AI wins Round One of copyright lawsuits.

Judge Orrick throws out most of the class action lawsuit against Stability.ai, Deviant Art and Midjourney. But the plaintiffs will be back with an amended complaint.

Hello Reader! In recent weeks. I’ve been halfway around the world, working with clients on strategic planning projects and making presentations about future trends.

I just returned home from Amman, Jordan, where I was invited to address the Governor and senior staff of the Central Bank as well as the management of the Arab Jordan Investment Bank on the subject of artificial intelligence in banking and financial regulation. It was a lively session followed by several hours of discussions. My visit occurred during a period of intense conflict in nearby Gaza, but Jordan remains peaceful and calm.

Three weeks ago, I gave the opening remarks about collaborative creativity at the financial regulator FINRA’s annual hackathon event, where 28 software teams compete to invent novel solutions for financial regulation. This year the focus was AI and cryptocurrency. It was an exciting and upbeat event. FINRA is moving nimbly to embrace AI in every part of the organization.

I also gave the opening remarks on the topic of “The Inevitability of Corruption” at the Minnesota Blockchain Association’s annual event. I was joined by a group of Web3 experts, including Dr Wulf Kaal, with whom I am collaborating on a decentralized governance project. The title of my talk refers to the fact that the USA seems to have retreated from free market competition by allowing rent-seeking oligopolies to dominate many fields. During the past 20 years, a wave of consolidation has occurred across 75% of American industrial sectors, accompanied by regulatory capture, legislative gridlock and a series of rulings from the Supreme Court that favor unlimited dark money in politics. As a result, many US industries demonstrate characteristics of cartels. Significant negative consequences follow, including higher prices, fewer jobs, a 30% decrease in investment in innovation and the slowing growth of real GDP in the US. My opening remarks about these trends set the agenda for a full day of workshops to explore alternative consensus mechanisms based on transparent, decentralized Web3 governance.

I intend to post these talks here because they each point to important trends that will shape the near future… but first I plan to finish my series on copyright and AI.

One more housekeeping note: on November 10 and 11, in partnership with Creative Startups, I will convene a group of thinkers and leaders in New Mexico at CXSF Creative Experience Santa Fe for a series of facilitated discussions to address the question, “Can new technology help us build resilient communities?” Our goal for this event is clear: we encourage attendees to move past fear and resistance of artificial intelligence towards an understanding of how AI together with real-time 3D, immersive media and possibly decentralized ledger technology can foster better outcomes for local groups. Participants include technologists, sociologists, civic data experts, designers, artists and representatives from several tribal communities. If you are interested in attending, please let me know. As the event curator, I’ve conferred with each speaker and session facilitator, and I can assure you this is going to be a worthwhile gathering.

---

A human artist faces a monstrous robot in the final showdown between humanity and Big Tech. Her secret weapon? Copyright law.

This newsletter is about the future. Lately I have been keenly interested in how generative AI seems to be on an unavoidable collision course with copyright law. The way I see it, copyright law is all about preserving the past at the expense of the future, which sets up an inevitable conflict.

Will the laws of the past constrain the future? If so, how?

To understand where this conflict is going next, it’s necessary to begin with an understanding how things got started. That’s why I began by writing a series of articles about the evolution of copyright from its early origins in the Middle Ages and the Reformation. Next week I will resume with that series, shifting the focus towards North America, where I will explain how copyright took root during the colonial era and later evolved into the modern copyright industry. There is a method to the madness.

But first, let’s attend to the news.

This week, news broke about the copyright lawsuits that were filed against generative artificial intelligence startups. I feel compelled to share it with you here with commentary.

Results from the US Court in the Northern District of California.

On Tuesday October 31, Judge William Orrick of the US Court in the Northern District of California issued an order on a motion to dismiss the complaint brought by Plaintiffs Sarah Anderson, Kelly McKernan, and Karla Ortiz against two of the leading generative AI startups, Stability AI and Midjourney, along with with art web site Deviant Art.

Judge Orrick wrote: “Finding that the Complaint is defective in numerous respects, I largely GRANT defendants’ motions to dismiss and defer the special motion to strike.”

Ouch, that’s gotta hurt.

You can read Judge Orrick’s order here. It’s a worthwhile read because it concisely summarizes both sides of the issues in contention, and it describes clearly the defects in the complaint, and it sets forth the legal framework that guides the judge’s reasoning. For anyone interested in generative AI, this is a must-read.

In light of the many defects in the complaint, the judge’s order seems generous to the plaintiffs, providing them clear instructions and the opportunity “to amend to provide clarity regarding their theories of how each defendant separately violated their copyrights, removed or altered their copyright management information, or violated their rights of publicity and plausible facts in support.”

The plaintiffs have 30 days to submit an amended complaint. Their attorneys have already indicated that they intend to do so. I am doubtful that they will be entirely successful.

Only the copyright infringement claim by one artist, Sarah Anderson, survived the Motion to Dismiss. Naturally, the plaintiffs’ attorneys contend that this is the most important part of the lawsuit. They have to say that, because otherwise they look like they got clobbered. Below, I’ll talk about the implications of this surviving claim.

We’ve been following this case from the outset. Judge Orrick’s ruling comes as no surprise. This case was deeply flawed since the beginning.

• For starters, two of the plaintiffs had failed to register their work with the copyright office, so they had no standing to bring a lawsuit for copyright infringement. Oops. Sloppy. These claims were dismissed with prejudice, which means they cannot be introduced again.  

• Second, the plaintiffs failed to cite the specific works that have been infringed. That seems like a baseline requirement for an infringement case, since copyright only pertains to specific, fixed works. Their attorneys will need to fix this in the amended complaint, which may or may not be easy because they’ll need to show that the defendants had access to the works and that they used them in training.

• Copyright only applies to specific works. Copyright does not apply to an artist’s style or a technique. Also, you can’t copyright an idea or a concept or a theme. The only thing that can be copyrighted is the expression of an idea in the form of a fixed work. Yet the Anderson complaint included claims that an individual artist’s style was infringed. Those claims were dismissed, and there’s scant likelihood they will be amended to the court’s satisfaction.

• Fourth, the plaintiffs made a lot of assertions in the complaint without indicating that they have sufficient supporting evidence, which is why the judge sent them back to the drawing board. 

• Finally, it seems to me that it is going to be difficult for them to come up with evidence that will cure all of the defective claims, but who knows? Maybe we will see some legal creativity. That’s what makes this case so interesting!

In sum, it was a poorly constructed complaint, which was noted by some observers when the suit was filed. So it comes as no surprise that the judge threw most of it out.

Copyright law is the first line of defense against technological disruption, but it doesn’t always work.

For those who have not been following the brewing conflict between publishing and generative AI, it may be useful to begin with the observation that copyright law is the first line of defense that old-school media firms always use to stop disruption from new technology. It’s not even a strategy, more like a knee-jerk response to innovation.

Big media companies routinely deploy copyright infringement lawsuits as a way to hobble any new, fast-moving innovation in digital media consumption. They have sued to stop VCRs, DVD players, DVRs, file-sharing, search engine indexing and streaming. With mixed results.

These cases often end in failure, followed by the paradoxical outcome that the media industry and content consumption continue to expand as a result of each wave of technology. So why do media companies reflexively sue innovators?

The media industry deploys copyright law to preserve the past at the expense of the future. That happens because media executives fail to see the potential in new technology and they are incentivized to focus on next quarter’s revenue and this year’s performance bonus. They prefer to make a dollar today from familiar distribution patterns rather than two dollars tomorrow from something new. Anything that threatens to upend the existing business model is considered a threat.

Slow it down, shut it down, sue it and stop it. That’s the mantra.

Generative AI presents the mind-blowing possibility that billions of consumers may someday generate their own stories, pictures, animations, games and films. That would completely demolish the business of publishing content. Or perhaps create entirely new business models for bundling and packaging content.

It’s easy to imagine that the major media companies are gearing up to file a flurry of lawsuits against generative AI companies.

But not in this particular case.

What’s different is that some of the first wave of genAI copyright lawsuits are not being filed by motion picture studios or book publishers, but rather as class action suits on behalf of individual artists. And so it is with the Anderson case.

Class action lawsuits won’t make individual artists rich. But they could pay off handsomely for the law firm.

It’s uncommon for an individual author or writer to file an infringement lawsuit because the cost is prohibitive. Most individual artists lack the money to gamble on the legal process.

That’s where class action lawsuits come into the picture. Some attorneys offer to waive some or all of their fees in exchange for a substantial cut of the damages or settlement payment.

Even though they will be obliged to give away up to 50% of the proceeds, this proposition may sound appealing to the artists and authors. Free money for no effort, what’s not to like?

Anderson is one of the first cases to reach a courtroom because a class-action attorney named Joseph Saveri filed the lawsuit in a hurry. Saveri’s firm is relatively new, founded in 2021, but he previously spent 20 years bringing class action suits against tech companies at another law firm.

It’s a good business, but some people consider it sleazy. Just ask any tech industry CFO: such lawsuits are an endless nuisance. It’s often cheaper and less distracting to pay off the plaintiffs, regardless of the merit of the suit, rather than the alternative of slogging through the expensive and time-consuming process of fighting off a suit. That’s why class-action suits grow on a public company like barnacles on a ship’s hull.

To skeptics, the Anderson case smells like a shakedown. An opportunistic money grab. Why? Because in class action suits, the big winners tend to be the attorneys, not the plaintiffs. The complaint is wrapped up in the noble language of “artists rights” but the reality is that there is a lot of money at stake to be divided up if successful.

This is particularly relevant to generative AI, where the model is trained on millions or billions of works by different authors. If a class action suit agains a GenAI company were to prevail, the proceeds from the lawsuit would be divided among many artists, potentially millions of artists. The money paid to any individual artist in the class would be tiny. It would make Spotify’s miserly streaming royalties look like a huge bounty by comparison.

But the law firm would take a huge chunk off the top, sometimes as much as 50% of the prize. Ka-ching!

Copyright law doesn’t do what you think it should

Six months ago, I predicted that many of the opportunistic copyright infringement lawsuits against the generative AI companies would be dismissed or would fail.  

But I did not begin with that notion. Quite the opposite.

When consumer-facing generative AI apps were first introduced last year, I was on the side of the artists. After all, I am a painter and an author so my sympathy lies with creative artists. It seemed to me outrageous that an AI company could scrape the web and make use of the entirety of an artists' work, and then re-sell that artwork to us as a GenAI service for $20 a month. 

But instead of getting upset, I got curious about copyright law. What I learned changed my mind. I was invited by Creative Commons to participate in a series of public discussions about copyright law, media and AI at ASU’s campus in downtown Los Angeles. There I met several intellectual property attorneys who were involved with AI, including some who are involved in litigation. We had lively exchanges about AI and its potential impact on the media business.

My conversations with IP experts led me to understand that there are many reasons to believe that our current copyright laws are a poor fit for generative AI, and that many of the copyright arguments against genAI are based on flawed assumptions.

That's not to say that the artists don't have a valid moral claim. They may. Their problem is that they may have a weak copyright infringement claim. US copyright law is indifferent to morality.

You see, copyright doesn't do what most creative people think it does. It doesn’t give an artist blanket monopoly power over all uses of the work, even though that is what many people would prefer to believe.

Copyright law is very clear about the fact that the monopoly right is conditionally granted with certain carveouts that are spelled out clearly in the 1976 Copyright Act. These carveouts are known as Fair Use. The carveouts make it possible for other people to make use of the work without permission from the author or artist. That’s intentional. Fair use is your right as a US citizen. You have the right to quote the work, cite the ideas in the work, teach it, train with it, analyze it and critique it, and even create a parody of it. You can do all of these things without the permission of the author or artist.

(In the UK, Canada and Australia, there is a somewhat similar legal concept called “fair dealing” but it differs meaningfully from US copyright law, so if you live in those countries, I caution you not to rely on what I’ve written here, which only pertains to the US).

Authors and artists don’t love permissionless fair use. Some of them are opposed to the idea that other people can make use of their work freely, without even the courtesy of notification or a thank you. That’s why their guilds and unions often fulminate against it.

But those who oppose fair use fail to understand that, without the fair use carveout, copyright law would be unconstitutional. If the artist or author had total veto power over other people’s ability to quote the work or discuss it, that would be an unconstitutional restraint of free speech.

At the core, copyright is a simple economic deal. A trade-off. The government agrees to grant a time-limited monopoly to an artist or author, but only in exchange for the right of the public to make use of the work in the specific ways defined as Fair Use. You can’t have the monopoly without Fair Use. They go together. This principle is established in the US Constitution, Article 1 Section 8. It is bedrock law.

For two centuries, copyright maximalists have worked relentlessly to reduce or eliminate Fair Use and expand copyright. They’ve been very successful. In a future article in this series, I’ll examine the sordid history of the campaign to expand copyright. It is a dismaying example of how savvy attorneys can turn a well-intentioned law against itself.

For now, let’s consider that one piece of the lawsuit that survived Judge Orrick’s red pen.

What about that one surviving infringement claim?

This lawsuit is going to come down to a battle about fair use. Is it fair use to train a large language model or a diffusion model on publicly-available copyrighted artwork?

It’s easy to predict how defendants will respond. Looking at artwork is not infringement. People look at artwork every day, in print, online, in galleries and shops and in museums. They can even make copies for personal use: for example, if you visit any art museum during a weekday, you are likely to see an entire class of students making drawings of classic works of art. This is fair use for education and training, which is clearly permitted in the 1976 Copyright Act. How is it different for a machine?

The lawyers defending the AI companies will argue that an AI model is “looking at art” just like a human student. That is the essence of machine learning. The remarkable capabilities of neural networks are the result of their ability to learn in ways that resemble how humans learn.

For imagery, we learn by looking. So does the AI. Defendants will claim that training the AI by looking at artwork is fair use.

In the Anderson case, only one claim survived the Motion to Dismiss intact. This is the claim of copyright infringement against one defendant, Stability AI. This claim survived because the artist in question, Sarah Anderson, had the foresight to register some of her work with the US Copyright Office, and therefore she has standing to proceed with the suit.

Specifically Anderson’s attorneys contend that Stabilty “downloaded or otherwise acquired copies of billions of copyrighted images without permission to create Stable Diffusion,” and used those images (called “Training Images”) to train Stable Diffusion and caused those “images to be stored at and incorporated into Stable Diffusion as compressed copies.”

The defendant rejected this claim in the Motion to Dismiss, but this matter is far too complex for the court to decide at this early stage in the process. The judge will need to see evidence and hear arguments. Which is why this particular claim survived the challenge, and therefore parties will need to go through the whole lawsuit to arrive at a conclusive decision.

In the process, important questions will be asked:

• Is it infringement to train an LLM or diffusion model on publicly-available copyrighted work or is this fair use?

• During training, do the large AI models make and retain a copy of the original work?

• Is it true that 100% of the output of a large language model or diffusion model consists of “derivative work”?

• Is is possible to generate output with generative AI that is “substantially similar” to the original work used in training?

• Is the use of the original work in training an AI “transformative” in the sense that it converts the original work into data that is used for another, entirely different purpose?

• Does the output generated by AI compete in the marketplace against the original work?

Watch this space. The Anderson case promises to shed light on these contentious issues. The answers to these questions may affect the trajectory of the entire field of generative AI.

Saveri is joined in this case by another attorney, Matthew Butterick, who happens to be a software programmer. Butterick has written extensively about the theory behind the claims in this case. You can read his commentary here. I recommend it because Butterick clearly understands the math behind the diffusion model and is adept at explaining it in plain English. That said, beware that this site is an open solicitation to artists to encourage them to join the class action suit, so it would be unwise to consider Butterick’s assessment as fair, neutral, even-handed or even necessarily accurate. It is persuasive.

One of Butterick’s interesting arguments is that diffusion models create copies of the work during the training cycle. This argument will be hotly contested in court.

The AI companies contend that no copies are made or retained. According to them, that’s not how generative AI works.

At the heart of the debate is the LAION data set, which is a collection of images gathered by a German company (at Stability’s request) used to train Stable Diffusion.

If Butterick’s claim is true, then this is direct infringement. If proven, then Stability will be liable. Case closed. Such an outcome would have major implications for every generative AI company.

How might they prove it? Butterick spells it out. According to his theory, a diffusion model contains a mathematical representation of a fuzzy (lossy) copy of the artwork that can be compared to a low-resolution MP3 or JPEG file. Low resolution or not, Butterick contends it is an unlicensed copy of the original. The defendants pushed back against this, arguing that no application could plausibly contain compressed versions of 5 billion images.

This matter is so complex that it must be decided in the courtroom and supported with evidence. In the Order, Judge Orrick left the door open to test this novel theory of image compression. First, he quoted this part of the Complaint on page 9 of his Order:

Because a trained diffusion model can produce a copy of any of its Training Images—which could number in the billions—the diffusion model can be considered an alternative way of storing a copy of those images. In essence, it’s similar to having a directory on your computer of billions of JPEG image files. But the diffusion model uses statistical and mathematical methods to store these images in an even more efficient and compressed manner.

Then he instructed Butterick to amend the Complaint in the following way:

Plaintiffs will be required to amend to clarify their theory with respect to compressed copies of Training Images and to state facts in support of how Stable Diffusion – a program that is open source, at least in part – operates with respect to the Training Images. If plaintiffs contend Stable Diffusion contains “compressed copies” of the Training Images, they need to define “compressed copies” and explain plausible facts in support. And if plaintiffs’ compressed copies theory is based on a contention that Stable Diffusion contains mathematical or statistical methods that can be carried out through algorithms or instructions in order to reconstruct the Training Images in whole or in part to create the new Output Images, they need to clarify that and provide plausible facts in support.

That won’t be easy.

It is way too early to predict the outcome of this case, but this novel theory has the potential to upset many of the assumptions that tech companies made about training large language models and diffusion models. If Butterick prevails in the Anderson case, then every GenAI company will be obliged to license their training data or pay astronomical damages in future class action lawsuits.

Indeed we can already glimpse the outlines of an emerging data licensing business, regardless of what happens in this particular case. The leading GenAI companies have initiated licensing negotiations with media companies and online communities that possess desirable training data. This is one reason why Twitter/X and Reddit have publicly declared that GenAI companies are prohibited from freely accessing the data created by their communities: they are hanging a sign in the shop window that says, “Open for business.” The same is true for the New York Times, CNN, the Chicago Tribune and the Australian ABC. Reportedly, OpenAI is in negotiations with Reuters and other media firms, although their negotiations with the New York Times seem to have soured. Rumors in Los Angeles suggest that Runway and other GenAI firms are in the process of negotiating with motion picture companies for access to their vast libraries of movies.

In the media industry, copyright is a moat that protects a lucrative business. These firms are content to let Butterick and Saveri do the legal grunt work.

---

As I wrote above, this newsletter is intended to address the future broadly, but lately I’ve devoted most of it to rapid developments in the field of artificial intelligence and the inevitable clash with copyright law because this area is evolving so rapidly and it holds the potential to affect so many industries and professions in the future.

I’ll be back next week the second part of my series of articles about the history of copyright. In the next series, I will consider the evolution of copyright in the United States and the rise of mass media, leading to the Internet and AI.

All images generated by Midjourney, and therefore not subject to copyright. (irony!)