AI as a tool
How I'm using NotebookLM in 'Grim Khonsu'
Last time, I mentioned I’d been using Google’s NotebookLM to help with my Grim Khonsu serial. Today I’d like to go into a few more details.
NotebookLM is an AI tool designed to analyse text files. These files can be arranged in various notebooks, and each notebook can contain any number of files (okay, there’s probably an upper limit, but I’m not aware of it). The AI ‘reads’ these files, then provides various tools for presenting the information they contain (in a ‘studio’ section). It also has a chat window where users can ask questions about the files. The ‘chat’ part is similar to ChatGPT, but seems to have a tighter focus (while it clearly pulls from data outside those uploaded files, the text within them is its primary source).
I’m uploading chapters of Grim Khonsu as I go, but before I publish them. I then have NotebookLM analyse the story so far. It gives me a full character list, with details (and, if I prompt it, will provide sections of the text it pulled these details from). It does the same for settings.
This is very useful, not only for keeping details of major characters and settings consistent, but also for keeping track of minor ones. Often, as I write, I’ll need a name for an incidental character, and I’ll come up with one on the fly. If I don’t make a note of this I find it easy to forget what name I’ve used. Rather than go back through the text, I can have Notebook LM do the grunt work. Then, when I need a character later in the story (or in a future story) I can go through this list of minor characters and maybe use one I’ve already mentioned.
It’s only a little thing, but it helps make the world of Grim Khonsu more contained and cohesive.
In many ways, I’m using NotebookLM to create an on-going story bible for Grim Khonsu. I normally do something similar in my novels, but using AI means I’m less likely to skip over the smaller details. And, as NotebookLM can tell me where it pulled information, I can always check if I believe it’s misinterpreted something.
Because AI isn’t perfect. From my understanding, most of these Large Language Models are designed to give the user what the AI ‘thinks’ the user wants — not necessarily what is ‘true’. As with any tool, AI has its limitations.
But let’s continue. There’s another part of NotebookLM that creates audio and video reports. I don’t reckon much to the videos (they’re little more than animated slideshows, and the graphics aren’t exactly exciting), but the audio reports are fun to play with.
It presents these as ‘deep dive’ discussions between two AI voices. They’ve recently added the ability to refine what these ‘deep dives’ analyse, so rather than having the audio dealing with the whole story it’s possible to have the ‘discussion’ focus on a particular element, or a particular part of the story.
As an example, I prompted NotebookLM to produce a shorter ‘deep dive’ that focused only on the first three chapters of Grim Khonsu, presenting the material in a way that would leave listeners wanting to know more. This is what it came up with:
I’m also using NotebookLM to act as a kind of editor. As I upload fresh chapters I ask it to look for any contradictions, as well as having it comment on how the new material might change a reader’s perception of the overall story (or of a particular character). I’ve prompted it to focus on the way I present the story, asking how it conforms to both sci-fi and detective noir tropes. Because I’m conscious that these types of AI want to ‘please’ the user, I’m careful to always ask for examples where I fall down — points for improvement, examples of poor writing, things that could be confusing for the reader.
And, because this is a tool, I don’t trust it implicitly. There might be places where I want to confuse the reader — Grim is solving a mystery, and I don’t want to make everything too clear for either my main character or the reader. If there are apparent contradictions, they might be intentional — a character whose actions seem to counter something they’ve said might indicate that they’re lying.
I’ve also prompted NotebookLM to analyse any loops I’ve opened, and as I add chapters I’ll have it check which ones I’m closing. As this story is the first of a series, some of the open loops (especially those concerning Grim’s back-story) won’t be closed until later stories.
NotebookLM seems pretty good at picking up on subtleties and subtext, though. At one point, when my detective Grim is meeting his potential client, I have this exchange after Grim offers Aveline Peron a drink:
“And I’m sure you’ll have a shot of something yourself. I’m almost surprised to see no evidence of narcotics on display.”
“I keep that in the back, with the sex dungeon. Got a young guy on the rack at the moment, if you’re interested. Been training him up with suppression theory, got him to the point where he lasts an aeon now…”
Nowhere did the AI mention that Grim actually did have someone in his ‘dungeon’. It seemed to ‘understand’ that this was a comment intended to wind Aveline up, not necessarily a statement of fact. In another section, where I have Grim commenting off-handedly about blood on his shirt, the AI picked up on his attitude — that this wasn’t anything he felt to be important, and that it wasn’t an issue.
So, I’m using NotebookLM as a way of creating a story bible for Grim Khonsu, an aid to editing an on-going story to ensure continuity, and as a fun toy to produce audio ‘discussions’. The responses it gives me will influence what I write and edit next, but I’m not using AI to do the actual writing. And while NotebookLM might spark ideas, I may well reject those ideas after a moment of thought.
As I said at the top, AI is a tool. It’s powerful, and I’m conscious of using it in ways I consider ethical and useful. But it’s a tool — like my laptop and phone, like spell-checkers and grammar-checkers — that I can use to create a story that is more coherent, more focused and more enjoyable. It’s a tool I can use to create a better story.