Using (and not using) AI for Making Movies (10)

Thoughts from a Writer

Screenwriters are afraid that AI is going to take their jobs.¹ The fear is real and the concern has some validity. How do we know? Because AI can produce screenplays. And what do I mean by “some validity”? It means … it’s hard to measure.

Measurement is necessary. It tells us the extent to which things can be done. It allows us to understand the depth and breadth of a problem. It allows falsifiable claims to be made about new phenomena that we don’t have much experience with it. And that experience matters when we’re talking about large numbers of people because there is no intuition yet; no shared observations; no general sense of the scale of a thing.

All of that might sound a bit fluffy to hard scientists who love good, solid, integral measure — 1, 2, 5, 12. But ordinal scales can help: A bit, some, a lot, tons.

Q: How much trouble are we in?

A: A lot.

That’s good to know. It might not be as accurate as C3PO’s calculated odds for successfully navigating an asteroid belt (3720:1), but … “not good” is news I can use. (Not that Han would want to hear it).

Another way to come to an understanding of something that hasn’t been measured one way or another (in this case, via integral measures or ordinal ones), is through practice. Practice is the act of engaging in something, and that “something” generally has a beginning, a middle, and an end. You are welcome to rinse and repeat, but then you’re engaging in the practice a second time.

Using AI to assist in writing a screenplay is a practice. To learn anything from it, you need someone who writes screenplays and is willing and able to reflect on practice to share some details of that experience so that we (people in the writing industry) can move beyond fear and metaphors and at least engage thoughtfully with reality — whether or not the findings are entirely generalizable.

So let’s pause on that term “generalizable” for a second. We all want to make claims about the world, but if we make claims that extend beyond the basis of our confidence, we might be making things worse rather than better. For example, many people still think a glass of red wine every day is good for your health based on a misunderstanding of popularized in 1991 in a 60 Minutes segment called “The French Paradox.” As much as I too would like this to be true, it isn’t.²

So what is generalizable and what is not? A thoughts on how to read and learn from my experience of turning my novel The Curse of Pietro Houdini into a screenplay:

A few thoughts:

• There are standard industry conventions for the creation and presentation of screenplays for both TV and feature films. Most people use FinalDraft to automate the formatting. So the rules and expectations of the industry and fixed and allow us to use AI within the confines of those rules.

• I’m using ChatGPT 4.0 in English. I have no idea if it works the same in, say, Korean (± 80 million speakers) or Hebrew (±9 million people) or Icelandic (±315,000 worldwide).

• What’s true now will not be true later.

• How much later? No idea.

• AI has never once — not in the entire process of writing the screenplay — delivered a single line of dialogue, let alone an entire scene that met my standards. There are no AI-generated scenes or lines in my screenplay.

• However: It sometimes delivered descriptions of locations that were somewhat useful or — if not entirely “drag and drop” — at least inspired me to consider something I had not. For example, I may not have considered a line of palm trees along the side of a road, or the prevalence of shredded truck tires on a highway. Whether or not I chose to use, or follow up, on these “observations” depended on my thoughts and mood and tastes and whatever other magic goes into creative decisionmaking. These descriptions became more helpful when the subject matter required rich vocabulary to explain the environment, like the inside of the 17th century basilica (e.g. nave, apse, transept, altarpiece, balustrade).

• Attempts to get ChatGPT to generate entire scenes from my novel all failed. It asked me to keep waiting and delivered the goods.

• I think, eventually, it will be able to but it will take a long time to do it “artfully.” A screenplay takes approximately 1.5 hours to read. A novel takes ten hours. So a screenplay is not a transcription of a novel but a distillation of story which requires that new material be written to achieve what the novel achieved using other forms. The book might have used interior thoughts to explain a characters motivation whereas on the screen that will need to be handled by dialogue or action or acting. Lighting and music and set design can and will assist with all of it but the feelings of the character will need to be — as it were — enacted. AI cannot yet manage all that.

• But here’s where it starts getting interesting: We don’t all have the same standards and soon, AI will meet minimum standards for screenwriting as it is already meeting standards in music production. A survey by Ditto Music found that nearly 60% of artists incorporate AI into their music projects and the music industry is now entirely engaged with AI as both a boon and a possible bust for musicians.

Automated tasks and rule-based tasks are handled extremely well by AI. I have regularly dropped pages of my screenplay into ChatGPT and asked it to improve the formatting and recommend alternative descriptions (but not dialogue). It catches clear errors (missing punctuation, grammatical errors, spelling, and other copyediting concerns) and makes helpful suggestions on how to cut things down. But trusting it is a mistake because often there is a logic or thought behind something unconventional that works because I understand how it fits into either the screenplay as a whole, or else aligns to a system of thought I have adopted.

Take the latter case: It is conventional in a screenplay to never describe anything that cannot be seen, heard, or acted. So anything that constitutes “backstory” will not be in a description. Screenwriters are told not to write, for example, “the woman in the truck handed the clothing of her missing son to the teenage girl and smiled sadly.” There is no way for a director to communicate that the clothing belongs to a missing son. Rather, the screenplay should be written so that the knowledge of the missing son is inferred through the actions of the actors. That line might therefore be written as, “the woman in the truck lifted a child’s sweater from a suitcase at her feet, looked it for a moment, smelled it with her eyes closed, looked at it again, and then handed it to the teenage girl.” Maybe that does it. If not, the writer’s room can discuss if there is a better solution or whether that moment is needed at all.

But: There are other philosophies here. Sometimes the writer is the director, and the director therefore always knows what the writer means, and the director wants a screenplay that helps get the heart to the scene rather than lay out all the shots. Other times, when a screenplay is going to be shopped around to studios or perhaps go out to actors, the real objective of the screenplay is “to be read.” In other words, it errors on the side of being engaging for the reader so they will get all the way through it and become charmed by the possibilities of it becoming a movie or whatever. Later, once the commitments are made, the screenplay can be edited so it becomes more formal and is useful to a director as a “shooting script.”

AI could, I suppose, be told all about one’s preferences and thoughts and goals and systems could be established to turn it into a tool that is better refined for the writer’s needs. In time I expect this will happen, but all that brings us back to human-machine interaction rather than the hand-off model where AI takes over and steals our jobs. In this case, the writer needs to work on AI to ensure it serves the needed functions later rather than simply works with AI as though it is a single-minded hammer that knows exactly how to perform in all cases (yes, hammers do know what they’re doing).

Industry uses

There are uses for AI that are not about the writing, though, and this is where it has prove fun at worst and quite interesting at most. Here are a few of those uses:

Casting

Who should play the staring role of Pietro Houdini? I turned the completed screenplay into a PDF, uploaded it to AI, and asked that very question. The director and I were very surprised to find that about 75% of the men on our list (draw up at the bar over a beer and a lot of talk and laughter and BS) overlapped with the AI list. Sure, some of the AI suggests were terrible and the actors were ten or fifteen years too young or old in some cases, but it was affirming to see how clearly those 75% “popped” and how the logic provided by AI was similar to our own reasoning. It also produced that list in about three seconds whereas we bounced those names around for hours.

But there’s a dark underbelly to that. I can’t prove this, but I (very) strongly suspect that the AI system is going to have a tremendous bias towards the well-known, the well-respected, the well-heeled actor and not the “best person for the role.”

Now: For a title role by a 55-65 year old actor who will speak almost 45% of the lines in the movie and who is “bankable”, the AI bias mostly works in our favor. But for smaller roles where performance is key, actors who have not had a big break or might have been unfairly typecast, or who would otherwise not be expected to fit that role will most likely be overlooked. AI will work against them. This is a machine that sends the expected and familiar back to you. If that’s what you want, great. If not … we’re not there yet.

Comps

A “comp” is a comparison but the word comparison is too long for people in LA to use, so they talk about comps. NYC suffers the same affliction when it comes to novels. Trust me, I know these things.

When I uploaded the screenplay into ChatGPT, it did an excellent job creating comps. It produces a tight list and each entry is followed by “why” which describes the reasoning (yes, I know it doesn’t reason …) and the “common elements.”

One such comp is The Book Thief. This is how AI presents it:

“The Book Thief (2013)

• Why: Both narratives explore the transformative power of culture (art, books) in the midst of war. The young protagonist and mentor dynamic, as well as the exploration of identity and loss, align closely with The Curse of Pietro Houdini.

• Common Elements: WWII backdrop, a child’s resilience, and the significance of cultural preservation.”

After listing five others in the same format, it then creates a summary analysis:

“Unique Selling Points for The Curse of Pietro Houdini:

• Historical Accuracy with Personal Drama: The screenplay fuses rich, authentic historical settings (e.g., Montecassino) with personal stakes, such as Pietro’s quest for redemption and Massimo’s transformation.

• Multifaceted Tone: A blend of humor, action, and poignancy, much like Tarantino or Waititi’s works.

• Complex Protagonist: Pietro Houdini’s enigmatic mix of charm, intellect, and vulnerability makes him a captivating lead, comparable to characters in The Monuments Men or Inglourious Basterds.

• Art & Identity: The thematic focus on art as both a symbol of identity and a survival mechanism offers a fresh angle for war dramas.

This project could appeal to studios and directors aiming for prestige films with commercial and critical appeal, especially given the screenplay’s balance of spectacle, narrative depth, and unique character dynamics. Would you like further assistance refining this for a pitch or expanding on any comparisons?”

It even spells out the word “comparison.” Isn’t that nice?

— DBM, 27 January, 2025

1. https://www.brookings.edu/articles/hollywood-writers-went-on-strike-to-protect-their-livelihoods-from-generative-ai-their-remarkable-victory-matters-for-all-workers/#:~:text=An%20animating%20concern%20for%20many,get%20to%20be%20a%20showrunner.

2. https://www.health.harvard.edu/heart-health/want-a-healthier-heart-seriously-consider-skipping-the-drinks?utm_source=chatgpt.com