Training AI with 139,000 Scripts: The Massive Data Set Angering Hollywood Writers
More than 139,000 TV and film scripts were converted into datasets and used to train AI models by Apple, Anthropic, Meta, and Nvidia without the knowledge of their authors, raising fears that their creative work is being used to train machines that could potentially replace them.
In addition to the 39,000 TV and film titles, more than 53,000 additional movies and 83,000 TV episodes were used to train AI, including a vast array of Best Picture nominees and TV episodes of The Simpsons, Seinfeld, Twin Peaks, The Wire, The Sopranos, and Breaking Bad.
The dataset “even includes prewritten ‘live’ dialogue from Golden Globes and Academy Awards broadcasts,” said the Atlantic’s Alex Reisner, who broke the story.
Dialogues as Datasets
The datasets used to train the AI models did not comprise the original scripts, but subtitles extracted, compiled, and uploaded to OpenSubtitles.org. Using subtitles instead of the more technical scripts is more concerning to some critics as subtitles offer a more natural flow of language used in conversation.
Generative AI models trained on well-written dialogue could not only mimic films but generate new ones entirely, which means AI could conceivably compete with the human writers on whose works it trained without their permission. This lack of transparency by AI companies has prompted artists, authors, and publishers to file lawsuits to defend the intellectual property rights of their creative outputs.
“For as long as generative-AI chatbots have been on the internet, Hollywood writers have wondered if their work has been used to train them.” Reisner wrote. “The chatbots are remarkably fluent with movie references, and companies seem to be training them on all available sources.” He created a search tool for the Hollywood AI database to help writers determine whether their work was used.
Response from Scriptwriters
Unhappy to learn about the alleged theft of their work, Hollywood writers responded angrily, as the WGA and SAG-AFTRA unions have contended the use of AI in recent strikes.
“I’m livid,” said David Slack, who wrote the TV show Teen Titans. “I’m completely outraged. It’s disgusting.” Slack discovered 42 scripts credited to him in the AI database. “It’s a huge amount of my work . . . These are things that I poured my heart and soul into.” Other popular writers whose work was used to train AI included Grey’s Anatomy creator Shonda Rhimes, who had 508 episodes in the dataset; American Horror Story creator Ryan Murphy, who had 346; and Matt Groening—who created The Simpsons and Futurama—who had 742 episodes.
AI’s lack of intentionality makes it unable to produce creative works solely on its own—rather, it relies on the work of human authors in a way that many consider plagiarism. However, the issue is even more complex, because in many cases, the studios own the copyrights of the scripts rather than the writers, giving them even less agency for legal recourse or compensation.
Learn more about the complex legal, ethical, and privacy issues surrounding generative AI technology.
The post Training AI with 139,000 Scripts: The Massive Data Set Angering Hollywood Writers appeared first on eWEEK.