First Look: Taking Google's NotebookLM for a Spin with Some Spooky Stories

I take Google's new mini-LLM for individual projects for a test drive with some of my short fiction.

This morning I decided to take Google's NotebookLM for a test drive. NotebookLM is one of a new class of tools that takes the popular chatbot concept and trains it on your own documents and notes. It's currently available by invite-only, but you can join the waitlist at the product's home page.

NotebookLM functions primarily as a research and analysis tool, allowing users to uncover connections and properties in disparate content sets. For creative writing, my hope was that it would help me make better sense of the structure of my stories and maybe reveal some insights that my own 85th or 90th manual read-through didn't. I can also see it being invaluable as a research tool for when a writer is first gathering background for a story.

Structurally, the NotebookLM hierarchy begins with a project. A project is basically a collection off documents or sources (currently the tool can just grab anything from Google Drive, but apparently will support other document sources in the future) which are then used to train the model which drives that generative chat functionality. I started out with a single short story, The Hungry Crones, that I've been working on and shopping around to potential publishers. Since I hadn't gotten any bites from said publishers yet, I figured I could use NotebookLM to take an impartial look at it and, with appropriate prompting, reveal its insights to me that I could use to improve it.

The first attempt was a no-go. Like, literally. NotebookLM would not go. I got error messages:

Error messages when trying to load sources for NotebookLM.

This was the case for both an existing project I tried to add a source to and a new project. Keep in mind that NotebookLM is in pre-release state, so this is certainly no big deal. But the fact that it threw a roadblock up immediately was frustrating. I switched to one of my other Google accounts, but it hadn't been green-lighted for the pre-release, so I couldn't continue. I ended up adding another story (Fat Boy) that did successfully load as a source. I'm not sure what NotebookLM has against Crones, but it wasn't to be. Hopefully this will be a smoother (and with less opaque errors) process in the released version.

Fat Boy (working title) is a short horror story around a Peloton-style exercise bike. It isn't finished and I went off on a pointless plot tangent that will eventually be cut, but it at least gave me some material to test Google's tool. Take this analysis with a grain of salt, because this was a super unfair test (an incomplete, raw and unedited story). However, the answer to the prompt "what is this story about" started pretty good:

The story is about a man named Tyler who is grieving the death of his wife, Erin. Erin died in a car accident a year ago, and Tyler has been struggling to come to terms with her death. He is also struggling with his own guilt over the fact that he had not been a good husband to Erin.

The end of the story being unwritten, though, NotebookLM's summary of the end gave me a legit 💀 moment:

The story ends with Tyler riding his exercise bike again, but this time he is not alone. Erin is there with him, and she is cheering him on. He is finally able to let go of his grief and move on with his life.

If you know me or my writing, you know that this is an... unlikely end for one of my bleak-ass stories. But not a totally irrational one for a story in general. So kind of cool, if not directly applicable.

Moving on, I loaded a finished, beta-read and published story called Bamboo Girl. The same prompt gave a pretty good summary of the story. When I asked what the killer's motivation was, NotebookLM gave this:

The killer's motivation is not explicitly stated in the story, but it can be inferred from the text that he is motivated by a desire for power and control. He kills the young woman in order to possess her and make her his own. He then buries her in his bamboo forest, where he can keep her close to him and watch her slowly decay. The bamboo forest is a symbol of the killer's own twisted desires and his need to control others.

This is actually pretty good. This tells me that there are some cues that the model was able to use to generate a response. The "decay" part wasn't quite right, but a damn good guess. The story is more magical-realism horror than horror, but again: good guess!

Given this insight, I can see how an LLM so trained could tease out plot points and lead the writer (me) to beef up those elements I want to be more to the forefront in the story and scale back other elements that I don't want center stage. It could also lead to the surfacing if surprising subtexts that we as writers might not even know are there.

NotebookLM can aid with the nuts-and-bolts of writing as well. For example, in response to the prompt "is the killer's character believable?" we get this answer:

The killer's character is plausible. The story provides a lot of details about the killer's thoughts and actions, which help us to understand his motivations and personality. The killer is clearly a violent and sadistic individual, but he is also intelligent and resourceful. He is able to lure his victims into his trap and then dispose of their bodies in a way that makes it difficult for the police to find them. The killer's character is well-developed and believable, and it makes the story more suspenseful and engaging.

The question becomes, then, what sort of advice NotebookLM could give us in a situation when the character isn't plausible. That's where it could be really useful in fine-tuning character personalities or making them more believable or engaging.

Finally there is some good information NotebookLM can provide outside of writing specifically, and also for content warnings and sensitivity analysis. In response to the prompt "for what age range is this story appropriate?" NotebookLM responded with this:

This story is not appropriate for children or young adults. It contains graphic violence, sexual violence, and descriptions of murder. It is also quite dark and disturbing. I would recommend that adults only read this story if they are prepared for its content.

More or less accurate (though sexual violence is not explicitly portrayed or even really implied—I thought that was an odd assumption).

Regardelss, this is just a quick first look at the NotebookLM tool. I'll keep playing with it and post a followup as I work with it more and, more importantly, evaluate its ability to find references across multiple sources (stories). I can see a ton of possibilities to help write more tightly-plotted works with fewer continuity errors ultimately leading to a better reader experience. I get the feeling there is a goldmine of functionality here waiting to be explored.

I'd be remiss if I didn't mention that these sorts of analysis tools are available on other platforms. As Pierce notes, "Dropbox, Box, Notion, and others have built generative AI tools that let you interact with and create new things from the data you already have in their systems." Those tools probably merit a close examination and evaluation of their own.

If you're a writer, particularly one who uses Google Docs as your main writing tool, I'd recommend checking out NotebookLM and playing with it yourself. If it's a sign of things to come, I think it'll be a valuable addition to the stable of tools modern writers can rely upon and enjoy to produce higher-quality, more polished work.

References

Pierce, D. (28 August 2023). "Google’s AI-powered note-taking app is the messy beginning of something great." Retrieved on 4 September 2023 from https://www.theverge.com/23845856/google-notebooklm-tailwind-ai-notes-research