Writing and Reading in the age of AI

Like many people who dreamed of a career in publishing, I’ve always loved books. For years the allure of the author enchanted me: slaving away for months or years on a manuscript before finally handing it over to a publisher. What happened to it after that was somewhat of a mystery to me, and it wasn’t until university I realised it takes a village to raise a book. It passes through many human hands before it ever reaches a reader, and when done well, the reader will never know.

The danger with this mystique is the process may seem unnecessary – to the reader or the author. The rise of AI has only added to the problem, with many tools promising to do the work of editors, proofreaders, typesetters, and designers – even the work of writers themselves.

Historically, standing in the way of technological advancement has been a fool’s errand, often the butt of jokes years later. So is there any point in trying to stop this? Can we? Should we? The pitfalls of rapidly advancing technology are often not examined or discovered until years later when they became entrenched and difficult to change. AI is still in its infancy. The more we know about how our favourite art and books are made and how AI (poorly) imitates this, the better.

The logical place to start is asking how Large Language Models (LLMs) like ChatGPT actually work. These programs are usually fed large samples of data (be that books, internet forums, social media posts and blogs like this one) which it searches for patterns of language. This makes these programs excellent at spitting out the most standard answer for simple questions or prompts. Not only can it not produce anything new, but it will always seek the average answer. If fed bad data, that answer could be wildly inaccurate, or in the best-case scenario, simply boring.

When applied to the task of writing, LLMs shave off all the distinguishing features that make a work unique by design. Even if trained exclusively on an author’s previous own work, it will reproduce their most common patterns and tropes in a Frankenstein-esque amalgam of style. And that is charitably looking at the consensual use of an author’s work to try and produce AI texts.

In September of 2023, a group of high-profile authors including John Grisham, George R.R. Martin, and Jodi Picoult filed a lawsuit in US court suing the owners of ChatGPT for copyright infringement – for training their LLMs on the authors’ material without their permission, and certainly without licencing the material to do so.

The use of LLMs to produce writing has been the most high-profile discussion happening around publishing and the new technology, but what about its applications that are less visible? The same problems with LLM’s inability to write outside the box persist when applied to editing a work, perhaps in more insidious ways. It is far easier to accept a suggested word or phrase without thinking, or to simply tell one of these language models to edit a story to its standards without examining the changes it’s making. It’s cheaper, it’s easier, it’s worse. As these tools become more integrated with our writing software it would be all too easy to cut editors out entirely.

But what do editors really do for a book? A LLM may eventually be able to technically correct the grammar and syntax of a novel, but what about cases where less than perfect choices add to the work itself? Looking at the subgenre of hilarious predictive text fails and some of the more questionable suggestions from your word processor should be enough evidence that the most common answer isn’t always the right one. A novel isn’t written the same way as an instruction manual, and a thriller might be written in a very different style to a romance novel. Editors can understand intent and style in a way that a LLM likely never will, and work to help differentiate an author’s voice rather than standardise it.

The greatest challenge from LLMs and AI is ultimately not their newness, but the fact that they are cheap and fast. How could an author or an editor compete with effectively free and the ability to spit out writing in seconds? Is the end result just a race to the bottom as endless AI content is churned out, leaving human-driven content to be buried under the avalanche? It’s certainly a concern, and the effects have already been felt across many areas of publishing.

Online marketplaces have been flooded with low-quality AI work attempting to scam readers into buying what looks like a legitimate work with potentially life-threatening results. AI-generated mushroom gathering books told readers to pick deadly mushrooms for consumption. While the original works identified were removed from sale on online sites, new suspected works started cropping up faster than they could be reviewed … spreading like a fungus.

Given the ease and relative lack of visibility through which AI can permeate into writing and reading spaces, the outlook is grim but not hopeless. As an industry and as readers, we can vote with our capital and stay informed on what goes into what we read, even supporting human-driven work if it’s more expensive.

Here in Australia, there has been a consistent shift towards shoppers buying free-range eggs despite higher price tags, with cage eggs now only representing 35 per cent of all sales. In 2001, only eight per cent of all eggs sold were free range. Closer to the topic at hand, ebooks were predicted to completely replace print books. However, print sales continue to make up the vast majority of all book sales.

We all want to keep quality books, film and art alive and thriving into the future, and it’s important to remember that their downfall is not inevitable. Like any technological advance, LLMs are not going away. They inevitably will form part of the creative process in the future, but we can control which parts.

Sarah Neilsen
Editing Intern