Starving the Gods

On AI

This book was written with AI assistance

I would rather set that out here, in my own words, than leave it to be discovered and described by someone else.

That sentence is accurate and almost useless. It can describe a grammar checker, and it can describe asking a machine for a novel and publishing whatever comes back. What I did was neither, so I would rather set it out myself than leave people to invent a version.

Three tools, three jobs. I developed the book with Gemini. I drafted the chapters with Claude. I made the pictures with ChatGPT. They are not interchangeable and I did not treat them as though they were.

Developing it

Gemini was the working partner for everything upstream of the prose. The closest comparison is a writers’ room, a developmental editor, a researcher and a critic, all available whenever I wanted one. I could bring a plot problem to it at midnight, argue for twenty iterations, reject nineteen, take one idea out of the twentieth, and come back the next morning having decided the whole thing was wrong. That happened a lot.

I used it hardest on worldbuilding, mostly by trying to break my own rules. If the Lattice works this way, what else has to be true? If a civilization loses instantaneous communication but keeps ancient transport infrastructure, which institutions survive that and which quietly die? If a technology grants an advantage at a physical cost, what does three hundred years of politics build around the cost? I could propose a rule and then spend an hour attacking it from angles I would not have found on my own.

Plot worked the same way. I would give it a scene, an outcome, or the gap between the two, and it would suggest routes across. Sometimes one was useful. More often the useful part was finding out why five apparently sensible routes were wrong.

That turned out to be the real function. It is very cheap to be wrong with. I could explore a bad idea for an hour instead of spending three days writing it and then discovering the same thing.

The same conversation handled pacing, exposition, scene structure, terminology, continuity and repetition. Structural problems are much easier to see when something is arguing with you about them at the point you are still willing to change the answer.

Writing it

The chapters themselves I drafted with Claude. That is a different kind of work and it needed a different tool: not what should happen in this scene, but what this paragraph sounds like, how this character talks when they are lying, where a sentence should stop.

Some passages I wrote and had it criticised or tightened. Some started as its draft language and I rewrote them, combined them, or threw them out. In places it went back and forth often enough that trying to assign individual sentences to one side would be meaningless.

Where it is bad

All of them share the same failure modes, and they are worth knowing about if you are going to use one. They like plausible answers, and plausible is dangerous in fiction, because plausible can sound excellent while being completely generic. They over-explain. They find a pattern and then repeat it. They will make three characters sound like the same intelligent person wearing different clothes. They will confidently repair a contradiction by creating a larger one somewhere else. After enough iterations they forget what made the first version interesting.

A surprisingly large part of using them well was telling them no.

The pictures

The visual work made that obvious faster than the prose did. ChatGPT helped me study contemporary science fiction covers and work out what I wanted this book to signal from across a room, and its image tools produced the concept art, the cover iterations, the interior illustrations and some of the artwork on this site. Most of that process was rejection. One cover direction drifted into fantasy. Another turned into a faction logo. A chapter insignia generator became obsessed with tree rings and refused, repeatedly, to draw the mechanical symbol I was asking for.

The cover that survived, three small figures bracing a monolithic door against a violet breach, arrived only after I stopped trying to decorate the book’s mythology and asked instead what the story looks like from fifty metres away. The same turn produced the interior illustration system: not pictures of scenes, but fragments of ledgers, machinery, records and structures that could plausibly have survived inside this civilization.

Having three of them open was useful in its own right. I could take a critique from one and put it in front of another, let them disagree, and decide which reasoning I actually agreed with. Less consulting an oracle, more watching very fast consultants argue in front of me. There were ordinary sources too: current and classic covers when deciding how the book should present itself, public publishing information for Canadian ISBNs and print distribution, comparable novels to work out which shelf this belongs on. None of that wrote the story. It told me what I was walking into.

The question I cannot answer with a number

People want a percentage. How much of this is yours. I do not have one. I never tracked it, and I am not going to reverse-engineer a figure now to make myself sound better.

What I tracked instead was whether I believed it, and whether I wanted to keep reading. I went through my own drafts looking for the exact point where I stopped feeling anything: for the characters, for the situation, for the turn I knew was coming. When that feeling went, I did not tidy the paragraph. I took the story somewhere I had not planned, because predictable was the actual failure and flat prose was only the symptom.

That is not a measurable standard and I am aware of it. It is the one I used, and it is why the book has the shape it has.

What changed, and what did not

What this changed was not that I had ideas. It changed how many times I could interrogate one. There used to be a practical ceiling on exploring every branch. Friends get tired. Editors cost money. There are only so many versions of the same paragraph a reasonable person will discuss with you at two in the morning. A machine has no such objection, which does not make its answers good. It makes the iteration cheap.

The Great Betrayal, the Lattice, the three dynasties, the Surge, the people caught underneath all of it, the idea of a debt inherited by people who did not incur it, and eventually the question at the centre of the book all went through versions no reader will ever see. Some got better. Some got simpler. Some disappeared entirely. The machines were very useful in that. None of them was the reason I cared which version survived.

The part I am not waving away

Everything above answers the question people usually mean, which is whether this was lazy. There is a second objection and it is more serious, and I would rather meet it than let this page quietly skip it.

These systems were trained on an enormous quantity of writing, much of it by people who were not asked, not told and not paid. Writers have lost work to this. Both of those things are true, and neither of them is answered by pointing out that my book turned out fine.

I do not have a clean resolution. I use the tooling and I think the people angry about how it was built have a real case. Anyone who tells you that tension resolves neatly is selling you something. I would rather sit in it honestly than pretend I have thought my way out.

What I will commit to

One part of this site is different

The Ledger is generated automatically, and should be labelled as such. A script reads real news headlines each day and has a language model rewrite one event three times, once in each faction’s voice. I set the constraints and I read what comes out, but I do not write those items and they are not reporting. Every one links to the real story underneath it. It is there because it amused me and because it demonstrates the book’s argument on material you can go and check.

And then

If you have read this far you are owed the actual thing rather than more discussion of it. Chapter one is here, in full, free, and asks nothing of you. The book stands or falls on the reading.