ChatGPT, science fiction novels, and the limits of machine learning

ChatGPT, the OpenAI ChatBot launched in November last year, is everywhere at the moment. It's been used to write song lyrics, debate the Trolley Problem, and even get a passing grade on English essays.

One wag on Twitter even claimed:

Well, I'm a smart person (whom, admittedly, he doesn't know) and don't have a ChatGPT tab open. But, then again, I don't work for a research company named  LITERALLY after an evil scrying ball in Lord of the Rings.

Don't get me wrong, ChatGPT can do some stuff very well. It's great at rewording existing text, such as award nominations, and writing pro-forma (sometimes evil) letters. It's also good at generating nice-looking text from widely-available internet resources (although it often gets facts wrong).

The basic problem, and the reason it won't put most knowledge workers out of a job is the crucial word KNOWLEDGE.

Writing is not flower arranging for words.

Software like ChatGPT (and the image AI equivalent) works by recognising recurring patterns in huge bodies of written (or visual) material on the internet. For ChatGPT, this means learning which words typically follow each other in a paragraph, or series of paragraphs. Then, when it needs to communicate similar words later, it can arrange them in the best order.

Some of us who are "true" picture thinkers (eventually) learned how to arrange our thoughts in text on paper in the exact same way, i.e. by systematically analysing the common patterns underlying sentences, paragraphs, news articles and features, and even entire novels. It's why I'm a good, fast writer now, but there's no magic to my art - just brute 'machine' learning.

What this means is that ChatGPT, like me, has great knowledge of how to structure text. The writing is well presented and, if you're not confident at arranging your thoughts in words on paper, this can be very helpful. However, if you're already confident using words to communicate ideas (and have ideas to communicate) ChatGPT doesn't really add much.

ChatGPT can't synthesise pages of research notes to write an academic paper. It can't (currently) turn an interview into an article in the house style of a specific publication. It can't write an explainer on a highly-technical topic that relies on interviews with experts. It can't attend a conference to look for topical stories, and it can't make a personal story interesting unless a human has already shared it.

Most of my time as a technology journalist is spent working out WHAT to write, not HOW to write it.

There are, sadly, jobs at risk from ChatGPT. These are mostly word-smithing roles, which often involve rewriting existing copy. I recently spoke to a copywriter friend who is, legitimately, worried about ChatGPT. His job mainly involves reworking brand documents at a shorter length or in a different style, based on client instructions in a short briefing note. He believes, justifiably, that some clients will find it easier (and cheaper) to brief ChatGPT instead.

My conversation with him led to a great example of the apparent potential - and limits - of current text AIs. He'd asked ChatGPT to write a short story about a copywriter worried his job would be taken by a robot. Despite the basic prose, the story featured a robot called 'Sparky' who wasn't specified in the task description. There was a funny joke at the beginning, and the story had a clear middle and an end.

Going by my experience with the MidJourney image AI, I was certain his good results were down to ChatGPT being given a loose task description, which gave it the freedom to recycle tropes off the internet. I figured it would fall apart as soon as it was asked to do something specific and original - the same problem that MidJourney and Artbreeder had with Syer Roth's complexion and Tirim Kraygor's nose.

I bet him that ChatGPT would struggle to write the opening scene of my current work-in-progress novel, Ace of the Usfan. Based on my experience with AI-generated art, I expected a text AI would create what was effectively 'fan writing', i.e. a mash-up of every giant flying airship/Star Destroyer/aircraft invasion trope available on the web.

With this in mind, I created an account and asked ChatGPT to:

Write the opening scene of a dieselpunk novel where a 130ft-long flying aircraft carrier invades a small island called Belorn.

I don't have the screenshot, but ChatGPT came up with:

The sky was filled with a mix of orange and purple hues as the sun set over the horizon. The peaceful evening was shattered by a loud rumbling sound, like a deafening roar of thunder.

The people of Belorn looked up in shock and awe as a large 130-ft long flying aircraft carrier descended from the sky. It was a sight unlike anything they had ever seen before. The aircraft carrier was equipped with an impressive array of weapons, from cannons to machine guns.

The people of Belorn were terrified as the aircraft carrier hovered above the island. The loud engines of the aircraft carrier seemed to shake the ground beneath them.

Suddenly, the aircraft carrier released a wave of fighter jets, which began to circle the island. The people of Belorn could only watch in fear as the jets began to fire missiles at the island.

The people of Belorn had no choice but to hunker down and wait as the aircraft carrier continued its attack. The sound of explosions filled the air as the island was engulfed in chaos.

It seemed as if Belorn was doomed.

When my husband saw this (over my shoulder), he got interested in ChatGPT and decided to see if it could write a novel outline. He asked it:

Write another scene - give me five possibilities

And it came back with:

1. The people of Belorn fight back against the aircraft carrier by launching their own fighter jets and shooting down the invaders.

2. The aircraft carrier lands on the island and a group of soldiers disembark and begin to search for survivors.

3. The people of Belorn are able to repel the aircraft carrier with an unexpected weapon, such as a hidden cannon or a secret weapon buried beneath the island.

4. The people of Belorn take refuge in an underground bunker and wait until the aircraft carrier leaves.

5. A mysterious figure appears and uses magic to repel the aircraft carrier, allowing the people of Belorn to escape.

We mutually decided '2' was probably the closest what happens in my novel. My husband then asked for:

Number 2 please

And it replied:

The aircraft carrier slowly descended towards the island, its engines reverberating through the air. The people of Belorn watched in terror as the aircraft carrier landed on the island, its heavy weight shaking the ground beneath them.

A group of soldiers disembarked from the aircraft carrier and began to search the island for survivors. The soldiers were heavily armed and wore menacing masks that obscured their faces.

The people of Belorn were terrified as the soldiers began to search their homes and interrogate the survivors. The soldiers seemed to have no mercy, as they destroyed homes and threatened to harm anyone who resisted.

As the soldiers continued their search, the people of Belorn huddled together in fear and prayed for a way out of this nightmare.

Suddenly, a loud roar filled the air as a mysterious figure appeared in the sky, riding atop a giant dragon. The mysterious figure descended from the sky and confronted the soldiers.

The figure seemed to be wielding some kind of magical power, as the soldiers suddenly stopped their search and began to retreat.

The figure smiled at the people of Belorn and waved them goodbye as the dragon flew away, taking the soldiers and the aircraft carrier with it.

And I said to my husband, quote:

"Where the F&!K did that dragon come from?"

I'm sure some people will read the ChatGPT text above and think "OMG! Novelists are DOOMED!" and "ChatGPT will be writing War and Peace within a decade." I'm here to explain why it won't, and it's nothing to do with the romantic human notion of powers of creativity.

Below is the (rough early draft) opening of my dieselpunk novel.

The Fall of Belorn (partial) by Vivienne Raper

It's (hopefully) more interesting than the ChatGPT version and that's not because I write better prose (I don't) or I'm unusually creative (I'm not). It's simply because the density of ideas is far higher in my text.

Writing is merely the expression of ideas in a text form

Buried within my scenes are pages of notes on the operation of World War II-era radar and the armed forces of Tonga. I measured the distance halfway across the Pacific and the travel time of a rotorcraft cruising at the speed of a Mil V-12. Lipari in the Aeolian Islands was an inspiration for Belorn, which has a similar population to Cambridge in the UK. I've stayed in both places and, to refresh my memory, I viewed pictures of Lipari's castle fortifications.

Folks, I even skim-read the history of the retractable pen. And that's just for the first 1,000 words...

To get something from ChatGPT that resembles my story would require ALL of my notes AND my detailed scene planning. Transferring that information to the chatbot, however, would be a remarkably complex process - and probably doomed to failure. I could give the AI my long list of hyperlinks, but I'd also have to provide explanatory notes and a detailed text prompt to help organise them. And, to be honest, by the time I'd done all that research, compiled the notes, written the text prompt, run ChatGPT, and edited the output text, it's be quicker to just write the story myself.

I'm sure ChatGPT (or another language model) will soon be writing disposable erotica to capitalise on popular trends. One day, it will doubtless write Young Adult novels featuring an ordinary teenage girl in a love triangle. It will author Regency romances between a handsome duke in a wet shirt and the servant girl who clumsily trips into his arms, and medieval fantasy novels where a farmboy discovers he is destined to save the world.

However, the world's more interesting writers generally don't rely on hackneyed plots and common fictional motifs. There's a reason that J.R.R. Tolkien, who served in World War I, could write Lord of the Rings. Even good authors of apparently 'pulpy' novels enrich fictional worlds with personal interests and details of their offline lives. ChatGPT, for example, could never own a gun shop like Larry Correia, author of the gunplay-filled Monster Hunter books, and isn't interested in writing a book defending the US Second Amendment.

It's personal beliefs and the fine details of specific situations that bring great novels to life, and ChatGPT simply can't replicate that. Not because it hasn't read anything, but - simply - because it's not alive. It can't get divorced, report on assassination plots, or even pop to the shops for a croissant - all experiences that inspire novels and create the relatable details in good fiction writing.

For me, there's no question about whether current AIs can 'replace novelists'. ChatGPT and MidJourney have neither lived experience nor the motivation to make - and share - their own experiences. They can, however, make trope-driven content written by humans worthless by rapidly generating similar, better-written, but equally derivative material. And that, doubtless, will have an impact on the authors of relatively trop-ey work.

The question is, thus, not "will AI replace novelists and artists?" but, rather, "humanity is now facing the greatest fan-writers and -artists the world has ever known. Can we, as humans, up our game to compete?"