Silicon Valley’s Copyright Crisis: The $50 Billion Lawsuits Threatening Generative AI

The storyline of the most recent Mission: Impossible movie contains a peculiar irony. To prevent a rogue AI from taking over the world’s systems, Tom Cruise spends more than two hours jumping off trains and motorcycles. It’s really exciting. However, as you watch it, you can’t help but wonder: wouldn’t a copyright lawsuit have been quicker? Because that could be the exact thing that destroys the actual AI industry if the courts get their way.

OpenAI, Anthropic, Google, Meta, and other generative AI companies are currently facing a legal reckoning that has been developing covertly since ChatGPT first came to the public’s attention in late 2022. It was first noticed by writers.

Next, artists. Next are the big news outlets. They all came to essentially the same conclusion: without ever asking or paying, these businesses had taken their work, scraped it, fed it into massive models, and then sold the results as a product. Then came the lawsuits. As of mid-2024, there were 19 of them, with more planned, and estimates of legal exposure exceeded $50 billion.

The most powerful venture capital firm in Silicon Valley, Andreessen Horowitz, didn’t wait for the courts to make a decision. A16z submitted a comment that read more like a distress signal than a legal brief when the U.S. Copyright Office opened a policy review into how copyright law applies to AI.

The company cautioned that applying current copyright laws to generative AI would “kill or significantly hamper” the advancement of these models. They likened the technology to the development of the microchip. They contended that allowing this to go unchecked is crucial to America’s standing in the AI world. It’s a daring claim. Many observers believe that it’s also a convenient one.

It is necessary to take a step back from the branding in order to comprehend what is truly at stake. Generative AI programs are not intelligent in any meaningful human sense, despite what the marketing suggests. They are unable to envision. They are unable to produce anything from nothing. They consume vast libraries of pre-existing text, images, and other media, find statistical patterns in it, and then, in response to cues, generate outputs that resemble those patterns.

This is what they do, and they do it impressively. Large portions of Reddit, extensive ebook libraries, and the entire English-language Wikipedia served as training data for early ChatGPT. Permission was not requested from any of those sources. The businesses may have honestly thought that this was within permissible legal bounds at the time. It’s also possible that they were fully aware of what they were doing and continued to advance.

The legal issue is multifaceted. First, is it infringement in and of itself to use copyrighted content to train a commercial AI model? The tech sector claims it is “fair use,” a legal theory that permits restricted use of copyrighted content under specific circumstances.

Opponents contend that selling access to outputs derived from someone else’s creative work after training a commercial product on it is the furthest thing from fair use. The courts have not yet provided a conclusive response. Every Silicon Valley boardroom and pitch deck is currently plagued by this uncertainty.

The second concern is the legal definition of AI-generated content. AI cannot possess a copyright, according to the U.S. Copyright Office. After learning that the author had created the images using Midjourney, the agency partially revoked a copyright it had granted for an AI-illustrated comic book. The decision was similar to one made by a federal court in 2016 regarding a monkey that had used a photographer’s camera to take pictures.

Because copyright requires a human author, the court determined that the selfies belonged to no one. The fact that generative AI companies are currently dealing with both issues at the same time presents a philosophical paradox: their tools may have been developed using someone else’s copyrights, and they are unable to own them.

Former Google CEO and current AI evangelist Eric Schmidt gave what he likely thought would be a private lecture to Stanford students in April 2024. If they wanted to create AI companies, he advised them to download the necessary content, ship the product, and then “hire a whole bunch of lawyers to go clean the mess up” if it took off. He continued, “it doesn’t matter that you stole all the content” if no one uses it.

The comments were leaked almost instantly. Later, Schmidt took them on a walk back. However, he seemed to have unintentionally expressed what Silicon Valley had been operating under for years.

Now, the industry is attempting to anticipate the issue. By paying for access to archives it may have previously trained on, OpenAI has been cutting licensing agreements with news publishers. Training on copyrighted material should be legally protected as fair use, according to a campaign started by a tech lobby group called “Generate and Create” that represents Amazon, Apple, Meta, Google, and other companies. Whether any of this will be sufficient is still up for debate.

The courts take their time. The number of lawsuits is growing. Furthermore, copyright law, which was created decades before any of this was possible, was never intended to address the fundamental question at the heart of it all: who owns the raw material of human knowledge and what rights exist when a machine is trained on it.

Observing this from the outside, it’s amazing how much was presumably assumed rather than proven before these companies grew to billions of users. The investor decks were never as solid as the legal underpinnings. There won’t be a dramatic train fight scene when the reckoning occurs, if it does. Most people won’t read the filings, decisions, and settlements until the harm has already been done.