In the likeness of a human mind

As I reach Chapter 10 by the seat of my pants, I am at the limit of at least one kind of improv. This strikes me as a good spot to stop and check while also giving some consideration to use of AI/LLMs in the process (and by extension the role, ethics, and limit of “AI assisted writing”).

While I have a specific number of chapters in mind for the project, this based on a ‘fixed number’ first draft, the three stages in life I am building on top of it are new to this iteration. I have been tracking these stages in a chapter sheet, colour coding each story by location and age, the idea being to tell the story in part through movement from one place and time to another. With ten published, how to position the rest of the lily pads to get across the pond?

This lead me to think more broadly about shifts of perspective in novel writing and how other writers have managed to balance or otherwise organize changes from chapter to chapter, act to act, book to book.

My first experience of this was as a very young child having got through The Hobbit once or twice and now old enough for Mum to read the next story to us as a family, The Lord of the Rings. We are with Frodo and his pals reliably - with the exception of Gandalf in Woden mode - all the way through to Lothlorien. It is only at the fall Boromir beyond the Falls of Rauros that the party is split. For the remainder of the novel, our time with Frodo is necessarily time away from the rest of the party. No matter who we followed, we found ourselves missing our friends, concerned for them at some cliff edge.

GRRM’s riff on LOTR, the Ur-text of high fantasy (an accident), ASOIAF, itself producing low fantasy (another accident), also uses change of perspective to dramatic effect. Each chapter is told from a strict, limited, varying first person, a tactic I believe he borrowed from The Sound and the Fury and As I Lay Dying* (famously, people used to skip the Sansa chapters, a rookie mistake).

There is an extensive fan literature on Martin’s character changes, particularly for the first three novels; I gather Tyrion and Jon get a bit more airtime. And of course Faulker is the subject of a century of forensic literary criticism. The Sound and the Fury considered “highly symmetrical and framed” in its changes in perspective and As I Lay Dying by contrast “kaleidoscopic and journey-driven,” this according to Grok though presumably at least in part a gestalt of academic opinion rather than a quick reading of the novels in expert mode (I did not provide access to the novels, I am surface skating with this).

My problem is simpler, one narrator recollecting moments from three passages in his life and an occasional reference to his present. I have stories to tell and an architecture to build them on but what I have come to need is an order to tell them in. Here geography and times of life line up neatly - though perhaps not entirely, we shall see - so the underlying question is to ask what rhythm I am after and to what effect with a corollary of figuring out what pattern is most likely to achieve it.

This brings me to an AI component of the process. Here I have considerable latitude; as Flowers in Asphalt is being written and published live, it would not be of much interest to most publishers even without its subject being a niche interest to begin with. I can, therefore, do exactly as I please.

Let us start with two simple things. If I am using Google to search, I am effectively using Gemini as that is front and center in returns. While there may be a way to disable this function I would not trust it; if I am working on a project that requires a strict AI process firewall, I use Bing. That said, it is entirely possible Bing’s returns are in some way AI assisted at the back end but, if so, this is black boxed; what is visible to me are straight returns for pages.

For example, just now Bing’s first return for “william faulkner pattern analysis in chapter character perspective. sound and fury. as I lay dying” is a LitCharts study guide for As I Lay Dying.

Who knows? Some of the returns I am looking at might themselves in turn be the product of AI assisted research and writing. But if this were a discussion with a student, I would say, with perhaps an aside on Dead Internet Theory, look, there is a limit to the decontamination protocol that is realistic or even possible with the matter a moving target. Models six months from now will likely be used for things we have as yet to imagine, let alone consider, so at this point I think it is a question of best practice. To some extent, it is a question amenable to the same tools we have used for some time. If I am quoting someone or crediting someone, I cite a source. If I am making a literary reference or allusion, I have to trust my judgement it is sufficiently recognizable to a reader for the egregore to do whatever work I am conjuring it to do.

This is to say, at least to some extent our problems are not new. They have to do with attribution and to a large extent even this is a culturally mediated challenge. If I were to cite Plato to a French academic, this would be an insult. If I were to fail to reference the same allusion to a Canadian academic, I might be committing plagiarism. It is irrational to expect the line on AI to have a clear demarcation that will be universally accepted.

Preamble done, you can call this next bit a note on process or a confession, depending.

Further detail, for context: I am using SuperGrok (whatever model is current as of July 4, 2026) (Happy Birthday, America!). I should likely be on Claude as this appears to be the current prestige model and I would guess what anyone studying AI work processes is using to study it with. My choice does not have a methodological basis, it is just baked into X/Twitter and so ready to hand (a Heidegger reference, in case I should say).

Here, I thought, is a potential use case. Other novels and other novelists have thought through the problem of how to balance shifting perspectives. How did they do so? What did they consider? How did it work out? See above re ASOIAF, books 1-3. This as a question I think would not trouble any but the strict purists as my choice, whatever it would be in the structuring, would be based on a research summary. If Wikipedia had ever considered the question (and they may very well may have done precisely so), my Grok return would be comparable in output (quality of said output, a separate issue). What would be the difference?

Thinking this through as I am typing, I remember an old friend who was working as a research assistant on a work of nonfiction for his employer. In the end, he felt the work published was a light reworking of his draft and his credit limited to one of many ‘thank yous’ in the introduction. I have been involved in at least one project along these lines with my work submitted word for word for grant review under the lead researcher’s name. I have done straight up ghost work and have no problem with the practice; it is as lindy as it gets for professional writing. What bothered me, and my friend with his uncredited book, is not understanding up front just how much work we were doing for so limited a return. We were compensated but were not paid ghost rates.

Is a Grok summary the effective equivalent to the work product of a hired research assistant? To what extent does that research summary itself deserve credit? With a human agent, it is a matter of contracts and clarity. With an LLM, is it a matter of disclosure? Again, I believe we are dealing with a moving target.

My next step is where I might have stepped over the line and why I believe I haven’t.

Again, ten chapters on the fly, each told from one of three general times and places. How to most effectively shift from place to place and time to time in a balanced way, with a specific effect in mind (rising action, etc etc), over a set number of chapters?

To this end, I provided my existing chapter outline with a link to the first nine published here. Could Grok, taking on the role of a critic familiar with the way novelists have approached the problem effectively in the past, suggest how to proceed?

The short answer is no. At least, not with my prompt game.

My first return split up the remaining chapters into all Montreal, then all the UK, then all Florida, split into three sections. This is specifically not how I have been hopping around the place and not, I thought, what I had asked for. And this quite aside from the whole point of a non-linear story structure. Grok didn’t get it.

But of course, why would it? Grok does not know how to parse these considerations ; I had not explained them. Once I attempted to do so, throwing notes at it, I found it did not make the slightest difference to its output. Worse yet, as I have Grok’s “personalized” conversation history active, Grok tried to insert suggestions based on my Miami real estate searches. Not at all what I was after and a waste of limited time.

Thinking about this overnight, I decided the problem might be asking an LLM to weigh in on a specific project. Grok did not need to know anything about the story, it just needed to help me decide what order to do things in. In other words, the help I am after is not narrative at all, it is pattern analysis.

Which, as it happens, I have some slight experience with. Once upon a time, for my first unfinished PhD, I took a directed seminar on the subject for a quantitative analysis credit. I cannot say I deserved the A+ but for anyone curious, I shall likely include that paper in the academic volume I am slowly editing alongside the narrative design book. More to follow.

The issue was not strictly about character progression but keeping a reader’s interest while also keeping my line of flight unpredictable. Transmitting a signal while frequency hopping, in other words.

I gave Grok the pattern of my first ten chapters composed of a sequence of three letters, each letter standing for a broad place and time of life. I then directed Grok to take the total number (of chapters) in the sequence with the caveat I wanted to end the series with a specific letter (a specific time and place for the close). One final (I think) instruction was to create two alternative complete patterns that would echo, extend, and build on the first ten. Letters only. Just the pattern.

This Grok managed handily, presenting two patterns and - unprompted - providing a “feel” for each. What was particularly fun was Grok’s guess as to why I had asked. Was I working with a genetic code, a musical sequence, with something linguistic or purely mathematical?

With these, we were cooking with Claude Levi-Strauss (The Raw and the Cooked) and I was on familiar territory. Both patterns Grok provided still looked mechanistic and predictable to me but - and here is the process part - looking at them broke whatever block I had and this specific problem. I produced a third version, my version, with reference to what I have in mind for our protagonist. Consequently, I now have a working guide for my my next weeks as they unfold. I did ask Grok what it thought of my decision. It said the result felt conversational and more like jazz than what it had offered. Reassuring to me, at least.

There isn’t a chance I am going to stick with it. These matters have a life of their own. They get away from us, sorcerer’s apprentice style. And this is after all the precise magick I am calling up for the invocation. Possession by the process, by a daemon in Roman-speak, is as lindy as the human and near human world. We could be sitting around an ancient fire, the stars themselves thirty thousand years displaced from where they now sit, and be up to much the same business.

Even so, it is a pattern, I used the steps as outlined as inspiration to produce it. For some that would take the future structure of this book over the line from the strictly human. I do not believe so personally but I am not married to my conviction. I am more curious how far I will stray from what I have come up with as the project unfolds.

* I do not know that Catelyn Tully - we know her as Catelyn Stark - was a riff, a nominative allusion to be technical, on Cora Tull but it fits his MO (respect). Cora is pious and judgmental by contract with her practical and grounded husband FWIW. I did once pose the question at Not a Blog but Mr. Martin either did not see or or did not decide he would like to answer. I remain curious!