How We Used Artificial Intelligence to Write a Crowdpleaser Horror Script
Will the world’s first feature co-written by AI change the way movies are made?
I have been fascinated by the idea of artificial intelligence since I was a kid. The idea of an AI doing all the work for us and computers “thinking” by themselves just sounds amazing to me. Lately, with the advances of machine learning technology, that idea seems not far from reality. But, can computers be creative?
A few weeks ago a short film called Sunspring came out and it is made by filmmaker Oscar Sharp, but written by an AI that could write screenplays! I was so excited to see this, but after a few minutes into the film, I realized that the story doesn’t make any sense at all.
The AI failed to recognize the deep plot structures that are essential to creative work. Why did it not work? In short, writing a full script is probably too complex of a task for today’s AI. Today's AI is very good at recognizing patterns in a large dataset, but it is hard for AI to be "creative" enough to create good content.
As humans, we may not notice how much information each of us are processing everyday and our ability to extract structure from this information. This happens when we read books, blogs, watch movies, videos, even when you are reading this article right now. When we engage in creative works, we take full advantage of all the data that we have processed in the past. Computers are very bad at doing that, especially if we want computers to write each word, each sentence on their own. There are too many combinations of words, sentences, and thus the problem becomes too complex for computers to solve, at least for today’s AI.
If we use computers to store and process useful information into consumable forms that human creatives can easily understand, what would happen then? Pure magic.
So, is there any way that today’s AI can help us in creative work? The answer is absolutely yes, there is a way. Let’s re-examine our “resources” again. On one hand, we have computers that are very good at storing and calculating massive amounts of information, but they are bad at creating stories in a creative non-rigorous way. On the other hand, we have people, who are good at doing things in irrational creative ways, but when it comes down to memorizing and calculating data, we are not even comparable to the slowest computer sitting in your grandma’s basement. Why not combine both?
If we use computers to store and process useful information into consumable forms that human creatives can easily understand, what would happen then? Pure magic. At Greenlight Essentials, we have been using this methodology to develop the first feature film co-written by AI. How did we do this? We first simplified the task for computers, and told our AI to take advantage of the decades of movie data and find the plot combinations that have high odds to draw in audiences. Then, our creative team worked closely with the insights generated by our AI to turn these plot elements into a creative story. IMPOSSIBLE THINGS is born.
The early signs are showing a huge success! Remember how our objective for AI is to draw in audiences? In a short month, our teaser trailer had over 200k views and 20k followers on Facebook, and the viewers are giving unbelievable positive engagement and comments (By the way, we didn’t tell them it is made using AI; for them it is just another indie film).
We simply asked our AI: what combination of plot elements within the horror genre would have the highest odds to draw in audiences?
How did this happen exactly? What did we do differently than Sunspring? In short, instead of ordering our AI to write each word and sentence, we ordered our AI to work on a more conceptual level: focus on the key plot elements.
For example, one of the plot element that our AI told us to include before a single word was written was that "ghost" and "family relationships" both need to appear in our horror film to make it a success.
It was also able to tell us what to include in our trailer to better match audience taste. Scenes involving "bathtubs" and "pianos" were two of the many that our AI suggested.
This method simplifies the complexity of the problem for computers by quite a bit. Because there are over 1 million English words out there, the combination of these words into sentences is well beyond trillions. Working on the plot level, we were able to shrink down that number to a manageable level. It is still big—we have about 40,000 different plot elements—but it is manageable. Then, we correlated these plot elements to the success of films that shared these elements. After that, we simply asked our AI: what combination of plot elements within the horror genre would have the highest odds to draw in audiences?
Using our big data algorithm, we are also able to find where the prime audiences are for that specific combination of plot—demographics, geographical location, etc. (Some details can be found at the end of this article). This enables us to draw useful information to tell us what to make and who to make it for. We are able to feed the right content into the right hands. Compare that to the traditional creative process, where this information is largely unknown. You will have to guess where your audiences are, and hope that they will like what you created. But using technology, we are able to have answers to these key questions before anything is even made.
We understand that currently not everyone has the access to the AI software that we have, and that is also the reason why that we put IMPOSSIBLE THINGS on Kickstarter, where our backers to have a chance to get their hands on it and use it for their own work. This is the first AI co-written feature length film, and we believe that this project will have its unique place in history and possible change the way films are made.