A hundred years ago, the word “robot” appeared in a play by the Czech writer Karel Čapek, which tells the story of workers in artificial factories created to serve people. Now, in a metanarrative twist, AI has written the play itself.
“It’s kind of a futuristic little prince,” explains playwright David Kosnyak, who oversaw the script. As in the classic French children’s book, the 60-minute production tells the story of the journey of a character (this time a robot). He goes out into the world to learn about society, human emotions, and even death. The play itself is called “AI: When a Robot Writes a Play.”
The script was created by the widely available artificial intelligence (AI) system GPT-2. Created by Elon Musk’s OpenAI, this “robot” is a computer model designed to generate text by filtering out the vast storage of information available on the Internet.
Until now, this technology has been used to write fake news stories, stories and poems. The play is the first theatrical production of GPT-2, the group says.
This is how it works. First, a person enters a hint into the program, a clue to start the play. Researchers at Charles University in Prague began with two sentences of dialogue in which one or two characters discuss human feelings and experiences. The program then uses the hint information and generates up to 1000 words of additional text.
The result is far from William Shakespeare. After a few sentences, the program starts writing things that sometimes don’t fit the logical storyline, or statements that contradict other passages of text. For example, the AI sometimes forgot that the main character was a robot, not a human. “Sometimes, in the process of dialogue, a man turned into a woman,” explains Charles University computer linguist Rudolf Rosa, who began working on the project two years ago.
This is because the program doesn’t really know the meaning of the sentences, adds Chad DeChan, an artificial intelligence expert at Columbia University. “It just connects words that are likely to be used together, one after the other.”
As this continues, more and more room for nonsense appears.
To prevent this from happening, the team did not allow GPT-2 to write the entire play at once. Instead, the researchers split the show into eight scenes of less than 5 minutes each. In addition, scientists sometimes changed the text, for example, changing passages in which the AI changed the gender of the character from line to line, or repeating the original text clue until the program produced reasonable prose. However, 90% of the final script remained untouched, and only 10% had human intervention.
Completely independent writing of the play will be possible in 15 years, experts explain. Then the technology became good enough to create a complex and coherent text, like a theatrical play, from start to finish. However, the experiment is still a good way to show audiences what artificial intelligence is capable of, scientists conclude.