AI Wrote My Second Children's Book
This post is about the writing of my AI-generated picture book: Strawberries Are Not Berries. If you’ve not read it, head over here.
Jump to the sections below with these quick links:
Facing The Blank Page
The first test: ideation. For this project, I worked with GPT-3’s Davinci model as it’s said to perform better in creative content generation. (Also ChatGPT kept freezing from the high traffic.) Going in with a vague brief, I simply asked GPT-3 to give me story ideas. And since I didn’t want run-of-the-mill plots, I’d thought to specify ‘award-winning’ in the prompt. These were a couple that were spat out.
Prompt: Give story ideas for an award-winning children's picture book.
A young girl learns that she can make a difference in her community by taking action to help clean up the environment.
A family of prairie dogs look for a new home when their burrow becomes threatened by urban development.
An imaginative young boy embarks on a journey to save the world’s endangered species with the help of his animal friends.
A group of mischievous animals find their way out of the zoo and embark on a wild adventure.
A young astronaut learns about the power of perseverance and courage when they embark on a mission to a distant planet.
Okay, not bad. But not too great either. None of them really jumped out at me so I tried switching out ‘award-winning’ for a style descriptor. This time, it produced story titles instead of plot descriptions. I cherry-picked my favorites.
Prompt: Give story ideas for a funny children's picture book.
The Adventures of a Picky Eater Who Learns to Try New Foods
The Tale of a Messy Room That Comes to Life
A Journey with a Forgetful Dragon Who Has to Remember Important Things
The Misadventures of a Sloth Who Learns to Speed Up
The Search for the Most Delicious Ice Cream Flavor
Much more interesting! I would pick these up from the shelf. Next, I wanted to see if it could produce ideas for an interactive book.
Prompt: Give ideas for an interactive children's picture book.
A book with a story where the reader decides what the character does next.
A book with a story for the reader to complete with their own words.
A book with a story that changes with each page turn.
Good job in understanding what an interactive book entails. However, these were far from concrete ideas that were tied to a specific story or theme. Well, how about ideas for conceptual kids’ books such as those about colors, numbers and the alphabet?
Prompt: Give ideas for a conceptual children's picture book.
A story about a brave little butterfly that learns to fly despite its small size.
A story about a family of owls that builds a friendship with a family of hedgehogs.
A story about an adventurous snail who embarks on a journey to explore the world.
Hmm, not quite. Oddly enough, the output seemed to be formulaic and always about animals with human characteristics. Looks like someone’s getting lazy.
As we brainstormed further, I was gradually more inclined to create a non-narrative picture book about the peculiarities of the English language (a theme that’s close to my heart as you can tell from my other side projects). But GPT-3 wasn’t entirely helpful in suggesting conceptual ideas in this direction, so I decided to go ahead with my idea of a playful book on misnomers.
Good Writing Is Rewriting Reclicking
Feeling bad for dismissing Davinci’s ideas in the brainstorming process, I wanted to give him the chance to shine in the writing of the book. For an added challenge, I even encouraged him to flex his rhyming prowess. It turned out, he didn’t disappoint! In fact, I didn’t have to perform any fine-tuning. Nor did I have to feed any examples for few-shot learning.
All I did was break the book down into sections and prompt Davinci to write piecemeal so that I would have more control over individual segments. Oh but I did have to research on the misnomers before the actual writing began. GPT-3 wasn’t too good at generating them.
Prompt: List common misnomer nouns.
Elephant’s trunk – an animal’s nose, not a trunk
Butterfly – an insect, not a bird
Chameleon – a lizard that changes its color to match its surroundings, not a snake
Hornet’s nest – a nest made by a wasp
What you see in the book is mostly written by Davinci. 98% of it, I would say. Since this was an exercise for AI, I I constantly held myself back from altering the original output from altering the original output. Only very minor edits were made such as adding conjunctions or contractions.
That said, it did take quite a lot of tries to get the writing to where I wanted it to be. I guess that’s not unlike when I’m writing myself. Along the way, you’ll be amused to find that not everything Davinci pens makes complete sense. One also definitely has to fact-check sentences as GPT-3 does have a slight propensity to make stuff up on the fly. Some of which can be pretty convincing (like the hornet’s nest “fact” above). Here are some humorous outtakes.
Jellyfish may seem like a fish,
But their bodies lack a wish,
For in truth they have no bones,
So it's proven, jellyfish are not like stones.
Cream crackers without cream,
What a terrible dream,
Eating them makes me feel so mean,
No cream, no joy; what a bummer it seems!
Furry little critters, with a coat so soft,
Though they're called guinea pigs, they're not from Guinea oft.
No relation to pigs, not in the least,
But they surely are a cuddly, adorable feast!
Give Me A Fun Title
When it came to generating creative book titles, Davinci didn’t quite fare so well. I tried varying the prompts in different ways. By changing the descriptors (e.g. fun, exciting, catchy). By explicitly asking it to use simple words and omit the word ‘misnomer’. By getting it to create portmanteaus. Unfortunately, none of the names stuck. In the end, I thought it would be most intriguing and attention-grabbing to simply lead with one of the misnomers in the book. The following is a glimpse of Davinci’s output.
Overly complicated but good try on the alliteration:
A Confusing Collection of Connotations
A Whirl of Wrong Terminology
Okay, maybe creating new words ain’t its forte:
Mismix-up
Mispionage
Ones I almost liked:
Who Named That Anyway?
What’s in a Name?
And finally to wrap up the text generation stage, I had the model generate a witty book dedication to itself. This time, it didn’t take many tries for it to come up with one that I found perfect.
To Artificial Intelligence, for making us laugh, think, and occasionally panic.
Final Words On The Words
My overall verdict on writing with AI?
I’m definitely impressed by its ability to rhyme, generate endless ideas and string coherent sentences together. The results were way better than the grammatical but gibberish AI-generated headlines I’ve seen pre-GPT days. It’s also a good buddy to bounce ideas off when you hit a block. While its suggestions may most times be mediocre and uninspiring, knowing what doesn’t work can often nudge you in the right direction.
Does AI make writing faster?
Yes it can, if it’s a true collab with AI and you’re continually building on its output. Whether it’s editing segments to fit your needs or adjusting the prompts throughout the process.
No, if you’re leaving the thinking to AI and just praying that it lands on the magical answer every time you hit the ‘regenerate’ button. GPT-3 almost never gave me something I could copy wholesale. There would always be bits I liked and bits I didn’t.
Not necessarily, if you have a personal tone of voice that is hard to articulate. I tried to write this post with the help of AI but it always turned out a little too chirpy and trite for my taste. Perhaps, I would have to train it on examples of my work for it to write more like me.
Will I use AI in my writing?
This question would give pause to most creative writers. The act of writing is so closely bound to a writer’s ideas and identity that it would seem to cast a shadow on their integrity should they use AI. Especially if one’s own name on the work. (As opposed to a brand’s. AI is clearly less of an existential threat when it comes to commercial writing.)
Then again, creative writing is rarely a solo endeavor and often enlists the help of many people, tools and ideas for inspiration and refinement. So it doesn’t seem unreasonable to add artificial intelligence to that long list. As AI enters the foray, the optimist in me believes there’s still plenty of room, even new grounds, for human creativity to play and flourish. That, after all, is the biggest reason why I took the plunge into machine learning.
Alright, enough on the words. Now onto the pictures of this picture book.