This morning, I finished building out the basic Story Bible for World War 3.X and used it to nail down the six-book arc for the rest of the series. The Obsidian vault is already more useful than the ‘old’ GPT I built to use as an Axis of Time bible last year. For one thing, it doesn’t hallucinate.
When I trained the original GPT on all of my Axis books to use as a story bible, it kept insisting the central character was Matthew Kolhammer.
His name, of course, is Phillip Kolhammer. He’s the beating heart of the whole fucking series. There’s no Matthews anywhere near him. Yet the bot spoke about “Matthew Kolhammer” with the calm assurance of a $10K-a-day barrister reciting precedent.
The problem arises because GPTs don’t retrieve facts the way a search engine does. They predict/guess. Every answer is just the robot guessing the most likely next word based on its training. And if the probabilities line up wrong — if “Matthew” feels just as plausible as “Phillip” — it will confidently invent Admiral Matt’ for you, right at the centre of your canon.
Ninety per cent of the time, the AI is fine. It’s brilliant at ideating… fast. Ask a general question about Kolhammer’s command style or whether Buster Cherry died in Hawaii, and the answer will probably come straight out of my own text. Probably. But the other ten per cent? That’s when the improv show kicks off and your story ‘Bible’ turns into a lot of plausible nonsense. And because the bot delivers those hallucinations in the same smooth voice it uses for the truth, they’re dangerously easy to fall for on lesser queries.
That doesn’t make GPTs useless as data management tools for writers. They’re lightning fast, endlessly fucking patient and willing to trawl through a million words of back catalogue without complaint. They’ll happily talk all day with you about the sort of character/story arc questions that most human beings will eventually run screaming from. But you have to treat them as what they are: unreliable narrators. If they’re doing anything other than spitballing ideas, if they’re ‘retrieving’ story data, for instance, ask for direct quotes. Demand citations. Sanity-check anything that doesn’t feel right.
Unless you want to retcon yourself out of a hole in the next book, explaining that Matthew Kolhammer is actually Big Phil’s identical twin who’d snuck aboard the Big Hill to give him a birthday card just before the Transition and…
Wait a minute.
Just had an idea…
Check out `notebooklm` from Google.
You can load all your work into it, then ask it questions (similar to ChatGPT). BUT IT HAS REFERENCES!!!
It will show you exactly where its getting its answers from, in the documents you uploaded, so you can double check.
I've heard ChatGPT basically makes stuff up. There was the woman (on the Gold Coast maybe?) who used it to develop her defense in court. The judge in handing down his decision against her, slammed her defense referencing cases that actually didn't exist. Had to laugh. Welcome to the brave new world of AI.