The downside of Generative AI involves the danger of copyright infringement. And the threat to the livelihoods of many, if not all, working visual artists and writers. If you need to refresh your memory about Generative AI and/or copyrights, see: A Few Things To Know About AI
We distinguish first between copyrighted images and copyrighted texts.
The Downside: Infringement of Copyrighted Images
Two current lawsuits involve copyright infringement of images. In both cases Stability AI is the defendant.
In one case Getty Images is the plaintiff. And the other is a class-action lawsuit on behalf of visual artists. As I understand the cases, the plaintiffs claim that Stability AI infringes their copyrighted images in two ways. First, Stability AI made unlawful copies of the plaintiffs’ images and used them as inputs to train an image generator. And, second, when users prompt the generator, the output images are infringing derivative works.
Call it a double downside.
Although I am sympathetic to the plaintiffs, their battle may be uphill. For discussion of potential problems in their claims, see: Pamela Samuelson “Generative AI Meets Copyright” Science, July 14, 2023.
Moving now to the current actors’ strike in Hollywood. We all know the big actors who make the big bucks. But we have no reason to know the “middle-class” actors who play minor roles. Or the background actors, aka “extras.”
The trade association representing film and television producers recently put a proposal before the actors’ union. They offered to pay background actors to work on set, say, for day. Thereafter the studio would own the actors’ likenesses and movements which would then be digitized into scenes where necessary. In other words, human background actors would no longer really exist.
Needless to say, the actors’ union shot down the proposal. However, I now see the prospect of films with real actors/digitized extras looming on the horizon.
The Downside: Infringement of Copyrighted Texts
Here the problem involves large language models (LLMs) like Chat GPT developed by Open AI. Like other kinds of Generative AI, LLMs train on existing masses of data, in this case, written texts. Some of which may still be under copyright. For example, let’s say that AI has ingested all the screenplays in the world. A click of a button could then produce a new screenplay.
Now, let’s return to Hollywood and the writers’ strike.
Last week Joseph Gordon-Levitt pointed out in an op-ed that if AI uses your work, it should pay you. He also pointed out two wrinkles that would arise when AI eventually generates an entire feature film.
One. It’s not just the actors and writers who should receive compensation. It’s everyone: camera operators, costume designers, sound mixers, etc. He acknowledged the difficulty of finding a fair way to pay them all.
Two. There’s an even bigger problem.
Q: Who owns the copyrights to films?
A: The big studios.
Here’s Gordon-Levitt’s take:
“When I do a movie, and I sign my contract with a movie studio, I agree that the studio will own the copyright to the movie. Which feels fair and non-threatening. The studio paid to make the movie, so it should get to monetize the movie however it wants…. But if I had known that … I would then be allowing the studio to use that intellectual property as training data for an AI that would put me out of a job forever, I would never have signed that contract.”
In the future big studios might not have to pay humans for their labor. They already own the benefits of millions of hours of that labor.
The Downside: No Resolution in Sight
Adam Conover provides a fascinating and passionate discussion of the Hollywood strikes with Jon Favreau. Conover provides rays of hope that writers in the entertainment industry might continue to make a living.
However, the landscape is fraught.
The Downside: Final Note
In 2005 the Authors Guild sued Google, Inc. for copyright infringement when Google Books digitized millions of books from research library collections without permission. The suit ended in 2015 in Google’s favor, citing fair use. It’s nice to know that Google gave digitized copies to the libraries that supplied the books so that the libraries could lend them out in accord with copyright law.
The case of LLMs discussed here is different. The books/texts are not ends in themselves. They are inputs, means to a different end.
See also: The Upside of Generative AI
This post was written by Julie Tetel Andresen