By Erinola Enoch Daranijo
How the language model from OpenAI can help write poetry, fiction, and screenplays
“modern writing desk featuring a legal pad and pen, with a laptop and office supplies,” Midjourney Image by Author
If you have been reading my posts on Medium, you will know that I have been writing about creative uses of AI since August 2020. I often write about generating digital art, but I occasionally write about using AI for other creative purposes, like writing prose and composing music. When I heard that OpenAI released a new language model called ChatGPT [1], I jumped right in and tested it for creative writing.
Overview
This is the first of a three-part series on using ChatGPT as a writing partner. The series will cover writing prose, composing music, and creating picture books.
In this article, I will review some background info on ChatGPT and then show the results of my experiments in generating text for various creative projects:
Writing Haikus
Creating plot summaries for new novels
Creating new screenplays for TV shows and movies
I’ll finish by giving you my general observations on using the model with some possible next steps for future exploration.
ChatGPT
ChatGPT is the latest text generation model from OpenAI. They offer the system as a “Free Research Preview.” You can sign in with an email account or use an existing Google or Microsoft account. When you sign in, you see the home screen with examples, capabilities, and limitations.
ChatGPT User Interface, Image by Author
To start a thread, I asked a question like this.
ChatGPT Answers a Question, Image by Author
Note I used the same question in my AI 8 Ball project from 2020. You can see above that ChatGPT is familiar with Mattel’s Magic 8-Ball, and started the answer using one of the 20 canned responses. But ChatGPT went further to give some background info, which was nice. When I hit the Try again button, it generated similar text. But I used the cool new feature and asked the system a follow-up question.
ChatGPT Answers a Follow-up Question, Image by Author
Notice how ChatGPT remembered the context from the first question, so it knew what “it” meant and answered appropriately. It answered as the Magic 8-Ball would have and provided some helpful background information.
OpenAI added a nice feature of keeping a history of your prior chat sessions. This allowed me to pick up where I left off to continue the chat thread if needed.
ChatGPT Origins
ChatGPT is based on the GPT-3 large language model from OpenAI [2]. The model is huge. It has 175 billion artificial neurons and was trained on about 500 billion pieces of text gathered on the web.
OpenAI further trained GPT-3 to create a model called InstructGPT that tries to follow the users’ intentions when generating text [3].
Starting with a set of [human]-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. - Long Ouyang et al. from OpenAI
ChatGPT is OpenAI’s sibling model to InstructGPT, which was trained to work interactively with users [1].
We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. - OpenAI
Societal Bias in Language Models
Note that, like most large language models, InstructGPT and ChatGTP both suffer from exposure to implicit social bias and toxicity in the original training data. To combat this, OpenAI actively worked to “align” the models to be helpful, honest, and harmless [3].
We make progress on aligning language models by training them to act in accordance with the user’s intention. This encompasses both explicit intentions such as following instructions and implicit intentions such as staying truthful, and not being biased, toxic, or otherwise harmful. — Long Ouyang et al. from OpenAI
However, the authors state that the new model shows only small improvements in toxicity over GPT-3, but no improvement in bias.
How OpenAI Fine-tuned ChatGPT
Here are the steps OpenAI used for training their latest system. PPO stands for Proximal Policy Optimization [4], a new form of reinforcement learning for improved outputs.
Diagram Showing Supervised Fine-tuning, Reward Model Training, and Reinforcement Learning via Proximal Policy Optimization, Source OpenAI
The “labelers” are actual humans that created ideal outputs in Step 1 and scored samples of generated text in Step 2 that were used to train a Reward Model (RM.) In Step 3, the RM was used to predict which output the labelers would prefer, and the results were used to refine the GPT-3 model using the PPO algorithm.
Writing Haikus
“golden leaves drift down, Autumn breeze blows through the town, nature’s tapestry,” Midjourney Image by Author
My first experiment with ChatGPT was creating Haikus, a form of poetry that originated in Japan. As I discussed in a previous post, language models seem to have difficulty generating prose with a set meter, like the 5–7–5 syllable structure typically used for Haikus. I noticed a distinct improvement with ChatGPT, however. You can see this in my interaction below, with the responses in italics.
Write a Haiku about Autumn.
Leaves falling softly down
Autumn breeze rustles through town
Nature’s tapestry.This is good, but use five syllables in the first line. The second and third lines are fine.
Golden leaves drift down
Autumn breeze blows through the town
Nature’s tapestry
The Haiku is pretty good! The imagery of leaves falling down and the autumn breeze blowing through the town captures the feeling of the changing of the seasons. But notice how the first attempt has a 6–7–5 syllable scheme. I then prompted it to refine the first line, and, sure enough, it created the second attempt with a 5–7–5 scheme, and it retained the essence of the thought.
Here are a couple of additional examples.
Write a Haiku about laughter.
Laughter like a bell
Ringing out across the land
Bringing joy to all.Write a Haiku about the Python programming language.
Python, serpent-like
Slithering through code with grace
Power in simplicity.This is good, but use five syllables in the last line. The first and second lines are OK.
Python, serpent-like
Slithering through code with grace
Power in its ease.
The one about laughter was pretty good with the metaphor about ringing a bell. And it hit the expected 5–7–5 count for syllables in a Haiku. The one about Python, is good, too, continuing with the analogy with a snake. It came out initially with a 5–7–7 scheme, but when I prompted it, ChatGPT generated one with the correct syllable count and kept the original meaning intact.
Next, I’ll share the results of using the system to create new titles and plots for novels.
“stack of books on a desk with a pen and pad of paper,” Midjourney Image by Author
Creating Titles and Plots for Novels
Similar to my PlotJam project, where I used a fine-tuned GPT-2 model, I asked ChatGPT to generate new titles and plot summaries for novels in a given genre. Here are the results.
Create a title and plot summary for a new historical fiction novel.
Title: “The Last Stand at Little Bighorn”
Plot summary: In the summer of 1876, a band of Lakota Sioux and Cheyenne warriors, led by the legendary chiefs Sitting Bull and Crazy Horse, make their last stand against the advancing U.S. Army at the Battle of Little Bighorn. As the two sides clash in a brutal and bloody fight to the death, the young warrior White Buffalo must choose between his loyalty to his tribe and his love for a white woman named Sarah, who has come to the battlefield as a nurse for the wounded soldiers. With the fate of the Sioux nation hanging in the balance, White Buffalo must navigate the complex and dangerous world of warfare, honor, and love to find his own path.The setting is OK, but the plot seems a bit corny. Try again with a new type of struggle for the White Buffalo character.
Title: “The Last Stand at Little Bighorn”
Plot summary: […] As the two sides clash in a brutal and bloody battle, the young warrior White Buffalo must confront his own personal struggles and doubts. Torn between his loyalty to his tribe and his desire for peace, White Buffalo must find a way to reconcile his beliefs and his actions in order to defend his people and their way of life. Amidst the chaos and violence of the battle, White Buffalo must make difficult choices that will shape the future of his tribe and his own destiny.
All right, not bad. I like that the story seems to be told from the perspective of one of the Native Americans. But, as you can see, I had to steer the system away from the clichéd “Romeo and Juliet” plot for the second iteration. I deleted the repeated part in the second plot summary for readability. As they say in “the biz,” ChatGPT is OK with taking notes.
Here’s another example with three rounds of interaction.
Create a title and plot summary for a new science fiction novel.
Title: “The Last Sentinel”
Plot Summary: In a distant future, humanity has spread out across the galaxy and established a number of colonies on other planets. However, a great threat looms on the horizon in the form of a hostile alien race known as the Xarga. As the last remaining sentinel, a lone soldier named Jake must embark on a mission to discover the source of the Xarga and find a way to stop them before they wipe out all of humanity. Along the way, Jake must overcome incredible odds and make difficult choices in order to save the human race.OK, but use a more futuristic name for Jake.
Title: “The Last Sentinel”
Plot Summary: […] As the last remaining sentinel, a lone soldier named Jaxon must embark on a mission to discover the source of the Xarga and find a way to stop them before they wipe out all of humanity. Along the way, Jaxon must overcome incredible odds and make difficult choices in order to save the human race.OK, explain briefly how Jaxon stops the Xarga.
As the last remaining sentinel, Jaxon is the only hope for humanity in their battle against the Xarga. After traveling across the galaxy and facing many challenges along the way, Jaxon finally reaches the source of the Xarga threat. There, he discovers that the Xarga are controlled by a single, central intelligence. Jaxon uses his advanced technology and combat skills to defeat this central intelligence, thereby freeing the Xarga from its control and ending the threat to humanity.
This is a good start, too. However, the summary hints at a lot of action and suspense, but it’s missing some important details of what the book is really about, who the main character is, and the motivation driving the protagonist. As a side note, I think it’s funny that ChatGPT thinks it can make a name seem more futuristic by adding the letter “X.” And I like that a “single, central intelligence controls the Xarga.” ChatGPT seems to be following that old piece of advice for new authors, “Write what you know.” 🙂
“movie set with a script on a table and coffee,” Midjourney Image by Author, based on a Photo by Jakob Owens on Unsplash
Writing Screenplays
I found that ChatGPT can be a good writing partner for creating screenplays. Similar to what I did with my GreenLit project, I experimented with the new model for creating a title, plot summary, and the first page of a script for a show given a genre. Here’s the initial interaction.
Create five titles and plot summaries for a new streaming series about a bank heist.
“The Heist” — When a group of skilled thieves team up to pull off the perfect bank heist, they quickly realize that the real challenge is getting out alive with the loot. As the police close in and the stakes get higher, the team must rely on their expertise and trust in one another to outwit the authorities and make a clean getaway.
“Inside Job” — When a group of disgruntled former employees hatch a plan to rob the bank where they used to work, they enlist the help of a seasoned thief to pull off the heist. As they dig deeper into the bank’s operations, they uncover a web of corruption and greed that puts them in the crosshairs of both the law and their former employers.
“The Bank Job” — When a group of friends decide to rob a local bank in order to pay off their debts and start anew, they quickly realize that their amateur heist skills are no match for the security measures in place. As they struggle to outsmart the bank’s high-tech defenses, they must also navigate a series of double crosses and personal betrayals that threaten to derail their plans.
“The Perfect Score” — When a group of brilliant students decide to pull off the perfect bank heist in order to finance their education, they quickly discover that their elaborate plan is much more difficult to execute than they anticipated. As they navigate the challenges of pulling off the heist and evading the authorities, they must also confront their own personal demons and insecurities in order to succeed.
“The Big Score” — When a group of seasoned thieves come together to plan the heist of a lifetime, they quickly realize that pulling off the job will require more than just their expertise. As they work to gather the resources and personnel needed to carry out the heist, they must also navigate a series of personal and professional setbacks that threaten to derail their plans.
These are good, but most of them are fairly generic. They seem like they have all been done many times before. It’s almost as if ChatGPT wants to play it safe and not go out on a limb with an odd take on a story idea. However, I did like the one about the group of students/thieves because it took on an interesting angle. But it needed a unique title.
The Perfect Score looks good, but suggest some titles that haven’t been used before. Lean into the student angle. And just the titles, please.
Here are five potential new titles for a streaming series about a bank heist featuring students:
“The Student Heist”
“The Class of Crime”
“The Grades of Greed”
“The Academic Heist”
“The Campus Caper”
These were all pretty good, but bland. And not too surprisingly, they all start with the article “the.” Note that if I didn’t ask for just the titles, it would have created new plot summaries.
Overall, I liked “The Grades of Greed” the best, but mostly for the alliteration. I then requested ChatGPT to write the first page of the script, after I decided to ditch the “the” in the title.
These are pretty good. Let’s go with Grades of Greed without the “the.” Please write the first page of the script.
Sure! Here is the first page of the script for “Grades of Greed”:
FADE IN:
INT. UNIVERSITY LIBRARY — DAY
We see a close-up of a student’s face, deep in concentration as they study. They are surrounded by stacks of books and notes.
CUT TO:
INT. UNIVERSITY CAFETERIA — DAY
We see a group of four students, JESSICA, MIKE, AMY, and NICK, sitting at a table together. They all look stressed and tired.
JESSICA: (frustrated) I can’t believe we have midterms next week. I’m never going to be able to afford tuition if I don’t get at least a B+ in Calculus.
MIKE: (sarcastically) Oh yeah, because that’s the only thing standing between us and financial stability.
AMY: (serious) Actually, Mike, it’s a lot more expensive to repeat a class. If we don’t pass, we’ll have to pay for it again next semester.
NICK: (determined) Well, we can’t let that happen. We need to find a solution.
JESSICA: (sarcastically) Yeah, like robbing a bank.
The group falls silent for a moment before breaking out into laughter.
FADE TO BLACK.
OK, not bad. The dialog seems natural, albeit in a bit of a rush to move the plot along. I liked the way the system added emotional cues for the characters’ lines. This brought some nice color to the script.
However, the characters are a bit bland, as it seemed to only have a bunch of white kids in the story based on their first names. This is probably because of the societal bias in language models, as I mentioned above. But the system knew what to do when I asked it.
This is great! But try something similar with a more ethnically diverse group of characters.
Sure! Here is the revised first page of the script for “Grades of Greed” featuring a more ethnically diverse group of characters:
FADE IN:
INT. UNIVERSITY LIBRARY — DAY
We see a close-up of a student’s face, deep in concentration as they study. They are surrounded by stacks of books and notes.
CUT TO:
INT. UNIVERSITY CAFETERIA — DAY
We see a group of four students, TARA, JAVIER, AISHA, and KEN, sitting at a table together. They all look stressed and tired.
TARA: (frustrated) I can’t believe we have midterms next week. I’m never going to be able to afford tuition if I don’t get at least a B+ in Calculus.
JAVIER: (sarcastically) Oh yeah, because that’s the only thing standing between us and financial stability.
AISHA: (serious) Actually, Javier, it’s a lot more expensive to repeat a class. If we don’t pass, we’ll have to pay for it again next semester.
KEN: (determined) Well, we can’t let that happen. We need to find a solution.
TARA: (sarcastically) Yeah, like robbing a bank.
The group falls silent for a moment before breaking out into laughter.
FADE TO BLACK.
Sure enough, the system changed the names to have a wider cultural representation. I then asked it to tell me what it had in mind for their last names.
Thanks. For the credit roll, what are their full names?
Sure! Here are the full names for the characters in the revised version of the script:
TARA — Tara Patel
JAVIER — Javier Rodriguez
AISHA — Aisha Khan
KEN — Ken Tanaka
Nice! I’ll send this off to both Netflix and Amazon studios to see if I can start a bidding war to fund the project. 😀
Discussion and Next Steps
As you saw in my interactions above, ChatGPT may not be able to create good long-form prose all on its own, but it works well as an interactive writing partner. The system is quite knowledgeable and is a helpful collaborator. When prompted, it will create many examples of what you may be looking for, although many of its ideas are bland and generic. It takes some work to draw out any interesting thoughts, but it can be done.
In the past, when I used large language models like GPT-3, GPT-J, BLOOM, etc., for creative writing, I found that I had to generate many versions of the text and then “cherry-pick” the best ones. Even then, the prose didn’t have a well-defined beginning, middle, or end. The system just rambled on, and I had to decide when to cut it off. ChatGPT, however, seems to respond well to prompts and produce nicely composed stretches of text. I found that I rarely used the “spinny” regenerate button. Instead, I gave some notes back, and it almost always did what I wanted to improve the text.
As you saw in my screenwriting example above, I used a top-down writing process, starting with the title and plot, and then moving on to write the dialog and stage directions. It would probably be helpful to get ChatGPT to write up and refine character sketches for the principals in the story and then map out the storyline broken down into acts and scenes. This would help the system create a more cohesive narrative over the course of the screenplay.
As I mentioned at the head of this article, my next posts will be on using ChatGTP for creating music, with help from Band in a Box, and writing children’s picture books, with support from Midjourney.
Acknowledgments
I want to thank Jennifer Lim and Oliver Strimpel for their help with this project.
References
[1] J. Schulman et al., ChatGPT: Optimizing Language Models for Dialogue (2020)
[2] T. Brown et al., Language Models are Few-Shot Learners (2020)
[3] L. Ouyang et al., Training language models to follow instructions with human feedback (2020)
[4] J. Schulman et al., Proximal Policy Optimization Algorithms (2017)