computers excel at generating answers, they falter in their ability to pose new questions’. ⁴ However, things once considered impossible are now a reality. We use computers to edit and research, and drive our cars. Microsoft Word and Google discards our bad grammar. Google Translate is constantly learning from user feedback to improve recognition of human languages. So the question becomes what aspects of the writing, editing and publishing process can be automated or otherwise digitised?

There are three key areas into which AI can intervene within the writing industry—editing, publishing, and hypothetically, writing itself.

The writing ability of AI demands the effective employment of Brynjolfsson and McAfee’s three principles—the exponentially of AI learning and processing, the digitisation of the laws of grammar, writing and the digitisation of libraries of written work as a database for machine learning, and finally the recombination of skills in editing, analysis, marketing, translation, predictive text and much more.

There are three key areas into which AI can intervene within the writing industry—editing, publishing, and hypothetically, writing itself. These three aspects all serve the same end and advances in each approach the development of true AI writers. An AI that can self-edit and analyse writing, conduct market research to determine the best story to write and market itself, and then write, fulfils the functions of a writer.

Editing

As an art, editing is more organised and consistently applied than writing. There are clear laws that enable programs such as Microsoft Word, however flawed it may be for advanced practitioners, to edit the spelling and grammar of most writing. At the level of language itself, machines are quite capable of generating clear logical prose. They may not know how to write the next line, but the laws of language are such that they can perform any number of satisfactory tasks at the basic levels of syntax, spelling and grammar.

Before we leap into the idea of machine editors, we should consider how technology has already become an indispensable interface for human practitioners. There are limitations to the screen and the way people read. Studies show that it is easier to engage with material in analogue format, allowing for deeper immersion and more consistent detection of grammar and spelling mistakes. In general, digital editing results in work where practitioners detect ‘fewer proofreading errors, read fewer pages, and experienced greater fatigue throughout the experiment than participants editing in the paper condition’⁵. However the use of on-screen editing (and writing in general) allows for faster reproduction of work, transmission and coordination between multiple parties and reduced arboricide. Any process becomes much faster with functions such as copy-and-paste, control-find-replace, spellcheck and track changes.

All these processes are vastly improved by the increasing intervention of AI, making more functions and resources available to the editor and writer.

Systems such as Word and Google Translate are constantly learning through user input and improved AI. Predictive text learns how a user writes, the phrases they most use, and their common word choice. Perhaps with enough advancement, AI will be replying to our emails for us. Such services require huge bodies of information. While costly to generate, once digitised these are incredibly ‘cheap to replicate, chop up, and share’.⁶ Google Translate scans all documents it knows across relevant languages for close matches—it is not the result of any breakthrough insight about language, but simply pattern recognition. Applied to the writing process, an AI simply needs to scan all known works for relevant sentences and recombine them in new ways.

Publishing

By early 2012 ‘Google had scanned more than twenty million books’ across several countries.⁷ It has amassed a wealth of literary experience beyond any human lifespan. Such a pool of digitised information is vital for all activities involved in market analysis, audience research and research into competing titles, distribution and social media. AI can analyse incoming content—either submissions to a publisher or competing titles, to capitalise on or pre-empt market trends. Schivas states that in scholarly publishing, ‘AI has been shown to outperform humans in predicting content that is likely to create a buzz’⁸; this is applicable to AI employed in literary publishing houses in uncovering new bestsellers, or commissioning them personally.

A key example of this is the Austin-based company StoryFit, who explored the use of AI to uncover ‘narrative trends and insights that inform market and sales strategies. Ultimately, its goal is to use AI to help producers battle the challenges of today’s entertainment market’⁹. The principle behind StoryFit assumes inherent structures to writing capable of digital expression in an algorithm. This works on several levels—first, you have the overarching trends of genre and story type, of tone and character. Secondly, you have the rules of a given work itself. The style of an author can be analysed according to word choice, sentence structure and character or story conventions. Systems such as I Write Likecompare samples of text to determine which famous writer’s style they most resemble. As Monica Landers, founder and CEO of StoryFit explains:

Books hold undeniable cultural weight—and they are also data. Unlike consumer data that, even formless, has proven itself so valuable, text has inherent rules and formations. And each deviation from one rule has its own set of rules, deviations, and acceptable actions that AI can learn, compare, and predict, just like humans can.¹⁰

StoryFit, as a marketing system, uses this data to target audiences, recommendation engines, acquisition assessment and much more.¹¹ If such functions and patterns of a text can be recognised by an AI—according to StoryFit’s founder, AI are more capable than humans at identifying the underlying components of a text—then AI can use this date to operate prescriptively, rather than descriptively, guiding themselves or human writers in the production of a new text.

Writing

The potential we see for AI in marketing, investment and consumer decisions also reveals the inherent ability AI has to step into the writing process. Monica Landers, founder and CEO of StoryFit explains that computers can analyse key aspects of a story through word choice and grammar—character, theme, tone, setting, conflict types, and more—and compare these to market trends and audience statistics.¹² StoryFit uses this analysis for marketing and research purposes, but this understanding of the way stories work can be employed (and is employed by human writers) to produce unique content. As advice often given to emerging writers, so too does the recommendation to learn to write by reading apply to machines—and machines can consume more content than a human can ever read. Though AI can excel at the analysis of written content, can it itself write?

This is the field where the limitations of AI become truly apparent. According to Brynjolfsson and McAfee in The Second Machine Age, ‘we’ve never seen a truly creative machine, or an entrepreneurial one, or an innovative one’¹³. We see software that can create lines of rhyming poetry, or can write clear prose and concise journalism—but ‘never one that can figure out what to write about next’¹⁴. As Brynjolfsson and McAfee call it, this is a failure of ideation—coming up with new ideas. A thousand monkeys can bang away at their typewriters and never produce a single Shakespearean play.¹⁵

The employment of predictive text, however, suggests that AI does know what to write next, provided the rules are outlined. And what better rule is there than the sum of all human writing? As outlined by Cortes in reference to StoryFit, AI can analyse all aspects of a written work and across multiple works determine what features they have in common—across countless works, these similarities become laws by which machine intelligence can learn to write.¹⁶

Such similarities between works of writing, and particularly narrative fiction, are well established in the literary field. As referenced by Christopher Booker, Samuel Johnson once endeavoured ‘to show how small a quantity of real fiction there is in the world; and that the same images, with very little variation, have served all authors who have ever written’¹⁷. In his work The Seven Basic Plots: why we tell stories, Booker explains that when comparing two texts, ‘the resemblances between [the bare outlines of their plots] are so striking they may almost be regarded as telling the same story’¹⁸. Across disparate texts, clear patterns encode thematic and narrative elements—wherever ‘men and women have told stories, all over the world, the stories emerging to their imaginations have tended to take shape in remarkably similar ways’¹⁹. We must not assume that every story ‘fits with mechanical regularity into one or other category of plot’²⁰. Despite regularities in large scale structure, the details of stories are infinitely varied, recombined in new ways, composed of original ideas that no AI can compose. If such details were programmed into a machine, it could only produce stories that have already been told.

Yet there are more books in the world today than any human could read in a thousand lifetimes. With this library of data, a machine can recombine any word, sentence or story beyond the human ability to identify its source. A machine has potentially infinite building blocks with to generate content—by the sheer volume of knowledge it possesses no needto be original. In fact, it will be able to uncover ideas long lost in forgotten works, buried by the misfortune of publishing or marketing.

Co-authorship

The human audience still values the authenticity of lived human experience, so the novelty of machine writing, at least for the purposes of narrative fiction, seems unlikely

The idea of autonomous AI writers seems farfetched and futile—how bleak a future populated by machine who write solely for an audience of other machines? The human audience still values the authenticity of lived human experience, so the novelty of machine writing, at least for the purposes of narrative fiction, seems unlikely. Somewhat more plausible, and already present in the contemporary world, is the collaboration between machines intelligences and human writers. AI are capable of analysing market trends to inform publishing decisions, but they are also capable of providing all such data to a human writer to produce the perfect work for any given audience.

It is not difficult to envisage commissioned writers working with machine-generated statistics and guides to produce books tailored to the market. They will work according to carefully formulated templates designed to maximise marketability and audience appeal. Why take the risk of an untested novel when the machines know exactly what kind of novel the market needs and exactly what writer is needed?

As Brynjolfsson and McAfee theorise, ‘a partnership between [machine and human] will be far more creative and robust than either of them working alone’ ²¹ . Examples of this exist today. Author Stephen Marche worked with an AI programmed with a series of thematic rules and stylistic goals, formulated after the AI analysed the writing and narrative content of a series of selected science-fiction works. ²² As Stephen Marche explains in an interview on The Current, the programmers created an AI:

Used to do deep stylistic analysis of massive databases of literary text. And basically what we did together is that we reversed the process. I selected my favourite science fiction stories and then by comparing them to two other sets of texts, which are like science fiction stories and then ordinary stories, we were able to develop a linguistic system— an algorithm— that would give us a short story. And then they built an interface for me to write it onto.’ ²³

Through the analysis of fifty texts, the AI set rules such as the number of adverbs per hundred words and broader details as setting and character. Even with just a small set of texts, the AI is able to fit a new text to that catalogue, even if it requires the imagination of a human writer to make such a work stand out. As Futurist Kevin Kelly theorises, ‘you’ll be paid in the future based on how well you work with robots’ ²⁴ , based on how well you aide by the rules lain down by the machine.

Section Two: The Market

Peter Cassidy, Co-Founder of content-marketing platform Stackla, cite studies in their 2017 Consumer Content Report that found authenticity to be a major driver of consumer loyalty and satisfaction. Audiences are resistant to content that seems commercial—they ‘disconnect with brands that try to fake it’²⁵.

AI used for recommendation engines and marketing will allow content to ‘become increasingly personalized, with leisure experiences growing ever more capable of becoming interactive and adapting to audience expectations’²⁶. Schivas sees the potential for personalised user experiences though the deployment of AI systems. However ‘most consumers aren’t looking for picture perfect ads or overly produced digital experiences’²⁷; they want authentic content, and in the case of narrative writing, this is the writer’s role, the writer’s voice and brand.

Will there be a market, beyond brief novelty, for AI composed works? ‘The one thing humans can do that robots can’t (at least for a long while) is to decide what it is that humans want to do’²⁸, to decided what it is that humans are for. The industrial revolution allowed a greater portion of the population to decide that humans were meant to be artists, or writers, or philosophers—this is not a gain we can afford to lose to the current digital revolution.

You look at all of these Twitter posts and all of these Facebook posts, the billions of these things that are pouring out every day and the only thing that makes them worthwhile is that a human being experienced those things and then wrote about them. If you automated that and had machines do it, it would just be meaningless.²⁹

While Twitter and Facebook are run by AI, user engagement depends on human content. Ultimately, there is a good chance that audiences will never fully embrace machine writing. The very aspects of stories that could enable machines to write—their overarching structural similarities—represents something of their nature, and of the nature of human society and psychological instinct to tell and be told stories, that makes apparent machine intervention in their production distasteful. The kind of repetition of stories that Booker discusses in The Seven Basic Plots reflects the way the human mind works in interpreting and expressing the world. The fundamental types of stories that we tell are so engrained in the physiology of our minds that ‘these structured sequences of images are in fact the most natural way we know to describe almost everything which happens in our lives’.³⁰

Conclusion

What ultimately separates the human from the machine? In some schools of thought, the human mind in all its brilliance is simply the emergent façade of mechanically structured neurons—consciousness is nothing that can’t be explained by sufficient study of a complex system. As Milford and Stratton state, ‘what’s quite clear now is that our best-performing AI is based on how we think the brain works’³¹. Sufficiently advanced, it should be indistinguishable—somewhere in their residual imagination.

Machine writers exist today. In narrative fiction they are rare, yet in journalism they are already producing financial reports, horoscopes and sports’ pieces. Systems like Narrative Science take data and produce natural language prose just as a real human would.³² Even in areas where human writers are still the industry focus, activities such as editing, marketing, reading and analysis are ‘supported and accelerated by computers [even if] none are driven by them’³³.

While we are likely to see AI sorting through slush piles, analysing the market and fitting stories to genre convention; or enabling audience engagement and personalisation of content, and even writing material for news-sites and copy—they are ultimately tools. While useful in specific situations, they are likely incapable of replacing people because they lack generalised intelligence, being inflexible and unsuitable for situations beyond their programming and physical design.³⁴. Like all tools, the way we use machines reveals what we value. Once, they were contraptions to plough our fields and grind our wheat and forge our weapons. Now, they drive our cars and edit our grammar—but still remain incapable of doing anything they have not been programmed to do. Ultimately, the rules an AI uses to write fiction tells us more about ourselves than the AI. They reveal our biases, such as the percentage of dialogue allotted to female characters. As Marche relates, the (co-authored) story is limited by the nature of the texts supplied to the AI in the first place, by the way we are trying to use them. While the process of automation ‘doesn’t always produce the best writing, it can teach us a lot about the essence of fiction’³⁵.

These trends in the publishing industry are driven in part by attempts to compete within an increasingly global and automated economy. If, as some experts believe, automation enables humanity to enter a utopia of artistic, scientific and athletic leisure, (Milford, M & Stratton, P 2017), the publishing industry will be freed from concerns of market consumption and investment return, and instead be enabled to focus on the pure artistry—safe from AI automation by virtue of its inherent lack of profitability.³⁶ As Kelly concludes in his 2012 speculation, ‘they will let us focus on becoming more human than we were’³⁷.

Notes

1 Brynjolfsson, E & McAfee, A 2014, The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies, W.W. Norton & Company, New York, pp. 7–8.

2 ibid.

3 ibid.

4 ibid.

5 Wharton-Michael, P 2008, ‘Print vs. Computer Screen: Effects of Medium on Proofreading Accuracy’, Journalism & Mass Communication Educator, vol. 63, no. 1, pp. 28–41.

6 Brynjolfsson, E & McAfee, A 2014, p. 61.

7 ibid.

8 Schivas, J 2019, ‘Don’t rage against the machine! How scholarly publishers can embrace AI’, The Bookseller.

9 Cortes, S 2017, ‘AI and how it’s revolutionizing the book publishing industry’, LSC Communications.

10 ibid.

11 ibid.

12 ibid.

13 Brynjolfsson, E & McAfee, A 2014, p. 191.

14 ibid.

15 ibid.

16 Cortes, S 2017.

17 Boswell, J 1831, The Life of Samuel Johnson, LL.D. Including a Journal of a Tour to the Hebrides: In Five Volumes, vol. 5, ed. Croker J W, Murray, Oxford University.

18 Booker, C 2005, The Seven Basic Plots: Why We Tell Stories, Bloomsbury Publishing Plc, London, pp. 7–8.

19 ibid.

20 ibid.

21 Brynjolfsson, E & McAfee, A 2014, p. 193.

22 Marche, S 2017, ‘Twinkle Twinkle: I enlisted an algorithm to help me write the perfect piece of science fiction. This is our story’, Wired.

23 The Current 2018, ‘Can an algorithm make science fiction better? Author Stephen Marche finds out’, CBC Radio, Quebec City.

24 Kelly, K 2012, ‘Better than Human: Why Robots will—and must—take our jobs’, Wired.

25 Cassidy, P 2017, Survey Finds Consumers Crave Authenticity - and User-Generated Content Delivers, Social Media Today.

26 Lamberti, F, Sanna, A & Montuschi, P 2015, ‘Entertainment Technologies: Past, Present, and Future Trends,’ Computing Now, vol. 8, no. 2.

27 Cassidy, P 2017.

28 Kelly, K 2012.

29 The Current 2018.

30 Booker, C 2005, p. 8,

31 Milford, M & Stratton, P 2017, ‘The future of artificial intelligence: two experts disagree’, The Conversation.

32 Kelly, K 2012.

33 Brynjolfsson, E & McAfee, A 2014.

34 Milford, M & Stratton, P 2017.

35 The Current 2018.

36 Milford, M & Stratton, P 2017.

37 Kelly, K 2012.

Joel Humphries28 October 2021