So, awhile ago, there was a 2015 TV show called SCREAM on MTV, about a slasher hunting high school students, and a complex mythology. Inspired by the feature films, the TV series was less a comedic, metatextual take on horror movies and instead explored the anxieties and trauma of school shootings through the analogous fiction of a slasher killing teenagers (played by early twentysomething actors).
Despite a massive launch, MTV's inability to market shows effectively and two short seasons (only 24 episodes in total) and about 300,000 viewers by Season 2 meant SCREAM never really had much of a chance. The show was beautifully shot and wonderfully performed especially with standouts Willa Fitzgerald, Bex Taylor-Klaus, John Karna and the brilliant voice actor Mike Vaughn as the Lakewood Slasher.
The show ended on a pleasant note but with a lot of loose ends. Inexplicably, MTV renewed the show for a third season, but then fired the entire cast and crew and started a new storyline with new characters that had no connection to the first two years, and then found themselves with a 'third' season that was unmarketable: if aired as a third season, the viewers would wonder if this new storyline would be abandoned like the oriignal and not bother. MTV ended up not even airing this third season, selling the six episodes to an affiliate two years later that burned it off over three nights.
Awhile ago, a YouTube vlogger who makes SCREAM commentary videos (footage of the films and TV show, his voiceover) declared he was writing a novel to wrap up the MTV SCREAM series, to be released in monthly chapters over his Patreon. After he finished this book, LAKEWOOD: A SCREAM STORY, I subscribed for a month and downloaded the whole ebook and... it is awful. Tenses jump from past to present. Commas and periods are missing. Names are mis-spelled. Characters do nonsensical things like knock on the door of the home in which they live instead of using their keys. At one point, a woman in a burning car decides to make a phone call rather than try to open the door and get out of the car. A police officer asks what he should do if he finds out the location of an active serial killer.
Out of curiosity, I tried putting this barely-readable volume through Google Gemini Flash 2.0, half-chapter by half-chapter to see what AI could do. AI could fix the tenses and punctuation and get them consistent and also spell characters' names the same way throughout. AI could even identify incredibly obvious errors like a woman knocking on the door to her own house. However, when it came to more complex issues, AI could be very hit and miss.
One impressive achievement for AI: the original novel had a sequence where a police detective learns that the serial killer broke into the police station and marked an abandoned train station on a map with a knife. For some reason, our detective in charge sends two uniforms to poke around instead of having the entire train station cordoned, secured and examined by a forensics team. AI was able to create a new chapter and address this, passing all the uniformed cops' dialogue to the detective character, adding in a forensics team, and following proper police procedure.
However, for scenes that consisted of bland exposition, AI could only clean up the grammar; it couldn't infuse them with deeper meaning. When scenes and sequences lacked a thematic purpose or point and consisted of characters reiterating the plot, AI could only clarify and refine what was there. It couldn't come up with jokes or deeper meaning or any meaningful statement on anything.
AI could broaden the descriptions with sensory details, describing carpets and lights and skies and coffee cups in excessive detail. However, these excessive details were often reused in other scenes; there seemed to be a narrow algorithm that often generated the same flickering yellow lights and cracked concrete whether describing a bowling alley or a house or a hospital. And none of it was purposeful in creating a lighthearted or frightening tone; it was just verbiage.
Reused descriptions were also very common in very uniform patterns of describing speech and body language.
"That's a possibility," she said thoughtfully, her tone pensive.
"It doesn't have to be that way," he said hopefully, his tone optimistic.
He emerged from the car, his movements cautious and concerned.
She stepped out of the office, her movements firm and unyielding.
The AI consistently ended each verb for speech wtih an adjective and then added a needless dependent clause, or repeatedly used the noun "movements" and then attached two adjectives. The AI consistently used phrases like "the air was thick with" various smells and "her furrowed brow", often multiple times in a paragraph.
However, I found it had its uses. There came a point when just touching up the original prose was no longer effective and it was necessary to write new scenes. But I didn't have time to do it, so I would write a bullet point list of events, actions and key dialogue, and AI would generate a prose version of my outline.
This prose would have all of the problems mentioned above, and either underdescribe or overdescribe actions, sensory details, settings tones of voice. However, it was a lot easier and faster to rewrite and replace poor but present prose than to write it on a blank page. The AI would sometimes make very cursory nods to the specific themes I tried to indicate in my outlines, or miss them entirely, but it still provided a framework for me to add it in whether through replacing or altering the descriptions and dialogue. At times, AI would provide only a two sentence summary of what I'd imagined as a more complex exchange; I could ask it to expand, but even that expansion would be repetitive and uninsightful until I rewrote it.
AI could help me get through a lot of the problems that often made me feel blocked in writing. At one point, I needed to know what police procedure would be for specific evidence. The AI, while a clumsy wordsmith, could describe the appropriate procedure and even put it in some poorly worded paragraphs for me to revise into my own voice. In another scene, I needed information that a medical examiner could plausibly provide from examining a body. AI could generate the jargon, although I stripped it back to the bare minimum.
At another point, I was stuck on how the characters could, outside a building, quickly get to the roof without going in. AI suggested some options (HVAC maintenance, skylight access) and a process by which the characters could ascend. When the characters were caught in a deadly situation and I couldn't think of a way out, AI could offer some ideas. As a result, problems that would normally stop me for days or weeks became something I could work through in an hour. Rewriting this person's fanfic would have taken a year and it wasn't worth that kind of time; AI allowed me to do it over the course of two weeks.
Slider_Quinn21 has talked about struggling with writing and I wonder if he might try AI. Not in these sense of generating a novel. But in terms of writing a bullet point outline for each scene, having AI generate a very poor but present version of the prose that he could then use as a raw draft to rewrite into a first draft.
Of course, there is quite a bit of find and replace needed as I went to work stripping out 172 variations on, "she stepped out of the car, her movements taut and tense" or "her voice troubled" or "the air was thick" or "the smell of stale beer" or "her face etched with worry" and all the highly repetitive writing tics.
I am not sure AI will get 'better'. These repetitive flaws might be amended, but it can't seem to address theme, meaning, insight or purpose unless a human provides those things first. Upon assembling the makeshift novel after a week, I determined that this SCREAM story was effectively a story about ableism and how society often stigmatizes people with disabilities. The AI could not see that, could not refine it, could not hone in on it independently.
When I pointed it out, AI could offer research and statistic and analysis and even refer to the theme in new content, but it couldn't create drama to explore and present it without me first providing an outline and key points for dialogue, and even when it produced those scenes, it was in repetitive prose that lacked intent and purpose and deliberate decision. It would emphasize scenic details that weren't relevant to the theme of stigmatization; it would focus on body language that was irrelevant to the scene being about societal disdain for disabilities.
I doubt AI writing will ever be good. It might have fewer repetitions, it might generate more verbiage pertaining to theme, but I suspect it will always lack artistic intent on its own. It doesn't create art. It generates words and presents data. I think more advanced language generation will be like modern CGI in movies: CGI has become more elaborate and more complex -- but it still looks like an advanced cartoon, you can always tell it's computer generated, and it never looks that real.
The best way to see AI writing might be the way CG can produce animatics. The AI writing is not a final product, but rather reference material to help a human write the actual story, just as Blumhouse will often use CG to produce a cheap animated version of a movie before deciding whether or not to spend the money on filming it with real actors and sets and locations.
AI will not make a bad writer a good one, but it can fix grammar and spelling and provide a subpar working draft, effectively a very rough and artless version of what a writer might imagine, and spare them the trouble of getting those first words down on the blank page.