The Arc Project – Final Report
This is the final report for The Arc Project which explored AI from a screenwriter’s point of view. If you don’t have time to dive into all of the findings, these are some tips to consider if you want to try it out yourself.
KEY LEARNINGS WHEN WORKING WITH AI AS A WRITER:
- Have a story/ plot line in place. Use a genre template if this feels right
- Use the AI model GPT3 but dial up the temperature to as high as you can go. This produces less predictable results.
- Experiment with your inputs and the language, length and timing of what you feed into the AI
- Take your time with it. Patience is key.
- Writer needs to have bigger input into the character traits. Writer needs to invent the flaws of the character.
- Dialogue was a big challenge because words are not reflective of character. So you are on your own there!
- If you really care about a theme, or if it is political, you might not want to use AI
- Best results are when you work in a group.
After some discussion with the good people at BoostHBG, we saw a unique opportunity to explore the use of AI technology as a creative writing tool.
The last 24 months have seen a kind of cambrian explosion in the development and use of AI in many industries. The creation of Dall E, and other closely related applications, have been particularly eye-catching. But they are also cause for refection from creators and consumers alike.
I became interested in AI after seeing a music project at CineGlobe in Geneva, where an AI application made a song with a human musician. I wondered how a writer could use AI to help them iterate faster. Then I wondered, if that was even a good thing?
Then comes the fear: what is the true role of the writer? Where is the value in being that thing that observes, that dreams up ideas, then translates them to screen.
Is it inevitable that AI will be writing about the human condition for human consumers in the near future? Will AI be a writer’s adversary or can AI be a genuine, collaborative companion to writers? If something rouses so many questions, you cannot afford to ignore them.
When I dropped this notion into conversations with other writers, just to test their reactions, the most common perceptions was that AI could, and most likely will, place a challenge on the shelf life for human creativity. Then it was commonly and quickly followed by the suggestion that AI data is biased. And indeed it is. Some of my colleagues took the question to its logical conclusion and suggested that we live in a simulation. Always fun to discuss!
But once we moved on from those ideas, a deeper discussion revealed a belief that AI modeling will outstrip the speed and iteration of human writing, and in about a decade (there is really no consensus as to when AI will be able to write a good story unaided by a human) film studios will be activating AI programs to dream up films to feed very specific audiences. So why encourage this? I felt as writers we should at least explore AI writing tools, get acquainted and collaborate so we understand its limitations. That at least means some power for the human writer.
Art has always been augmented by technology. So perhaps our fears as writers are misplaced. The question became: how could we use this new, ever-developing technology to support our ideas, or better still, to push us creatively and to get to the truth of the story we are writing. If there was something to learn I wanted to explore it and share with my fellow creators.
BoostHBG and I agreed to track my research, to publish it as a blog, and to report on what it is like to write a scenario for a film using the most current AI technology.
Outreach to the AI & Filmmaking community
Interest is one thing. Knowledge is another. Regarding AI, I had a lot of the former. Not much of the latter. I began to speak to friends and colleagues about their experiences with AI as creators. I spoke with local writer and technology advocate, André Hedetoft, VR journalist and philosopher Kent Bye, and David Košták, a very cool theater director from the Czech Republic who actually wrote an AI play and performed it. I wanted to understand their positions as creators. . It soon became clear: there is so much interest in using AI, but little use experience to refer to. (You can watch all of these interviews here).
They were very enlightening discussions because the interviewees were so divergent on their opinions. André wanted to learn how to make a game with AI (and by all accounts is succeeding). While Kent Bye was adamant that AI was just not up to the job yet. He stated that to write a meaningful screenplay or a story outline, was a kind of Holy Grail for AI researchers.
Yet one of my favorite discussions was with David who told me using AI was not so dissimilar to working with a real writer. The AI had good days and bad days – it just needed the right prompts. There was some management of the AI required to get the best results. David did not see AI as a way to speed up the iteration process. David insisted that you need to invest time in AI to really write with it. From his experience, AI writing is a long process, with significant periods of building and testing a relationship to an AI necessary to get the most out of it. And when managed well, there was a process at play where we kind of ‘dance’ with the software to help iterate on an idea and develop it to take it somewhere new.
For me this was exciting. The truth is that most of the time the creators we collaborate with have similar backgrounds to us: we mostly circulate in similar classes, with people with similar education or cultural upbringing. So when we create together, the chances are overwhelming that we are each drawing from the same wells of imagery and inspiration. An AI model, in comparison, is naive. How it responds to your work could be so left of field that it could be inspiring.
AI WRITING WORKSHOPS
Along with my friend and AI researcher, Mohanty Sharada, we set up a virtual workshop over Zoom with 3 local film writers: Måns Thunberg, Andreas Cliément and Filson Ali, plus three of Mohanty’s colleagues to seek out hard evidence: what are the problems that an AI program can solve for a film writer?
After an entire day of writing and debating, we learned what we already knew: time was an issue. Making time to write is always a challenge. Structuring a script takes practice and perseverance. But what about the idea itself: Together we learned how to actually write with an AI model called GPT3, which is hailed as one of the most impressive and realistic AI predictive models ever released.
We fed the GPT3 some scenarios from our writers, short sentences at first, then longer pieces of text. With Måns’ story he explained there was a block for him with a certain scene. He could not locate a good motivation for his characters, this affected their believable reaction. So we experimented and some of the results from the AI were almost entertaining. In fact, one of the AI’s limitations became its advantage: we quickly learned that the model had a short ‘memory’, in that it might repeat itself or start writing in a circle when the paragraph became too long. The AI made Måns’ characters have a car accident, then made them have another car accident while fleeing the scene of the first one. Hilarious. Much of the value for Måns came from the suggestions and ideas from the other writers and the programmers in the workshop. This is a great example of AI offering a twist on an idea, while the human interaction with the idea validates it and gives the idea meaningful context.
Encouraged by this mini-breakthrough in our mini-experiment we switched to a scenario by Filson. We became more daring and by adjusting some settings we increased the AI’s ability to become more unpredictable, to offer more unrealistic linguistic relationships in its answers.
This approach transformed the core of Filson’s proposed scenario. Her story began with a simple morning commute by a young copywriter in New York City, and transformed into a thrilling game of cat and mouse. With our prompts the AI came up with a scenario that saw the main character get stabbed on the train. It came out as almost subtle but intriguing. But it lead Filson’s story in a new direction completely as we grappled with who might stab the main character, what was this person’s backstory? Filson liked the idea but was clear her story would not revolve around violence, so ultimately her story might not go in that direction. But we were instantly able to see how an AI could in fact help a writer discover a new perspective on a story. In Filson’s case the AI helped us understand a major value for her work and for herself, while demonstrating firsthand how a script can come to life with a gentle AI push.
After a period of research and reflection I felt it was time to try to get a script together with just one writer. Being less involved with the actual creative process meant I could oversee the collaboration. We paired up with EliSophie Andrée, a Skåne based writer. I loved her shorter works and thought her dark sense of humor would be an ideal sparring partner for the AI.
Mohanty oversaw the tech side of the process with us on Zoom and we began to work on a story EliSophie had brought to the workshop. It revolved around two rival sisters, obsessed with social media, and ended with one cooking the other. (Dark but funny, believe me.)
Feeding scenarios to the GPT3 we found soon enough there was room for character development and depth that was lacking. Whatever story lines or dialogue we eked out of the GPT3 was passable but not believable.
We began to play with the prompts, making them more specific and experimenting with their word count and the language we used, and we came onto something that was unique, yet a familiar structure began to emerge. The AI introduced a mysterious character with ‘pale skin and long black hair’ who haunts one of the girls and her dreams. The new device worked because our new character (not yet protagonist, not yet antagonist), The Pale Faced Man, had no back story, no detail, but was instead a cocktail of evil and disturbing characters we connect with from pop culture. It was a great concept that I wanted to explore together. At the end of the day we had a new dimension to bring to this story. Due to Elisophie’s writing schedule she was not able to continue with the project. But the Pale Faced Man stuck with me.
The Final Step in the Writing Experiment
Originally, I wanted to use one of my existing stories to experiment with. The scenario revolved around a journalist from a small Indian town who goes on the run as he exposes government corruption. Not bad, I thought!
I began to input my own story points into a GPT3 AI model but the results were lukewarm at best. The responses the AI delivered to my prompts were hollow, simple and felt disconnected from the mood of the story. They were just too vague.
I began to feel that because the story was based in Asia, and the concepts themselves (freedom of speech, journalistic integrity, and government corruption) were not commonly written or talked about in culture, the AI could not create usable or ‘likable’ replies to my prompts. This realization was in itself quite telling.
When you consider the life of a normal mainstream film: upon release audiences will comment on it through social media, journalists write about them, data is tracked and analyzed, and film studios observe certain patterns and trends, then use that to create new films. At the other end culture consumes the film genres, and its surrounding culture, and infuses the symbols from that genre into other forms of media like advertising (for example). So of course the internet has a surplus of a certain kind of film, a story type that an AI model can draw inspiration from. It became clear my scenario did not fit the models that are popular with film culture. That is an issue. In way I felt this exercise was a futile. In order to write really well, to comment on your surroundings truthfully, the writer needs to be going against what common culture promotes. Artistic creations are at their best when they are critical of how we live, how we treat each other. AI is not able to be critical because it is a product. To survive a product needs to be liked. If anything AI is more likely to do the opposite and placate a consumer into not being critical of their surroundings at all.
At the same time perhaps my story was a bit… boring? Was its concern with freedom of speech, power of the press and fighting corruption just on the wrong side of popular to be appealing? Perhaps.
I had read that to create Deep Fakes and utilize AI in a convincing way, creators often worked with several different AI applications to fulfill various tasks. So I experimented with other AI applications to imagine the location my story was set in. Again, not enough data available. Try typing in ‘rural indian village by the sea’ and see what comes up.
I engaged with Philopsopher AI, to create more ‘profound’ sounding characters. Perhaps AI was good at writing dialogue? The lack of results, the lack of ideas I could really use, was frustrating me. The results from an AI were not as diverse or left-of-center as I experienced in the other writing experiments. I was not feeling inspired to use the answers as springboards.
Seeing a stark limitation in the ability of AI to help me write around the themes I want to explore, I realized it was time to shift gears. I switched and took up EliSophie’s story as source material, with her permission of course.
EliSophie writes from a personal space, but also for a specific audience. I wanted the script I write with an AI to connect to a wide audience when it is done.
Before you start any writing but particularly with an AI model, it is better to start when you at least have a setting, some character background, some sense of what the story is you want to tell: redemption tales, rags to riches, a vengeance story. This will help you work out your prompts and help you to start exploring a story world.
The art of the prompt was key: to work closely with AI one has to experiment with open-ended kinds of prompts to elicit usable answers from the AI. I invited the AI to explore the two main characters Cissi and Ela with me. Within an hour, with prompts like: CISSI AND ELA ARE SISTERS. ELA COMES TO VISIT CISSI FROM THE COUNTRYSIDE TO LIVE WITH HER IN STOCKHOLM, the AI soon fell back to the idea of the pale faced man. The idea reemerged even with different prompts. Disappointed.
Lackluster effort, GPT3.
So I turned up the heat to elicit more interesting responses.
The model gave me: a tall thin man dressed in black who introduces himself as a teacher in this boarding school.
I realized we must embrace and seriously consider whatever we are given so I go with it. I was faced with a couple of paths: either using this boarding school teacher as a kind of tool to create tension between the sisters, or make the character a villain. Or just use the suggestion for later and make him the evil, original Pale Faced Man.
This character became a problem up because I could not explain why Cissi and Ela live together by themselves in Stockholm when they’re so young. The AI forced me to answer this question.
I thought up something else: I wanted one character to be very vulnerable and be forced to leave home because her parents are getting divorced. Perhaps she is sent to a boarding school but escapes to her sister in Stockholm. The idea of an ‘Odd Couple’ became appealing for a moment and I toyed with the idea of them kind of destroying each other. Maybe I did not put enough time into the writing, but I instinctively felt: this scenario was not coming together.
I began to experiment with shorter prompts, and gave the AI problems to solve in the story. But today the AI had defeated me and began to just offer superficial words and illustrations of scenarios. This was not the sounding board I wanted.
I began to write blocks of text around the idea of ‘what the internet could be in the future’ -just to test out some ideas. The GPT3 came back with ‘Web 2. Web 3.’
I was now getting very frustrated with the GPT3. I start to enter new text, something around the girls taking a bus together in Stockholm, and the AI just spits out the car crash scene again. This is super frustrating. AI: you make my life harder.
But in time I came up with an idea: What if Cissi accidentally becomes the internet. The scenario I built up had motivations: Cissi leaves home because her parents are getting divorced. She locates her sister, Ela, in the big city, but Ela is utterly self-obsessed and is completely transfixed to social media and her phone. The period is set 25 years in the future, so I exaggerated the effect of the internet in culture with flying delivery cars, mind reading appliances and holographic advertising many kilometers high, a bit like Blade Runner.
I imagined a scenario where Cissi is left at home alone, discovers a huge fiber internet cable and by touching it she becomes the Internet. She controls it. The GPT3 liked this scenario. A lot.
It imagined how Cissi would throw society into chaos at a whim. How she would be able to breach military computers and launch missles. The GPT3 was excited. So I went with it and wrote up a short story where Cissi is in fact in control of the world’s internet and as she learns her power she has the world at her fingertips. She can do anything. The world is petrified of her and what she could unleash. It was quite fun to write. I had to invent a layer to the story where Cissi realizes she is adopted, and in the ensuing chaos of the tearing up society, her real mother sees Cissi on the news and sets out to find her. Cissi’s divorcing parents also set out to Stockholm to help ease the situation. I gave Cissi a best friend, a moral guide of sorts, who is ultra-wealthy and can also afford to do anything but is deeply detached and alone. But the two enjoy setting the world on fire together.
The culmination would be a redemption where Cissi finds love from her family, while society around her does not learn its lesson of being obsessed with the internet.
It seems the AI can make drama. But not emotion. Trying to write a scenario where the two young girls try to have fun is quite a challenge. When you ask the AI model to imagine two young girls have fun it imagines bike rides, playing a board game or baking. Is this what you would do if you had the entire internet at your fingertips?
I think we have come as far as we can go with GPT3.
I spent nearly a year with this experiment, trying various ways to understand and collaborate with AI. Some of my concerns were validated. AI is not well suited for stories based outside of Western culture. Perhaps it is not the right tool to help a writer create a memorable character, and it certainly is not good at writing dialogue. It was fun on a superficial level. I found some breakthroughs in the writing of my colleagues that potentially might not have been made. That is always a benefit. As they say, in art and as in nature: nothing is wasted.
But the limitations of AI (its short memory in particular, its lack of understanding of basic human conditions like joy) make it recycle ideas quickly. The surprise and originality I was looking for in AI’s responses did not materialize enough for me to think that a GPT3 is thoroughly pushing the boundaries of creative writing.
Ultimately, if I did not have anyone else to write with, AI could be a provocative tool to make a small degree of progress in the scenarios I write. Managing an AI requires a significant amount of human intervention, editing and energy to create a story that is, for me at least, just ok.
The story I ended up with was something I would have never otherwise written. The genre and structure of the story are not my preferred elements, but I appreciated pushing myself out of my comfort zone. This was part of the goal and that was accomplished: pushing yourself to write outside your area is hard to do, but this project helped me do it and reduce anxiety around this challenge. AI did help me experiment. But most importantly it really forced me to question why certain points in my story were not working. I learned a lot about … myself.
The problem is that the story I ended up with is not something I would show to a producer or a network, or even to other colleagues because when you read it, there is a lack of life. It actually reads and feel like a bit of a Hollywood set up, but even then, the touches that make you feel a story is describing life, an encounter with a life, is missing. It took a lot of re-writing and thinking for the AI and I to come up with this scenario. That energy might be better spent if I wrote what I wanted, then showed it to others I trust to get their feedback before starting the process again. Yes, they are in my circle, and I could be encouraging a feedback loop that many creatives suffer from. But you cannot escape the need for humanity in your writing and that comes from inviting and analyzing varying perspectives in your work.
It seems my quest to write something unique was buoyed for a moment by AI: my AI writing partner was free of history. I hoped AI could write more freely, but also have a sense of what was meaningful to a person. I think I asked for the impossible. Good writing is tied to your experience and that comes out in every word. Without it, your script is just going through the motions.
AI is evolving so quickly, so many applications for it are being imagined by super smart people. But now I know AI has a long way to go before it can accomplish what humans can do when they think hard, and write honestly and with vulnerability and put words on paper. Only then can you connect to someone else’s experience. I will definitely be experimenting with GPT4 when it comes out. I am sure it will have some surprises for me. But it is highly unlikely I will take a serious writing project and use an AI model to help me write it out again.