Hey ChatGPT, write an introduction
to my next YouTube video. [Music] "Hello and welcome to this video where I share
my experience as an author using ChatGPT to write fiction. As a writer I'm always on the lookout
for new tools and techniques to help me in my craft that's why I was intrigued when I first
heard about OpenAI's ChatGPT and its potential to revolutionize the writing process. In this video
I'll take you through my journey using ChatGPT to write fiction and share my insights on how it's
impacted my writing process from generating new ideas to fleshing out characters I'll show you
how ChatGPT has become an indispensable part of my writing toolkit so whether you're an aspiring
author or a seasoned pro this video is for you. Let's dive in!" So I don't know if I would say
/indispensable/ but we'll get into that. I can't tell you how many takes that took to read that
verbatim. So in my last video I talked about how I used ChatGPT to write a short story collection in
two weeks. I have so many notes from the process that for that video I just picked the high level
things to go over but I figured in this one I will do more of a deep dive on the process and all
the little funny things that ChatGPT tended to do. [Notes on Generating with ChatGPT] So at
first I could not get ChatGPT to make up a new word to name a fictional land based on Greek
mythology. Every time I would try, no matter what I prompted it to do, it would continually
give me names of real places, or places from Greek mythology. But later on towards the end
of the process I was actually able to get it to generate some names for a village that were not
based on real places. I chose Shadowdale from its suggestions. A very simple way of creating fantasy
town names, but it totally works for me. So when generating a ton of content if your prompts are
vague GPT will start re-spinning the same things and all of your paragraphs will be very similar
it doesn't take long to figure out that you have to be really specific with your prompts but you
can spend a lot of time tweaking your prompts using specific language is very important when
you're generating and you're kind of wrangling the AI at times. ChatGPT sometimes struggles
to understand your whole prompt if you're not very clear about what you want it to change where
especially when working with larger chunks of text like a few paragraphs or more if I ask it to write
something with a few paragraphs and then ask it to rewrite some details it will usually change those
details and then repeat the rest of it exactly how it was before and instead of taking that new
information into context. Along with having a very basic/generic/bland writing style, in my opinion
it also has some pretty poor writing skills. Sometimes the prose is full of passive verbs and
unnecessary wordiness now this is something that OpenAI states on their website as being a current
limit/issue with ChatGPT. Most of its sentences start with nouns and it does a lot of telling
instead of showing or sometimes it does both at the same time one after the other. That leaves a
lot of redundant content that you have to cut out later. It is very difficult to get ChatGPT to go
into detail and you have to tell it not only to add detail but what kind of detail exactly. "Add
detail about how the character feels afraid of the monster" or "add detail about the setting to make
it more ominous." Without this kind of information you get really basic output I ended up pretty
much putting two or three detail requests in every single prompt and while this does mean prompting
is slower I think it saves time overall. I found it really really helpful to make small adjustments
as I go sometimes I just couldn't get it to work with me and I just had to take two or three of
the different versions it created and combine them together to get what I was looking for this
really does slow down the process but at some points it was necessary and is faster than trying
to continually regenerate things to get something that was better. Something else that I learned
pretty early on is that if you want dialogue in the texted outputs you pretty much have to specify
that even if you say write a paragraph where these two characters talk about this it will sometimes
not give you dialogue I would say only one in three times will it actually put dialogue in the
paragraph which can be very frustrating usually I would have to prompt and say "add dialogue"
otherwise it would just say "And then they talked about this." Also if you want dialogue you have to
be really specific about how the conversation goes because a lot of the times it just can't make up
these details and fill in like a natural sounding conversation. So on January 11th I wrote that "I'm
not sure if ChatGPT learns to predict what you're going to ask but now it's adding title options for
scenes and stories after I prompt for an outline. Interesting and helpful but also scary." Agreed
Past Sydney, a little bit scary. I would usually have it outline a short story and then after that
I would ask it for some title suggestions for the story itself and then I would do the same thing
for each chapter so I don't remember when exactly this is um January 11th so this is a little bit
early on in the process but it was giving me title suggestions for chapters and stories while I was
still generating the outline which is something I had previously asked for so that's kind of
interesting. [Notes on Outlining with ChatGPT] The Collection follows a hero named Ajax in a Greek
mythology type world and while I was outlining the last story in the collection I was having trouble
getting it to suggest things that weren't already covered in the Percy Jackson series so in the
first story Ajax meets a griffin who was sent by the gods to help him fulfill his destiny of
destroying a great evil so after that I was kind of steering each of the stories towards that great
evil at the very last story I spent quite a long time getting it to generate ideas before it came
up with something that wasn't super overused or just like some of the most popular things from
Greek mythology and retellings. I keep wanting to say "we" like me and ChatGPT are just two
people working together creating a story. So one kind of strange thing is that when I was
prompting it to create options for a powerful object that the evil sorceress could use in the
story, the first few times I prompted it gave me all of the well-known ones that have been used in
lots of Greek mythology stories and lots of Greek mythology retellings but I wanted to push it to
get something at least a little different so I prompted many times I couldn't get it to come up
with anything except for pretty much the five most popular options, so I started prompting it for
more generic options, went through a few of those lists of ideas but then went back to asking it
specifically for things based on Greek mythology specifically things crafted by the god Hephaestus.
After one of my prompts its response was this: "Some of the most famous objects Hephaestus
created include:" And then it gave me a list of seven objects. Five of them were actually things
found in Greek mythology that Hephaestus was known to have created... um however two were /not/. Two
were MADE UP. From the list of things that it said were famous from Greek mythology, two of them
were made up by ChatGPT. ChatGPT: "Nah, I'm just messing with you, here you go, here's some
fake ones." That goes to show you always check ChatGPT's work. This actually ended up working
out for me because I chose one of those those things that was very interesting I might not have
even noticed that except the last many times I prompted it it gave me things I had already heard
of. I'm not a Greek mythology buff by any means I have read the Percy Jackson books but definitely
no expert on the subject, I don't know a lot of the detailed or lesser known characters but this
is one of those things that while I was having it generating I was like "where is it pulling this
from? what is this? I've never even heard of this? Let me look that up on Google and see what it's
about. Oh! That's not actually a thing. Huh." This is something that is mentioned on the OpenAI
website about one of ChatGPT's kind of limitations or known issues is that the information outputs is
not always accurate. I kind of wasn't expecting it to come out this way when I was generating things
it was kind of interesting. So throughout the process ChatGPT would frequently change point of
view and as I didn't feel like adding "from Ajax's point of view" to every single prompt I decided to
just kind of go with the omniscient head hopping of it and kind of get a little bit of insight
into lots of different characters heads. I wanted to work with what I had and I didn't want to feel
like I was influencing it too much I really wanted to see like what the output would be that was kind
of the whole point of this experiment. It seemed like sometimes ChatGPT is more likely to take
into context the more recent answers that it gave you instead of the kind of generating history, so
sometimes it would kind of reference something or work off of something that was only you know three
or four generations ago instead of the information that I actually gave it so that was a little
interesting to work around you can kind of scroll up a little and read your last couple of prompts
or output and see where it got that information from but that's not exactly what you asked for,
but you can see how it went there. As I was editing the first story is when I realized just
how much is really missing from the output. Along with there being little to no characterization,
the story generally lacks... a point? I'm not sure how you would go about getting ChatGPT
to write a character arc, this would I think be really difficult for you to just kind of have
it naturally do you'd have to figure it all out yourself kind of beforehand or at least heavily
influence it. [Notes on Writing with ChatGPT] And of course my experiment that I did with Legends
of the Shadow Woods is probably not the way that most people are or would be using ChatGPT to write
fiction right now. I think you would be using it a lot more as an idea generator or characterization
or or writing small pieces of text that you then heavily edit, whereas me with this project I
kind of wanted to have the least amount of my influence as possible on the generating side
of it and then just take what ChatGPT output and edit it enough so that people could read it,
but not so much that I was really affecting the way that it sounded and stuff I wanted to preserve
as much of the AI content as possible while still making it a readable story. And I think it would
be kind of interesting for me to do another one of these projects in the future but using it to
write something more in my style. I think it's kind of funny how something can be so unique
and so kind of basic at the same time. Legends of the Shadow Woods that I created is unique
because of the technology used to create it, but basic because of how bland it is. It is not
quality fiction. The driver in interest here is reader's curiosity about AI not necessarily
the story or the characters themselves. Right now generating with ChatGPT for making a book or
story is not as easy as prompting and copying and pasting. The output is clunky at best and if not
prompted specifically, bland and very repetitive. I keep saying "right now" and "yet" and "in the
future" because as much as some people might hate the idea of AI programs like this I don't think
there's any stopping them I think because there is money to be made here I think they will continue
to improve and be used. I can see in a few years ChatGPT or similar programs could output quality
fiction. "Quality Fiction" whatever that means. Something that requires little to no editing on
the human author's part this is most definitely a whole video topic or a video series topic by
itself so I will talk about that another day. It's not like I'm just sitting here telling it to write
a story about this, taking that, and publishing it. To get something readable we still need a
lot of human skills. I did a lot of editing and it is still not up to my standards as a reader or
writer. [Notes on ChatGPT Being Difficult] So here are some notes on ChatGPT being difficult. This
note is from generating the second story about the character's origins. "ChatGPT is in a mood today I
guess it keeps giving me glitchy results like just repeating the same sentence over a dozen times
at the end of the paragraph or repeating the same sentence a bunch of times with slightly different
versions like it's using a thesaurus I'm going to have to cut through so much unusable content for
this." I remember this and it was really annoying. The stop generating feature is very very helpful
and I'm so glad they added it but this was before they added it so I just had to sit there and wait
for it to be done which was a while sometimes. It would basically just repeat itself it kind of
got stuck in a loop or something so it would just repeat the same sentence for like I don't
know 50 times 80 times um so while generating the next story I prompted for a few paragraphs
about two characters sneaking into a king's tent on a spy mission when they discover a letter that
gives them some clues I then prompted ChatGPT to write the contents of the letter and here's what
it said, it said: "I'm sorry but I am not able to write the specific letter you have described as
the contents of the letter is a crucial component of the plot you have outlined and as such it would
spoil it but I can provide you with an example of a generic letter or document on a given topic."
So after this I regenerated the exact same prompt and it gave me exactly what I was looking for so
sometimes it's like it's telling a story and it doesn't want to spoil it to you but by what it
said it was like "I don't want to spoil it for you even though you already outlined the story
and know what the letter says" it was kind of weird so for a while it also kept switching tense.
It was not doing this before but it happened a few times within one hour of prompting it's easy
enough to re-prompt and get it to write the same thing but in past tense. Here's another note
"GPT refreshed or something as I was generating the LAST PARAGRAPH (that was in all caps) of story
three. basically got what I needed before it got rid of the chat history and GPT was down for a
while that day and I almost lost a whole bunch of work." save your work, kids. I also moved
into a new chat for the next story as it was difficult to keep organized near the end. so it
seemed like the longer the chat I'm working in the more ChatGPT gets slower and more chaotic
it had a higher tendency to give longer chunks of repeating text it sometimes restates something
a bunch of times adds paragraphs of summarizing so while working on chapter 3 of the last story,
there were five chapters in each story, I decided to try adding a small bit of foreshadowing just
to kind of play with different ideas in ChatGPT. I prompted it for a paragraph where the characters
notice an inscription on the cave wall at near the entrance later they come to a puzzle after that
I only prompted ChatGPT to write about the group trying to hop on different tiles trying various
combinations all on its own it tried to reuse the thing I had to add previously from the cave
wall. I accidentally hit stop generating and I was totally kicking myself for that I redid it and
reproduced it. It made the characters come to the conclusion I was planning on, all on its own. It
is important to mention here that it was just a few prompts ago and right before that prompt I was
asking it for information on the hierarchy of the Greek gods but it put together that that's where
I was going like it came to the same conclusion. I had it write "Follow the 12" which was the
"mysterious" inscription on the cave wall, 12 major Greek gods and they would go into the
cave they would go through they went through lots of tunnels and fought some bad guys and then came
to a puzzle room they had to step on the tiles in the correct order and tiles all had different
little inscriptions on them it figured out that that's what the inscription on the cave meant, is
the 12 major gods and all the titles on the floor like that, I gave it all that information, so it
makes sense that somebody could figure that out, but /ChatGPT/ figured that out and that kind of
blew my mind. I think I feel equally both ways it was freaky and awesome. That just goes to show you
though that the more information you give it the better ChatGPT works. [Notes on AI Detectors with
ChatGPT] Okay so my last section here is about AI detectors early on in the process I took a sample
of what ChatGPT put out and put it through a few different free AI Content Detector websites, and
they all said 95% to 100% what I put in was AI generated content. The only thing that I had
done to the text is switch around the position of a couple of sentences but the words were all
the same. So just to test them I put a sample of my own 100% non-AI generated book into the same
websites and they all said 5% or less chance of being AI generated text. So then after that I took
the original chunk of text and did a quick manual edit of the AI generated text and used a few of my
ProWritingAid tools, then proofread it and tried this again. So after editing that chunk of text
the AI detectors said 88.6% AI-generated content, 51% AI-generated content, 93%, 94%, and the
last one 47%. Like I said before my goal with editing this project was not to make it into my
own style at all this is very different from my usual style and I would basically be rewriting it
if I were to do that, not to create a story with Incredible prose or to have a deep moving story
but to see what this thing could do if I worked around it. I kind of approached it like I was like
an editor or a beta-reader or like a hhostwriter, almost. A ghostwriter and an editor and a beta
reader all very different processes but somewhere in there you get the idea. So of the edited 294
words, 198 of them were AI-generated and 96 of them were added by me. That's about 32% human
content I actually expected those quick edits to make a bigger difference as far as the AI
Content Detectors were concerned I figured the main thing they were looking at was the things
that I was seeing the sentence structure and repeated word usage clearly these things are
looking at more factories than that because I wasn't really able to fool them. I'm not really
sure how these content detectors work or anything but I thought this was really interesting. So
there is my experience using AI to create a short story collection although I published this
as an ebook I also formed it as a hardcover this is my proof copy I'm not sure if anybody else
would want a hardcover copy because the ebook is only 99 cents but I figured I will put it up
there anyway just in case anybody else does. So I think in my next video I want to talk about
AI and the future of writers and media so stay tuned for that all right that's all I got
for you so thanks for watching bye [Music]
