One of the less-than-gracious responses to the ongoing WGA strike is that writers could simply be replaced by scripts created by artificial intelligence (AI). There are many problems with this idea, one of which is the obvious: could a program like ChatGPT actually generate a decent sci-fi fantasy script?
We put the idea to the test, but it’s worth first noting that this scenario has real basis for concern. Predictive, generative writing programs, often referred to as AI, are a consideration for the Writers Guild of America and the exclusion of AI as a core writing software is included in the writers’ pattern of demands. The WGA wants to limit its use in writer’s rooms and make sure that any language generation models are used as a tool, rather than to create a final product, or even a developmental product. Another demand from the WGA is that AI will not be used to “rewrite” material. It also wants to guarantee that AI will not be trained on material written by guild members or covered under the minimum basic agreement (MBA). This last point might be the stickler, as public large language models, which are trained on a vast corpus of work, are not exactly forthcoming about what’s included in their library.
Even though ChatGPT—by far the most popular and recognizable generative text bot—is making huge jumps in development between its iterations (we’re currently on version four), it’s still a far cry from usable within the creative sector on its own. Or is it? Is ChatGPT, or any generative writing model, capable of replacing writers? Before we get into exactly what I did in an attempt to get ChatGPT to generate a script, let’s dive into the issues with AI generative models at a universal level.
First, a brief overview of how AI chatbots work. ChatGPT, for example, is what is commonly called a Large Language Model, because it is “trained” on a dataset (a corpus of work, texts, articles, etc.) and through mathematical formulas meant to imitate neural pathways, using an algorithmic average to produce the next word in a sentence. Using its artificial neural network, ChatGPT generates the next word that is most likely to appear based on previous words, but it does not know, or even have the ability to understand meaning—it only understands a word’s position as a datapoint within a sentence. ChatGPT has no intent whatsoever. To call this generation writing is a stretch of the imagination, as writing requires, at the very least, intention.
Currently there is also no way to copyright AI-generated text, another sticking point for people across the entire industry. Many involved in the writing process have brought this up as an issue with allowing AI writing into writers’ rooms. It seems that studio executives feel there might be a workaround somewhere—but there is no reassurance for writers that AI wouldn’t get credit or even be the sole credit on a piece, even if writers were forced to work from an AI-generated idea.
One of the problems I have with ChatGPT is that there are ethical concerns at nearly every level of its use and production. Not only are there issues with the library of text used to train the software, there are also ecological and labor problems at the core of the model’s use and development. According to Science Focus, the ChatGPT AI model “was trained using text databases from the internet.” The corpus includes data from books, Wikipedia, online articles, and “other pieces of writing on the internet. To be even more exact, 300 billion words were fed into the system.” To date, there has been no confirmation that OpenAI—the developers who programmed ChatGPT—have been given universal consent for these pieces of writing to be fed into their machine.
Another issue I had was that I assumed that by using ChatGPT I would be helping train ChatGPT. However, there is no indication that ChatGPT uses the text that it creates to “re-train” itself. It has a massive lexicon that is already slotted into its programming. Using ChatGPT does not make it better, as it cannot truly “understand” why you are asking it questions, correcting its output, or even “refining” the output. It is simply operating, not learning or developing a deeper understanding of the artificial neural pathways that have been coded into it. So that’s something, I suppose. Be reassured: ChatGPT does not eat itself. It does not think. It does not create. It just outputs. That’s a reassuring and frankly kind of gross way for anyone to imagine writing.
There are also ecological impacts for using the chatbot—it uses a lot of water to help cool down its servers, and, according to Gizmodo’s own reporting, “an average user’s conversational exchange with ChatGPT basically amounts to dumping a large bottle of fresh water out on the ground, according to a new study.” Not ideal. Additionally, Time recently discovered that OpenAI—the publishers of ChatGPT—took advantage of cheap labor in Kenya in the early stages of ChatGPT’s inception in order to keep costs down before they had funding secured. Another yikes. While a lot of the messaging around ChatGPT is centered around the elimination of labor, it’s clear that labor is an intrinsic part of training and keeping up the artificial neural network. Exploited workers programming the AI to create cohesive sentences, follow grammar rules, and even associate words with data input are are on the front lines of finding ways to develop a “smarter” AI.
While a core message around the use of ChatGPT and many other generative text models is that it will reduce the amount of work done by writers, Time’s reporting has made apparent that there is a huge amount of invisible labor being done behind the screens, often by underpaid or exploited workers. It is simply not ethical to use ChatGPT in any way, as its problems far outweigh its benefits, especially at this stage in its development.
Knowing all that, I’m still interested in seeing how it performs. While I am not looking to profit off any work that this test generates, I think it’s important to push its capabilities and see if it’s even at a level where it could potentially replace writers. My first entry: “I want to write a horror/fantasy story about demons possessing cars.” I’ve been on a horror kick recently, and Fast X is coming out soon. ChatGPT responded with a synopsis and then a breakdown of seven chapters and an epilogue. The synopsis read:
“In a quiet, unsuspecting town, a series of bizarre car accidents has left the residents bewildered and terrified. Little do they know, a sinister force is at play—a horde of demons has found a way to possess automobiles, turning innocent vehicles into instruments of malevolence. As the demonic presence grows stronger, a small group of courageous individuals must uncover the truth behind the possessed cars and find a way to stop the terror before the entire town is consumed by darkness.”
Less of a synopsis, more of a first draft of back-cover copy. This is the idea that someone has before they figure out what’s really happening in this story; it’s the idea before the idea is fully fleshed out. The chapter breakdown did follow basic narrative structures, but it was also… not all there. There were no real characteristics given to the main three characters (Emily, Natalie, and Marcus), there were no motivations, there was no depth. It was a shallow, empty husk of a story, the work of a model that understands the formula of storytelling but doesn’t quite know how to make it interesting, how to make it a story that means something beyond surface level plot progression.
I took the first chapter synopsis and asked ChatGPT to expand it. It’s a good enough opening shot; Emily is a mechanic working on her cars and she begins to hear the whisper of demonic forces in the engine she’s working on. The result was about 450 words that lacked style or even consideration. It was written from an omniscient point of view, something that has largely gone out of style in contemporary writing (one modern exception is Cat Rambo’s You Sexy Thing, a truly hilarious and wonderful sci-fi novel that deliberately engages with the POV). Close omniscient can lend itself very well to synopses—it’s a top-down view of literally everything going on in a scene, but it adds distance and removes the audience from the action if used incorrectly in the actual line by line writing. ChatGPT’s generation didn’t include a lot of emotion or presence either. It didn’t just lack style, it showcased a style vacuum. AI has an image of the language needed to develop a story, but no understanding of how to use it. It is a tool unable to use the tools that were used to build it.
Descriptions in the generated text are disembodied and removed from character. Some lines, as examples: “The atmosphere grows heavy with an unexplained sense of foreboding.” “A chill runs down Emily’s back, and her heart quickens its pace.” “As she draws closer, the whispers become clearer, weaving a chilling tale of dark desires and forgotten sins.” All of these are outlines of sentences, but they are not specific, they are not imminent. They require no context, they do not flow from one sentence to the next, they simply exist. It’s a poor way to write—but, you know what, we’re interested in a script. While this scene has Emily on her own, I ask ChatGPT to add in a phone call; “Let’s take the following scene and add dialogue. Have Emily chat with Marcus on speakerphone.” This was a chance for me to see if ChatGPT could take the established scene that shows an outline of something creepy, weird, and foreboding, and generate appropriate dialogue. I wasn’t going to hold my breath for characterizatio Source: Gizmodo