According to a new article in Fortune, Bill Gates is annoyed that “the AI revolution means everyone will have their own ‘white collar’ personal assistant”. Most writers I know don’t have billions of dollars of pocket money to hire the best and brightest ghostwriters, or a whole publishing company to do editing, distribution, and marketing. The best we can afford is a free, high-end AI writing assistant that produces instant results.
In the recent past, I have hired a large roster of contractors to help me with various projects. Often, I had to wait up to two weeks to find out I would not receive what I was promised. That led me to start double hiring so I could have a backup in case someone let me down.
What I want for my creative projects is forward movement. I do not want to be miserable, tired, spent out, and burnt out. Deep learning language systems like OpenAI’s ChatGPT and Facebook’s LLaMA are the technological upgrade we have been waiting for, and I encourage you to start using them as tools. In this post, I will use ChatGPT and my third novel (WIP) to show you how the system works. It is really easy to work with an AI writing assistant.
First, you should start a new chat and tell the system your objective. I said, “Pretend you’re a fiction author. Please be my creative writing assistant today.” Then, I explained that I would focus on three areas: character descriptions, realistic dialogues, and character motivations. After you do that, adjust the details and send one request at a time.
Generate character descriptions
My third novel has one principal character, a woman named Hurricane Nisto. Hurricane is extroverted but not very talkative. She prefers to observe her surroundings because she is always searching for inspiration for her next fantasy. But what does she look like? In the chat, I wrote notes on Hurricane’s personality, nationality, and work environment. I then said, “I don’t have any idea what she looks like.”
ChatGPT’s first response was a rehash of what I wrote, so I tapped “regenerate response”. This time, the system gave me a physical description of Hurricane.
Hurricane Nisto is a woman in her mid-thirties with Nordic features that reflect her Norwegian heritage. Her long, blonde hair is usually styled in a loose ponytail, framing her heart-shaped face, sharp cheekbones and piercing blue eyes. Her tall, slender build is often hidden behind loose-fitting clothes. She prefers comfortable, casual clothes because they allow her to move freely.
Draft realistic dialogue
After finding out what Hurricane Nisto might look like, I now focus on one of her conversations. She has a vivid imagination, which distracts her from the mundane tasks of her government job. Fantasising got her in trouble and this is the basis for the events unfolding in the story.
Most of Hurricane’s conversations take place online. At one point in the story, Hurricane buys an Italian castle for the unbelievably low price of ten kroner (roughly US$1). Then, she finds out she has to renovate it. She lives in Antarctica so she needs to hire an assistant who lives in Italy.
The details I gave ChatGPT were that a castle was purchased for US$1, that Hurricane is living on a remote island, and she used a digital bulletin board to advertise for an assistant living in Italy. The output told me that a candidate, Priya, travelled to Antarctica to meet Hurricane in person. This was a mistake, given the specific details I provided, so I reminded the system to pay attention. I said, “Priya never leaves Italy and Hurricane never leaves Antarctica. They have to work virtually. Please start over.”
My question was, “How does their conversation start? Do not include self-introductions.”
Apologies for the confusion earlier! Here’s a revised version of the conversation starter: Hurricane and Priya connect over a video chat app on a social media site. They’ve never met in person and are communicating virtually due to the distance between them.
And this is a portion of the dialogue generated:
“It’s a long story,” answered Hurricane, “but basically, the castle has been abandoned for years, and the government sold it for a very low price as part of a preservation project. I saw the opportunity and took it, even though it’s located so far away from me.”
“That’s fascinating,” was Priya’s response. “So, how can I help you with the castle?”
“I need someone to inspect the property and give me a detailed report of its condition,” answered Hurricane. “Can you do that for me, Priya?”
An interesting draft, and the tone is appropriate for a professional conversation. And now I have an idea of how helpful (and fast!) the software is for drafting dialogue based on a general prompt.
Brainstorm character motivations
Remember that Hurricane Nisto is an extroverted person who is not talkative. At the start of the story, we know why she ends up living in Antarctica. She breaks a rule at work, but does not get fired because her government job is secure. The workaround is to send her to Antarctica. In her new job, she supervises a weather station on a Norwegian outpost.
The incident was a “stairwell tryst.” The system already knows that Hurricane is a fantasist, and is somewhat detached from what’s happening around her. So, I asked simply, “Why would Hurricane have a stairwell tryst with a supervisor in her office?”
After reviewing the results, I told the software that there was no “affection” between the characters, and that Hurricane was neither materialistic nor ambitious.
ChatGPT’s 3rd attempt:
Hurricane Nisto has a tendency to daydream, and if she is not motivated by personal gain or advancement in the workplace, it’s possible that her affair with the senior colleague is driven by a need for validation or attention. She may feel overlooked or invisible within the organisation. And the attention and validation she receives from the senior colleague may make her feel seen and valued. Or, Hurricane may be bored or dissatisfied with her life, and the excitement of an illicit affair may provide a temporary escape from the monotony of her routine.
These were valid motivations, but this output required three generations. The model had a difficult time understanding that Hurricane has low self-awareness. I agree that Hurricane may be bored or dissatisfied with her life, but her character is nuanced. She acts like someone who thinks, but in reality, she is driven from one circumstance to the next. To fully develop the story, I will need to show the reader that Hurricane is likeable despite her lack of self-awareness.
One of the wonderful things about working with a deep learning language model like ChatGPT is that I have grown more confident as a writer. There are lots of people saying that AI will replace creative writers, but I am sure that won’t happen. Creative writing requires flexible thinking which present AI language models might not be able to handle.
Robots cannot tell a story with deliciously intriguing characters. Even with detailed prompts, they cannot present nuanced emotions and personality quirks. They may have a low tolerance for human flaws. But they can provide support with technical research, and provide feedback on storytelling techniques.
The secrets to generating useable output are to create detailed prompts and establish baselines for the output. Give plenty of feedback. Assistants don’t read minds and this software is not clairvoyant. Please practice, and tell me what happened.
Thank you for reading. Best of luck with your writing tasks.
You must be logged in to post a comment.