Writers, screenwriters, and content creators always seek new ways and tools to inspire and generate ideas. One of these revolutionary tools is AI, which can help come up with a plot for a book, film, AI roleplay, or video. But can a neural network handle this task well? Let’s find out.
History of the Question
At the end of 2015, Elon Musk and Sam Altman launched the OpenAI project, a commercial platform that develops artificial intelligence. The main principle of the company was openness, as the creators did not want AI developments to be concentrated in the hands of a specific country or government. One of the company’s projects was the GPT neural network, which learned to generate texts by predicting each word. Based on OpenAI’s developments, in 2018, Google launched the Bidirectional Encoder Representations from Transformers (BERT) neural network, which first goes through a pre-training stage — a lengthy training process on a considerable number of texts containing billions of words. In addition to dictionary meanings, pre-training includes working with the syntax and context of sentences.
Once a neural network has acquired basic language recognition skills, it can be quickly retrained for various tasks, one of which is language modeling. To solve a problem, a neural network studies billions of words and their arrangement in grammatically correct sentences, after which it can predict the next word in the text. In 2018, OpenAI released GPT-2 (and in 2019, GPT-3), the largest neural network for working with natural languages. It has more than 1.5 billion parameters with which input data is processed and is capable of creating entire pages of coherent text, and for this, it requires only one training stage, unlike BERT.
How AI Generates Plot
In fact, the system is not as complex in architecture as it may seem; it just learns from a huge amount of data. On average, about 600 GB of text is used for training.
A type of neural network called transformers is currently used. The transformer represents each word in the text as a set of numbers. Then, these sets of numbers exchange information with each other in a certain way several dozen times. The neural network can use the results obtained with the help of these calculations for various tasks, including generating texts on multiple topics.
The input data is a “seed” — the beginning of the text that needs to be continued, for example: “What is the answer to the main question of life, the Universe, and everything?” The network will try to continue this text and give an answer. At the same time, it does not express its own opinion; it simply perceives the seed as the beginning of some text already written by a person and tries to predict what will happen next.
Examples of AI Plots in Different Genres
There are already numerous examples of areas where AI is acting as a scriptwriter. Let’s see how successful this is.
Role Playing Games
In the role-playing genre, the plot comes first. Currently, there are AI chats in which you can play with a variety of characters. Thanks to AI’s capabilities, you can have a lifelike roleplay with a vampire, werewolf, Harry Potter, or Captain America. The neural network perfectly answers your questions and adapts to your desires. Many players rate such chats higher than a game with a real person.
Movies
Director Oscar Sharp and Google developer Goodwin Ross created a neural network called Benjamin. It is a type of artificial intelligence used to recognize text. To train Benjamin, Ross Goodwin fed it dozens of sci-fi scripts from the internet, mainly from the 1980s and 1990s.
The first film based on a Benjamin script was Sunspring, a sci-fi short film for the 2016 Sci-Fi London festival.
The script had incoherent dialogue, such as when the main character complained that she was growing a mustache. But after some time, Benjamin learned to create more logical stories and began to create films.
Books
In 2017, American artist Ross Goodwin decided to experiment. He hung a surveillance camera, a GPS module, a microphone, and a watch on a car and drove from New York to New Orleans—along the route of Jack Kerouac, described in the novel On the Road.
He also created his own modification of the LSTM neural network, which read all the information coming from the outside world through sensors and generated text on a roll of receipt tape. In style, the neural network imitated Kerouac. The peculiarity of this AI companion is that it does not convey the entire context but decides which part to use and which to forget.
Goodwin decided not to edit the resulting memoirs, did not even correct grammatical errors, and presented the world with the first attempt at AI in its pure form, as it turned out.
Will AI Replace Screenwriters?
A good story has conflict, characters, emotional baggage, and character development in addition to an interesting plot and dialogue. Artificial intelligence is not yet capable of analyzing such categories and is unlikely to ever be able to. To create a complex character, it is necessary to understand the meaning of the text being created, but no one has learned to do this since the time of neural networks.
It is an open, unsolved problem in machine learning, and at the current stage of technology development, neural networks’ understanding of text is still very superficial. Basically, neural networks only create the appearance of knowledge by remembering which words were most often encountered in which context. Therefore, the machine’s answer coincides with what a person wants from a neural network. This appearance collapses when modern neural networks are asked tricky questions because they are still poor at this.
Conclusion
Nowadays, it is foolish to fear artificial intelligence, but it is also not worth unthinkingly relying on it. Neural networks are a fantastic auxiliary tool that can help with the plot or create an attractive character for a roleplaying game. It is definitely too early to say that AI will replace scriptwriters or other creative professions. Today, technology is not the enemy of creativity.
