I wrote a book with ChatGPT — debunk the hype and get to work.
ChatGPT has a role to play in every writer’s workflow, but it ain’t no magic wand, kids.
ChatGPT is never going to write a book that anyone wants to read. Period.
Stop hoping, stop dreaming, and stop thinking there is a way to write a book that doesn’t involve you — a person — writing lots of words and organising them into a structure that resembles a book. Period. Full stop. C’est tout.
No matter what you do, the book is inside you, not the large language model or the neural nets of whichever AI tool you choose. Getting the book out of your head and onto the page is up to you. It’s that simple.
You may have seen articles (Medium is full of them) saying ChatGPT wrote a Steven King novel, or poems, or a whole bunch of stuff but it’s not true. Or rather, what it produces is a dumb facsimile of the source material. If you want a book with no meaningful plot arc and generic clichés, ChatGPT could get you about halfway there, maybe, but you’d need to do a lot of work to make it read like a real book.
Artificial intelligence and intelligence are very different things
I have been testing this out in-depth. I have been asking ChatGPT to emulate guests (on my various podcasts) and answer the same questions as the actual guest. It has never beaten a human in the answer, even though it has answered in almost exclusively that same human’s words. Yes. ChatGPT will locate a person, digest everything they’ve written on the web and serve up a summary word salad of clichés and headlines, not a real answer.
I have just published a book about the history of zombies and zombie films. I used ChatGPT to help me. I asked it to give me a list of zombie movies with synopses, information and trivia. What it actually gave me was dross. I asked for 50 movies, I got 27 and about 18 duplicates, plus a few that don’t exist. I asked for synopses — it replicated the same linguistic structure for each, the same turns of phrase, the same lame hyperbole. It also made up a lot of weird crap. It put Hollywood stars into every movie, even Sri Lankan musical dance zombie movies.
It offered me nothing but American movies at first. I had to work on my syntax and questions to get the needed information for each movie, which proved so unreliable I had to double-check everything. EVERYTHING. Imagine you’d tasked a TikTok scrolling 15-year-old to be your research assistant, and they’d got bored halfway through Googling for movies and just copied and pasted any old crap of some average Joe fanboy blog. That would be about the same result quality as I got from ChatGPT for my Zombie book.
And here’s why…
The opportunity — and value — behind the hype
Generative AI uses transformer neural networks (the ‘T’ in ChatGPT) and ‘large language models’ (LLM). Neural networks sound complex, but at a very basic level, they are data tables that predict the likelihood of what will come next in a sequence of numbers (the numbers could represent anything, like words or functions). This is how voice assistants like Alexa determine functions. For example, “Alexa, play rock…” will lead to Alexa’s neural network predicting “music” as the next most likely word, enabling it to identify the context action word ‘play’ as a music request, not a game request or video display request. This is vital in language recognition because if Alexa’s mic doesn’t hear the whole command word for word, it can still make a strong guess at what most of it means.
Transformer neural networks take this a step further by using vast databases of language — libraries full of books, magazines, newspapers, transcripts, etc. — to understand the intent of a whole sentence, not just a word at a time. So ChatGPT can take questions like “what is rock music?” and use its LLM neural network to work out the intent of the whole question, which is really asking for multiple things such as a definition of the genre and some examples of classic songs.
In that sense, ChatGPT-style AI is a simplified interface to access, query and manipulate complex data sets and computer functions — making advanced computing possible for almost anyone. The same is true of AI image generators, which use simple text to control an array of complex imaging software that would normally need a skilled graphic designer to use. A simple interface for complex systems? That’s got huge potential for both time-saving and skill enhancement. For example, the image above (and the one below) was made from my using the following prompt (and variations) in an AI called Midjourney.
/imagine a typewriter made of jello :: transparent surfaces, iridescent, glossy render, HD, Photorealistic, studio lighting, product photography, highly detailed, acid colours, Canon EOS 5D Mark IV camera with a Canon EF 14mm f/2.8L II USM lens low-angle tripod — ar 16:9 — q 2 — v 5.1
This is how AI works for writers — as a digital intern
Workflow. This is how to use AI platforms for a non-fiction project, using AI as a research assistant — and how to maximise the time saving and avoid the time-suck.
Step 1: Plan your project carefully
Yes, it’s boring, but if you just leap into action and start asking ChatGPT to do this and do that, you’ll end up making the same mistakes I did. Spend some time planning your project the same way you would without the AI. I found it really helps to have an action plan — how much? Very much.
Here’s why. I wrote the first book — a non-fiction movie guide called How to Be an Instant Zombie Movie Expert — in about 5 weeks, and I reckon the AI saved me about a week of research and compiling content. After that, I reviewed the process and started another project, the second non-fiction title in the series — How to Be an Instant Vampire Movie Expert — and I have already saved at least double that by planning how and where I would use the AI.
As you experiment with the tools, you’ll figure out new ways to use them effectively, feed your trial and error learnings into a project management spreadsheet of some sort — save your prompts, re-use them, adapt them for new tasks — keep a reference of how you got your results and re-use this next time.
Step 2: Be specific in the way you describe tasks (prompts)
Don’t ask for generic data or lists because you’ll get a bad result with errors and duplicates. Prompts like “Give me a list of zombie movies” will return errors and even fabricate movie titles that don’t exist. However, if you understand requirements at a granular level and ask for specifics, you get better quality results.
Remember, there is a fundamental lack of diversity in ChatGPT — in fact, all generative AI platforms have a statistical bias towards white American and European content because statistically there is a lot more of it in circulation than non-white art from other continents. This is because mass-market publishing and art markets evolved in Europe first, and mass media film and TV in America first — the rest of the world is playing catch-up in terms of volume. Sadly, this means there’s a bias towards American and European movies, authors, actors, writers, etc. So if you want a list of zombie movies that includes Asian titles, you need to ask specifically for Asian zombie movies to compensate for the fact they are a small dataset within the data category you are asking for.
However, if you get this right, you can save a lot of time. I got more granular results by asking by continent, language, popular genres and studios/movements (like Bollywood or Hong Kong Cinema). This needs a little more sorting to organise them into a curated list, but it still saves a lot of Googling the whole lot from scratch. So take the time to identify the vectors that describe your project — for non-fiction — and you’ll find ChatGPT can be a really useful assistant.
Step 3: Sort, organise and enrich your data — but don’t get complacent
You can definitely save a moderate amount of time using the ChatGPT engine to organise your data. In my case, I used it to index the content, took all the data outputs (a list of about 70 movies in the end) and asked the AI to put them in chronological order and cross-check each one for awards, find trivia and stories about each title and so on. It did a much better job of organising movies into chronological release order than identifying cast and crew, but it was pretty accurate except for some eye-watering howlers, which were impossible — like casting actors before they were born or just making up names. I think it got confused between originals and remakes at times too. So make sure you fact-check everything.
Simple data enrichment like that works okay, although fact-checking means it doesn’t save you as much time as you probably hoped for when you started. However, where it goes seriously awry is asking it to give you complex tasks, like a plot synopsis. There is a prevalence in LLMs (large language models) of bland, generic content, clichés and hype. Bear in mind the web and much of the data it accesses is keyword-loaded SEO marketing copy, thin on facts and rich in semantic word-washing. Every zombie film, according to ChatGPT was described in more or less the same terms as a “classic example of the genre” “considered to be one of the best” “a must-see for fans” and so on.
Worse than having to cut out all the generic marketing dross, the actual substance of most of the synopses was factually wrong. It usually gets names right but often confuses minor characters for major protagonists or swaps hero and villain roles around. It’s quite funny at times but useless. I had to fact-check everything again. And of course, make sure I watched the actual movies.
Step 4 (if you can be bothered): Loop back to step 2
Obvious, perhaps, but when you reach the limits of your patience with the gobbledegook and word salad, remember to apply the same thinking as step 2 in the process — be specific. If you ask “Who were the cast and crew of White Zombie, the 1932 movie starring Bela Lugosi,” you will get a much more useful set of information than asking a bulk generic question like “Give me a list of movies including their cast and crew”.
Similarly, tell the ChatGPT engine to write you an expert review, like a film critic, and the results improve. Ask it to expand on synopses can also help. However, unless you can’t face doing it yourself, I don’t see much time advantage in any of it. Except of course, you can dump the text into your manuscript to get a feel for length, look on the page and so forth.
Step 5: Take the easy wins and sacrifice perfect for good
One huge bonus was using the book content to automatically create inventory like Amazon listings, ad copy, short blog summaries about the book, etc. It takes a bit of doing with ChatGPT because you need to specify parameters and upload docs at times, however, there are many handy add-ons and tools to make this easier.
I am receiving no incentives for any of the tools I have mentioned by the way, it’s purely my opinion based on trying out everything. For example, I used a toolset called WriteSonic that allows you to upload files and URLs for digestion in the ChatGPT engine. This was my preference over ChatGPT and other AI-connected wordprocessors etc. I am not endorsing it, and if there’s a better one let me know in the comments!
Some editing is required, but I can see a huge workflow boost with these kinds of tools for marketing materials, and of course, if your book is a fiction project this would be a very useful set of tools to create synopses of different lengths, covering letters, query letters and so on. This is important to note — if you can train the bot with the right data, you can get a decent synopsis out of it.
So I wrote my book jacket copy, then fed it into Writesonic‘s pre-formatted Amazon Listing tool, added keywords, got it to search for adjacent keywords and then let it write it for me. Was it great? Well, it’s okay — it’s an Amazon listing, so there’s no great science to it. I tweaked it, saving me an hour or two of soul-searching, which is still a production win.
Conclusion 1: It’s economics, innit…
So I produced a non-fiction guidebook in about five weeks. It should have taken more than six weeks. That’s an efficiency saving. My next one, a similar book, I estimate will shave another week off that because I have used this workflow from the outset. That’s a pretty chunky 33% saving. Cutting your time by a third on a commercial project is a real win.
I’ve also started applying the same techniques to articles and other forms of freelance writing, including using the tools to script out podcast interviews, plan content for shows and talks, organise presentations and all manner of previously manual tasks. You could hand these jobs to an assistant, junior team member, intern, PA, or any colleague on your team — and they, like you, would do a better job — but why? These tools save time, they also allow you to pay humans for things only humans can do — like thinking creatively and problem-solving - rather than commoditise human labour down to Googling, filing, sorting and reading.
Conclusion 2: AI can’t write
It’s clever software. It’s like a scientific calculator for language as opposed to mathematics. It’s like having a LOG, SIN, COS and TAN button for conceptual stuff like CVs, rhyming or Amazon listings. It enables you to query and process data intuitively and naturally with your words. In the same way, generative image software enables you to access complex image processing and 3D rendering without spending years learning how to make that stuff in Adobe Photoshop or Blender. It’s a new kind of interface to help you process and problem-solve faster and smarter than ever before.
It can’t write. It can’t see. It can’t hear. It doesn’t feel or imagine. It doesn’t do the math for you any more than your Casio FX 82a calculator did back in the 1980s. It doesn’t make decisions — although it can be creative to a point based on the model it’s been trained on. It can write a bland, vanilla SEO keyword blog post, it can’t communicate what’s in your soul or on your mind.
In my history of zombies essay (in my book) I discuss the roots of the zombie in culture, relating it back to its origins in the cultures of the enslaved peoples of Haiti, and the subsequent uprising and revolution the cultural event that spawned the fear of the zombie — the faceless horde — rising against civilisation. It’s a story about white colonialism and the brutality of slavery, and the fear of enslaved people reclaiming their freedom from oppression. This theory builds on the fact nzambi in many of the dialects of the Haitian African slaves meant ‘deity’ or ‘god’, zumbi meant ‘fetish’ or ‘idol’, and Zumbi was also the name of a slave leader who led a rebellion in Brazil — which was misreported in the British press as Zombi. If you were an African slave, zombies were a good thing. They were the end of your way of life, if you were a colonial power.
ChatGPT didn’t make that connection. It thought Zombies were:
Zombies are fictional creatures that have their origins in various mythologies, folklore, and popular culture. The concept of zombies has been present in different cultures and has evolved over time.
It missed any meaningful historical context. Zombies come from the history of slavery. That is unarguable. To ignore it is to whitewash one of the greatest crimes against humanity in the history of the world. It also said this:
Similar to Haitian voodoo, the concept of zombies can be traced back to West African Vodun traditions. Some African tribes believe in various forms of animated corpses or spirits that sorcerers can summon and control.
This sweeping generalisation — not to mention ‘voodoo’ is considered an offensive term in Haiti, and should be Vodou — also ignores the obvious fact that the West African Vodun traditions were exported to Haiti by the slave trade. Also, the term ‘some African tribes’ is odd — it refers to the former Kingdom of Kongo, a vast multi-state territory dating back to the early 1300s that now contains Angola, the Republic of Congo and the Democratic Republic of Congo. Some tribes? Over 500,000 people from dozens of tribes. Plus the original meaning of Vodou (deity/god) comes from the Kingdom of Dahomey in East Africa originally and was transformed by cultural relations with Western tribes and the impact of the slave trade.
One of the most influential works in shaping the modern zombie archetype is George A. Romero’s 1968 film “Night of the Living Dead.” Romero’s zombies were reanimated corpses that craved human flesh and had no consciousness.
Romero never called his creations zombies, he called them ghouls. They also developed consciousness in Romero’s later films — in fact, more of his zombie films have conscious zombies than not (3 out of 5).
But how could the AI know? It’s never seen a movie. It’s view of the world is skewed by billions of words of generic truisms and generalisations. So why would we expect it to be capable of writing a book about them? Why do we expect it to understand history? It has no concept of time.
