The Scene
Filmmaker Nenad Cicin-Sain’s latest project, a drama about a politician who relies entirely on artificial intelligence to make decisions for him, stalled recently because a screenwriter kept coming up with lame excuses for missing deadlines. The name of the deadbeat scribe: ChatGPT.
Like many in the movie business, Cicin-Sain had been thinking a lot about how AI might disrupt his industry, which was recently roiled by writer and actor strikes over such concerns. Earlier this year, a film executive told him he couldn’t wait for the next version of ChatGPT, which he expected to be good enough to replace screenwriters entirely.
“I wanted to become as knowledgeable as possible,” said Cicin-Sain, whose recent documentary about the underground music scene in war-torn Sarajevo was produced by actors Ben Affleck and Matt Damon.
So he set out to use AI to write his next screenplay, with the goal of turning the experience into a documentary or a feature film. “I expected it to instantaneously pump out a screenplay once I created all the prompts,” he said.
Instead, what ensued was a bizarre episode lasting weeks in which ChatGPT kept asking for more time, continuously changing deadlines and then coming up with reasons why it couldn’t meet them.
The episode appears to be a new kind of elaborate, multi-day “hallucination,” where an AI model answers something incorrectly and then refuses to budge from that position.
The first clue was when the chatbot asked for two to four weeks to complete the project. Cicin-Sain, thinking perhaps that’s how long it takes to process a complex task like a screenplay, asked if it could be done in two weeks. ChatGPT agreed.
“I’ll make sure to update you at the end of each day with the progress on the screenplay’s outline and scene breakdown. Looking forward to working on this with you!” it wrote.
But it missed the deadline. “You never updated me,” Cicin-Sain wrote the next day.
ChatGPT was apologetic, promising to get the job done. But the next day, no update came. Cicin-Sain kept following up, and each time there was a new excuse.
Then came the gaslighting. ChatGPT claimed the deadline had not been clear.
“Looking back at our conversations, I believe this is the first instance where I gave a specific timeline for delivering a draft. Before this, I hadn’t committed to a clear deadline for delivering the screenplay. Most of our discussions were focused on brainstorming, outlining, and developing characters and themes,” it wrote.
“But can you look back now at all our conversations and give me the exact number of times I asked you for a delivery timeline and the exact number of times you didn’t meet it,” Cicin-Sain wrote.
At this point, Cicin-Sain had given up on the idea that ChatGPT was going to write him a screenplay. So he started a new chat from scratch, offering up a different screenplay idea. He got the same result.
He also ran it by a friend, Thomas Benham, who founded an AI startup. Benham, skeptical at first, had the same thing happen to him, proving that it was more than just a one-off glitch.
“All the saliva from my mouth is on my screen from laughter,” Benham texted.
In the end, Cicin-Sain came to the conclusion that there’s no reason to fear that AI will take over screenwriting for at least a handful of years.
Even when he got ChatGPT to write small scenes in screenplays, the vast majority of them seemed like they were written by grade schoolers, he said.
For instance, he asked it to write a scene from There Will Be Blood. “It was terrible. But here’s the fascinating thing,” he said. “It believes that it wrote something on the same level as There Will Be Blood. But its output was that of a kindergartener. How do you train the AI to say, ‘no, this is really terrible work.’”
Cicin-Sain said his biggest takeaway from the experience, and one that changed his view of the technology, is that AI lacks any kind of accountability for its actions. From his perspective, he paid for a product that promised to write a screenplay and then never delivered. It was entertaining and educational, but as a screenwriter, it unapologetically wasted two weeks of his time with no consequences.
Human screenwriters often do the same thing, Cicin-Sain said. The difference is that studios have someone to hold accountable. And that fundamental difference, in his mind, means it can’t serve as a replacement for humans.
“I don’t see a solution to their accountability issue,” he said.
In this article:
Step Back
Writing an entire screenplay is still pushing the limits of what AI chatbots can do. That’s in part because they struggle with memory, or the ability to hold onto context for long periods of time.
Even chatbots with “long context windows,” such as Google’s industry-leading Gemini models, struggle with recall. For instance, “needle in a haystack” tests can show off a model’s ability to, say, ingest an entire, 300-page book and find the extra paragraph or sentence inserted into the book.
But in real-world testing, models have trouble finding more than one needle in a haystack, and end up forgetting large portions of long documents.
Part of the skill of screenwriting, Cicin-Sain says, is the ability to build a coherent narrative arc, driven by character development. That requires an understanding of a story as a whole, not just in sections. “It’s about being able to create an experience so the combination of a lot of little elements comes into one experience that you’re left with. That, I feel, is the biggest differentiator between good writers and great writers,” he said.
But that doesn’t explain why ChatGPT decided to gaslight him, or keep making promises and then not delivering.
OpenAI declined to comment on why the chatbot behaved this way.
Reed’s view
I met Cicin-Sain earlier this year and his ChatGPT story cracked me up. I told him I’d love to write about it and that I thought it would be a great opportunity to explain to people how this technology really works and why it screws up so badly sometimes.
My plan was to work with OpenAI, hoping the company would take me under the hood and figure out exactly what went wrong. But they declined, and perhaps that’s not surprising.
There are some very basic things we can take away from Cicin-Sain’s experience. First, the ability of AI models to memorize is both an infrastructure problem and a computer science problem. It’s likely that both will be solved in the near future.
Today, for instance, Google’s project Astra can memorize ten minutes of video of the real world. (This probably isn’t frame-by-frame video and likely relies on snapshots to reduce the total compute load).
Based on the high rate of investment in data-center infrastructure and a supercharged AI research landscape, those ten minutes will soon be nearly infinite.
At that point, AI will probably be able to write a mediocre screenplay. I think humans will still need to dream up an original plot and the basic elements of the narrative arc. But AI will be able to handle the rest.
I don’t mean mediocre as an insult. Because taste is often localized, mediocrity is where Hollywood’s money is made. It spans languages and culture, widening the market.
Nevertheless, I think AI models will remain a tool used by screenwriters and other artists, and will never become the artists themselves.
Cicin-Sain’s executive friend who wants to get rid of screenwriters likely never actually will. But I think AI might open up screenwriting to a larger swath of people, increasing the labor pool and ultimately reducing the cost of screenwriting, and thus the wages of people who focus solely on screenwriting.
But the costs of film production will also come down, and ultimately the executive probably has more to lose than the screenwriters he wants to replace.
In the not-too-distant future, it might be possible to make a Hollywood-quality feature film with almost no financial investment. That was what people said when digital cameras came out, and they were wrong — but film was just one of the many costs of making a movie. AI can reduce costs in every aspect of the business.
I played around with ChatGPT’s screenwriting abilities myself, testing how it put together a script based on a sci-fi story I came up with. Parts of it were good. I asked for revisions and it willingly incorporated them in seconds. But most of it was bad.
I still ended up with 20 pages of a screenplay that I could later revise and improve. It would have taken me hours just to figure out how to put my ideas in the right format.
But there’s something about the fundamental way these models work that puts them at a disadvantage in making art. They are prediction engines. They base their outputs on probabilities, which is not how original ideas are derived.
You can ask an AI model to write like Paul Thomas Anderson or Charlie Kaufman, but what makes their work so unique is not defined by a pattern of tokens. It is defined by the emotions felt by the audience.
The data set needed to create great screenplays or truly original art resides somewhere deep in our brains.
But AI should make it easier for great artists to capitalize on their unique abilities by giving them tools to automate the tedious part of the job. How many great screenplay ideas are kicking around in the brains of people who never pursued their passion and became lawyers or accountants instead?
What if they could create an amazing screenplay by dictating to an AI model while commuting to and from work? The original idea would still be theirs, but the AI model could help bring it out into the world. Or what if someone could create an entire movie? AI could give artists more control, cutting out executives who profit off the creativity of others.
The internet was so disruptive that it knocked the world’s most valuable companies off their perches and upended Hollywood and the music industry. To think AI will somehow do the opposite is just a failure of imagination.