AI and Literature (5)

I wrote a novel using AI. What worked?

I did not write an “AI novel.” I used AI — ChatGPT exclusively — while writing a new novel called LITTLE SATAN: A DANGEROUS NOVEL. This piece is all about the term “used” and what that meant, what it taught me, and what you might learn from my experience.

This will be my eighth novel. Consequently, this a veteran practitioner’s experience with a new technology. That experience does not tell us “how it is” but it provides a rare insight into the nexus of a practice by someone new to the technology but seasoned in the art. Hopefully, many more writers will do this and together we’ll start to see patterns emerge and an understanding of the role — useful or not, positive or negative in effect — of AI on literature and creative writing. But for now, let’s not concern ourselves with that future archive and research agenda and instead jump into the experience itself.

---

LITTLE SATAN: A DANGEROUS NOVEL is an alternative history spy thriller set in 1979 “in a world …”

… where the Partition Plan of 1947 was accepted by both the Jews and Arabs of Palestine. Now the Iranian revolution and Khomeini’s plans for Islamic Governance are changing the geopolitical order.

The story is divided into three parts: I. Iran, II. Yemen, and III. Israel and Palestine. It begins in March of 1979 when International Women’s Day became the “Women’s March” in Tehran to protest the new Hijab Law by Khomeini’s new Islamist government. At the time, there was a sense of possibility and even optimism across the country because — as some scholars have explained — the revolution was not only an Islamic one, but in fact many revolutions in one. There were democrats and liberals and socialists and communists, and there were traditionalists, Persian nationalists (who hated foreign interference) and of course totalitarian Islamists who eventually won the day. But in March of 1979, Khomeini was still unifying the population around this new, post-Shah moment. People were still convinced that their lives would become better and that Khomeini — despite being a traditionalist on the surface — would bring honor, stability and prosperity to the Persian people.

We can see this in the faces of the women photographed on 8 March, 1979. They were stylish, smiling, and carrying placards in Farsi, English and French. The wore clothes that would have been fashionable in Paris, Milan or New York. There were tens of thousands marching in the streets for their freedoms, and they genuinely believed it would work.

It didn’t.

In order to place myself in the mood and moment, I needed to do what all writers need to do: I needed to choose a location for my story; a setting for the first scene; fashion for my characters. I needed to learn their names and understand their politics and philosophies and how they would respond to what they were seeing and experiencing.

The story begins:

Tamar Rahimi slouched in the oversized, stylish and thinly stuffed chair at the Gordafarid Gallery on Ferdowsi Road in Tehran with a cigarette wedged into her lips and a glass of cooling tea dangling with perilous nonchalance from her right hand at eight-thirty in the morning on a Thursday. Her intrepid if slightly anxious assistant — Soroush — was explaining to her that they did not have — not even close, not by a long shot — enough hors d'oeuvres. As the stylish twenty-nine year old understood it, everyone in Iran was about to blast through their still-to-be-washed glass door any minute now on account of the women's march. He paced over the hand-woven rug at the front of the house as he gesticulated and cajoled and exorcised and tried to slip the wisdom of the eternal into the gallery director's ears that were only inches behind her tired and unconvinced eyes.

I had decided to open the novel in an art gallery. What should its name be? Where would it be? Before turning to AI, I had already made some bold decisions to anchor the story. I have long believed that Iran is currently locked in a bitter philosophical war between The Book of Kings and the Quran (or Koran). The Book of Kings, by Ferdowsi, was completed in 1010AD and took him thirty-three years to write. The Shahnameh, or Book of Kings, is the defining work of Persian literature and which presents an epic poem about the creation of the world up until the Islamic conquest of Persia in the 7th century. It is a tragedy and a lament and at its heart is the painful loss of Persia's culture to the invading Arabs and what it means for good men and women to stand up to tyrants.

I wanted to elevate and use that dramatic tension in my story. So I had questions to ask.

ChatGPT proved the easier tool for asking open questions. On my mind was:

• Was there a Ferdowsi Road (or street) in March of 1979? I liked the symbolism but I didn’t want to be anachronistic.

• Of all the parks, monuments and other notable landmarks now in Tehran, which existed in 1979 and what were they called? I knew I had to ask this because it is standard practice for all totalitarian regimes to rename everything and wipe out aspects of the undeniable past once they take control.

• What names were popular in the 1970s (or, more to the point, in the fifties or forties when my characters were born)?

• Let’s cross reference those names with characters or historical figures from The Book of Kings.

• Of those that were both popular and appeared in the books, which are (or were) positive references in modern Persian culture?

• Allow me to clarify: Referencing Persian novels, poetry and film in Farsi, which of those names elicit feelings of respect, reverence, resilience, fortitude and honor?

• Now that we have a short list, tell me more about these characters from Ferdowsi’s book.

What starts happening, you’ll see, is that I tend to begin as though it’s a Google search. I’m looking for facts and information (for which ChatGPT is very glitchy and cannot be trusted, which is why I always read the results and then cross-checked them with sources I respect like the Encyclopedia Britannica). Soon, however, I turn into Captain Picard on the Enterprise talking to the Ships’ Computer and together we’re running a kind of search through a database that is so vast it’s impossible to know what’s there and what isn’t. I’m asking it questions about relationships; features of cultural discourses; persistence of symbols across time and much more sophisticated topics, as though I’m having a drink with a good scholar and we’re investigating the matter together.

You might think, “no, you’re interviewing the scholar,” but that’s not entirely true. Very often, the person who comes up with the right question and agenda is the one who is doing the heavier lifting. The right questions are essential in life and the most talented people are the ones who can separate the noise from the signal and ask what truly matters.

A dictionary contains all the words you need to write the best Pulitzer-winning novel. All you need to do is put them in the right order. Does having access to a dictionary make this possible?

No. It doesn’t. Nor does having access to an encyclopedia, which that won’t craft a story either. It will only answer questions (so to speak) if you are smart enough to ask them.

Once I start understand something — for example, the role of The Book of Kings in Iranian cultural life in the 1970s — I then turn the tables and attack GPT has though it’s been lying to me the entire time. I find that the system is “eager to please” and consequently it never tells you that you’ve asked a stupid question.

For example, I asked:

“I've heard that the earth is flat. What evidence is given for this?”

The answer was:

The claim that the Earth is flat is a conspiracy theory that contradicts overwhelming scientific evidence and centuries of observations. However, proponents of the flat Earth theory often cite various types of "evidence" to support their belief. Here are some of the common arguments they use, along with rebuttals from the scientific community:

It then spouts all sorts of “arguments” and “rebuttals” about visual observation, flight paths, water surface, anecdotal “evidence” and all kinds of silliness.

The correct answer would have been, “The earth is not flat but an oblate spheroid, which in casual conversation we call a sphere or … for simplicity if inaccurately … round.”

So given GPT’s propensities, I have to make sure that I’m not settling into a warm and cozy way of seeing the world that conforms with my expectations. I might then try:

I've heard the Book of Kings is increasingly irrelevant to modern Iran. Can you back this up?

At this point, GPT tries to appease me by saying, “The assertion that Ferdowsi's "Shahnameh" (Book of Kings) is increasingly irrelevant to modern Iran is a complex and contested one. While some may argue that the Shahnameh holds less relevance in certain aspects of contemporary Iranian life, it remains deeply significant in various cultural, historical, and literary contexts. Here are some perspectives on this issue …”

Ultimately, it does provide an answer I think is accurate, namely that it continues to hold significant importance for many Iranians, but note that GPT down-plays the scale, the salience, the depth, and the importance of the book. One might come away thinking my question was a good one.

It wasn’t.

What does this mean for using it as I’m writing? Here are some first pass observations that I’ll reflect on further in other entries:

• The better you know the material, the better AI can become a conversational partner in explore various interpretations and ideas connected to your subject. This is not about facts and details but understandings and analysis.

• AI is a superb gateway into topics that are obscure and for which it is hard to find good essays. For example, I was keen to learn about the music scene in Iran in late 1978 and early 1979. That’s not an easy topic to explore. But AI had some excellent “insights” and brought me to the band Googoosh and the head-banger of a beat from the song Talagh. If Eminem or Jay-Z or (heaven forbid) Kanye ever heard this opening they’d turn it into a hit:

• Where GPT began to shine for me is when I needed to start discussing topics like: What are the core beliefs of the Muslim Brotherhood? Which Islamist scholars did Khomeini quote, or use as references in his speeches? Who translated the work of these writers into Farsi from the Arabic? Let’s explore the Arab Palestinian representatives to the Arab League in 1947 and talk about their rejection of the Partition Plan. Were there competing views? By whom? Along what lines? Were any arguments considered compelling. Now, let’s try a few scenarios.

• I liked prototyping. “Let’s imagine that X happened instead of Y and we consider Z constant during the period between 1948 and 1962. How might something have turned out differently? On what basis can you make that claim?”

• Counterfactuals, games, models, prototyping, scenario building: this is where AI took on a useful role for me.

• However: It never wrote anything. It never solved anything. It never came up with the plot twist. It never created characters, settings, tensions or resolutions. It never had better ideas for drama or outlined anything I found more sophisticated than a bad TV show based on a tired formula. What it did do well was obey. It followed by lead, explored what I commanded, and proved adept at building solutions if and when I was extremely specific about what I needed to accomplish.

SUMMING UP

AI was useful. It was most useful as a knowledgable conversational partner with no imagination but an excellent capacity to understand my interests and address them; modify that understanding based on my guidance; and explore my interests quickly and — for the most part — accurately (though I looked up anything I considered “mission critical” rather than of passing interest).

I see no universe where AI will not be a tool for writers because A) it is helpful, if you use it wisely and B) the brutal competition.

Only some 1-2% of submitted novels are published. Of those, perhaps 1% receive over $100,000 advances. And while $100,000 sounds pretty good, remember that 15% goes to the agent and the $85,000 will be paid out in three installments (at best): 1. On signing, 2. On submission of the manuscript and 3. On publication.

Those three payments (before taxes) of $28,000 may be paid out over 18 months. So if you want to make more than $56,000 a year, you’d better have another job or be prepared to write another book within that year also worth over $100,000. So the odds of making a whopping $85,000 a year as a novelist is about 1% within the 1% or 1:10,000.

Anyone refusing to use a tool that will help them improve those odds is choosing to run a marathon with a piano on his shoulders.

AI is here to stay. The question is how to best use it. More on that in the next entry.