Category: Culture

In Opposition to Copyright Protection for AI Works

This is a response to “Paradise Rejected: A Conversation about AI and Authorship with Dr. Ryan Abbott” hosted by Professor Sandra Aistars at the Center for the Protection of Intellectual Property (CPIP) at George Mason University School of Law. It was first published on the CPIP blog in conjunction with Professor Aistars’s post. 

On February 14, the U.S. Copyright Office confirmed its rejection of an application for a claim of copyright in a 2D artwork called “A Recent Entrance to Paradise.” The image, created by an AI designed by Dr. Stephen Thaler, was rejected by the Office on the longstanding doctrine which holds that in order for copyright to attach, a work must be the product of human authorship. Among the examples cited in the Copyright Office Compendium as ineligible for copyright protection is “a piece of driftwood shaped by the ocean,” a potentially instructive analog as the debate about copyright and AI gets louder in the near future.

What follows assumes that we are talking about autonomous AI machines producing creative works that no human envisions at the start of the process, other than perhaps the medium. So, the human programmers might know they are building a machine to produce music or visual works, but they do not engage in co-authorship with the AI to produce the expressive elements of the works themselves. Code and data go in, and something unpredictable comes out, much like nature forming the aesthetic piece of driftwood.

As a cultural question, I have argued many times that AI art is a contradiction in terms—not because an AI cannot produce something humans might enjoy, but because the purpose of art, at least in the human experience so far, would be obliterated in a world of machine-made works. It seems that what the AI would produce would be literally and metaphorically bloodless, and after some initial astonishment with the engineering, we may quickly become uninterested in most AI works that attempt to produce more than purely decorative accidents.

In this regard, I would argue that the question presented is not addressed by the “creative destruction” principle, which demands that we not stand in the way of machines doing things better than humans. “Better” is a meaningful concept if the job is microsurgery but meaningless in the creation or appreciation of art. Regardless, the copyrightability question does not need to delve too deeply into the nature or purpose of art because the human element in copyright is not just a paragraph about registration in the USCO Compendium but, in fact, runs throughout application of the law.

Doctrinal Oppositions to Copyright in AI Works

In the United States and elsewhere, copyright attaches automatically to the “mental conception” of a work the moment the conception is fixed in a tangible medium such that it can be perceived by an observer. So, even at this fundamental stage, separate from the Copyright Office approving an application, the AI is ineligible because it does not engage in “mental conception” by any reasonable definition of that term. We do not protect works made by animals, who possess consciousness that far exceeds anything that can be said to exist in the most sophisticated AI. (And if an AI attains true consciousness, we humans may have nothing to say about laws and policies on the other side of that event horizon.)

Next, the primary reason to register a claim of copyright with the USCO is to provide the author with the opportunity, if necessary, to file a claim of infringement in federal court. But to establish a basis for copying, a plaintiff must prove that the alleged infringer had access to the original work and that the secondary work is substantially or strikingly similar to the work allegedly copied. The inverse ratio rule applied by the courts holds that the more that access can be proven, the less similarity weighs in the consideration and vice-versa. But in all claims of copying, independent creation (i.e., the principle that two authors might independently create nearly identical works) nullifies any complaint. These are considerations not just about two works, but about human conduct.

If AIs do not interact with the world, listen to music, read books, etc. in the sense that humans do these things, then, presumably, all AI works are works of independent creation. If multiple AIs are fed the same corpus of works (whether in or out of copyright works) for the purpose of machine learning, and any two AIs produce two works that are substantially, or even strikingly, similar to one another, the assumption should still be independent creation. Not just independent, but literally mindless, unless again, the copyright question must first be answered by establishing AI consciousness.

In principle, AI Bob is not inspired by, or even aware of, the work of AI Betty. So, if AI Bob produces a work strikingly similar to a work made by AI Betty, any court would have to toss out BettyBot v. BobBot on a finding of independent creation. Alternatively, do we want human juries considering facts presented by human attorneys describing the alleged conduct of two machines?

If, on the other hand, an AI produces a work too similar to one of the in-copyright works fed into its database, this begs the question as to whether the AI designer has simply failed to achieve anything more than an elaborate Xerox machine. And hypothetical facts notwithstanding, it seems that there is little need to ask new copyright questions in such a circumstance.

The factual copying complication raises two issues. One is that if there cannot be a basis for litigation between two AI creators, then there is perhaps little or no reason to register the works with the Copyright Office. But more profoundly, in a world of mixed human and AI works, we could create a bizarre imbalance whereby a human could infringe the rights of a machine while the machine could potentially never infringe the rights of either humans or other machines. And this is because the arguments for copyright in AI works unavoidably dissociate copyright from the underlying meaning of authorship.

Authorship, Not Market Value, is the Foundation of Copyright

Proponents of copyright in AI works will argue that the creativity applied in programming (which is separately protected by copyright) is coextensive to the works produced by the AIs they have programmed. But this would be like saying that I have claim of co-authorship in a novel written by one of my children just because I taught them things when they were young. This does not negate the possibility of joint authorship between human and AI, but as stated above, the human must plausibly argue his own “mental conception” in the process as a foundation for his contribution.

Commercial interests vying for copyright in AI works will assert that the work-made-for-hire (WMFH) doctrine already implicates protection of machine-made works. When a human employee creates a protectable work in the course of his employment, the corporate entity, by operation of law, is automatically the author of that work. Thus, the argument will be made that if non-human entities called corporations may be legal authors of copyrightable works, then corporate entities may be the authors of works produced by the AIs they own. This analogizes copyrightable works to other salable property, like wines from a vineyard, but elides the fact that copyright attaches to certain products of labor, and not to others, because it is a fiction itself whose medium is the “personality of the author,” as Justice Holmes articulated in Bleistein.

The response to the WMFH argument should be that corporate-authored works are only protected because they are made by human employees who have agreed, under the terms of their employment, to provide authorship for the corporation. Authorship by the fictious entity does not exist without human authorship, and I maintain that it would be folly to remove the human creator entirely from the equation. We already struggle with corporate personhood in other areas of law, and we should ask ourselves why we believe that any social benefit would outweigh the risk of allowing copyright law to potentially exacerbate those tensions.

Alternatively, proponents of copyright for AI works may lobby for a sui generis revision to the Copyright Act with, perhaps, unique limitations for AI works. I will not speculate about the details of such a proposal, but it is hard to imagine one that would be worth the trouble, no matter how limited or narrow. If the purpose of copyright is to proscribe unlicensed copying (with certain limitations), we still run into the independent creation problem and the possible result that humans can infringe the rights of machines while machines cannot infringe the rights of humans. How does this produce a desirable outcome which does not expand the outsize role giant tech companies already play in society?

Moreover, copyright skeptics and critics, many with deep relationships with Big Tech, already advocate a rigidly utilitarian view of copyright law, which is then argued to propose new limits on exclusive rights and protections. The utilitarian view generally rejects the notion that copyright protects any natural rights of the author beyond the right to be “paid something” for the exploitation of her works, and this cynical, mercenary view of authors would likely gain traction if we were to establish a new framework for machine authorship.

Registration Workaround (i.e., lying)

In the meantime, as Stephen Carlisle predicts in his post on this matter, we may see a lot of lying by humans registering works that were autonomously created by their machines. This is plausible, but if the primary purpose of registration is to establish a foundation for defending copyrights in federal court, the prospect of a discovery process could militate against rampant falsification of copyright applications. Knowing misrepresentation on an application is grounds for invalidating the registration, subject to a fine of up to $2,500, and further implies perjury if asserted in court.

Of course, that’s only if the respondent can defend himself. A registration and threat of litigation can be enough to intimidate a party, especially if it is claimed by a big corporate tech company. So, instead of asking whether AI works should be protected, perhaps we should be asking exactly the opposite question: How do we protect human authorship against a technology experiment, which may have value in the world of data science, but which has nothing to do with the aim of copyright law?

 About the IP Clause

And with that statement, I have just implicated a constitutional argument because the purpose of copyright law, as stated in Article I Clause 8, is to “promote science.” Moreover, the first three subjects of protection in 1790—maps, charts, and books—suggest a view at the founding period that copyright’s purpose, twinned with the foundation for patent law, was more pragmatic than artistic.

Of course, nobody could reasonably argue that the American framers imagined authors as anything other than human or that copyright law has not evolved to encompass a great deal of art which does not promote the endeavor we ordinarily call “science.” So, we may see AI copyright proponents take this semantic argument out for a spin, but I do not believe it should withstand scrutiny for very long.

Perhaps, the more compelling question presented by the IP clause, with respect to this conversation, is what it means to “promote progress.” Both our imaginations and our experiences reveal technological results that fail to promote progress for humans. And if progress for people is not the goal of all law and policy, then what is? Surely, against the present backdrop in which algorithms are seducing humans to engage in rampant, self-destructive behavior, it does seem like a mistake to call these machines artists.

Addressing Fair Use Rhetoric in Debate Over SMART Act

On March 18th, Senators Tillis and Leahy of the IP Subcommittee introduced the SMART Copyright Act. The major functions of the bill, as codified in a proposed new Section 514, would empower the Librarian of Congress to approve designated technical measures (DTM) for identifying infringing material via a triennial rulemaking process. For a detailed description of the proposed rules and remedies in the bill, see Copyright Alliance CEO Keith Kupferschmid’s post. In this post, I wanted to respond to one criticism of technical measures for copyright enforcement—namely that they cannot account for fair use—but first, a recap of the background.

The SMART Act is a legislative response to the fact that after almost 25 years, the OSPs have rarely held up their end of the bargain under the terms of the Digital Millennium Copyright Act (DMCA), ratified in 1998. The foundation for all of Section 512 was predicated on the argument by the service providers of that period that they needed a liability shield against civil litigation stemming from the inevitability that customers would post copyright infringing material to their platforms. Thus, the statute lays out the conditions under which a platform can maintain the “safe harbor” shield, and as many copyright owners know, the major OSPs since then have not always complied with these conditions in good faith.

For the past quarter century, the major platforms have consistently avoided compliance with the notice-and-takedown process by, for instance, erecting unnecessary roadblocks for copyright owners to submit requests. Or, as a Virginia federal court recently affirmed, COX Communications remains on the hook for a billion-dollar damage award due to its failure to comply with the DMCA condition requiring the removal of repeat infringers.

These and many other examples paint a picture of a service platform industry that has fostered a culture of turning a blind eye to infringement and a reluctant, scattershot compliance with the statues they themselves lobbied to write. But specifically in regard to §512(i), which requires collaboration with copyright owners to develop standard technical measures (STMs) for identifying infringing material, Big Tech straight-up ghosted on the matter.

Remember that when the DMCA was being debated and drafted, it was the OSPs who presented their own technological capabilities as an implied promise that STMs could be developed to identify and remove infringing material. But not only did those service providers, and the bigger ones who followed, never engage with copyright owners to develop technical measures, they also funded a network of organizations (you know their names)[1] to promote the general theme that online copyright enforcement is fundamentally bad for society.

Kevin Madigan, VP, Policy and/ Copyright Counsel at Copyright Alliance, describes in a new post how The Network predictably repeats the same unfounded talking points no matter what proposal is on the table. And they have certainly dragged these orcs out of the mud once again in response to the SMART Act. But even as we examine the pros and cons of the bill itself, we should not lose sight of the fact that SMART is a legislative response to the OSPs’ refusal to comply with the conditions their own industry negotiated in the days of Web 1.0. So, maybe we can put the hyperbole in the drawer and sit down like adults? I wouldn’t hold my breath.

Technical Measures Can Never Account for Fair Use?

One gremlin The Network likes to call upon whenever technical measures are discussed for identifying infringing material is that an algorithm can never identify fair use. It is an oddly defeatist argument coming from representatives for an industry that makes bold promises about AI, and hardly allows the lack of perfection to stop them from experimenting with new products. But even if it is true that no algorithm could ever account for fair use, I’ll be blunt and say that neither can most of the platform users The Network claims to represent.

Let’s be real. The minority of professional creators who endeavor to be informed and engaged on copyright matters struggle with fair use; the experts who have worked in copyright law their entire careers struggle with fair use; and the courts struggle with fair use. Naturally, The Network exploits this uncertain landscape to imply that we could never hope for an algorithm to get fair use right. But this logical leap, which is meant to end discussion, also obscures the fact that the average social platform user doesn’t get fair use right either. That is if he even considers the question at all.

The term fair use is bandied about by The Network to promote the idea that it is the default status of most uses of protected works. It is a rhetorical strategy that, when paired with the message that copyright enforcement is inherently a form of censorship, promotes an ideological agenda seeking to elevate the fair use exception to the status of a civil right. But although it is true that the fair use doctrine evolved in U.S. law partly in support of the speech right, it remains an affirmative defense to a claim of copyright infringement. And the distinction matters.

Not only does The Network consistently ignore the literal First Amendment safeguards in the DMCA (which would not be disturbed by the SMART Act), but when they assert that no AI could ever account for fair use, I suspect they are alluding to a much larger constellation of presumed fair uses than actually exists. In other words, the argument against STMs, on the basis of fair use, almost certainly encompasses the effort to expand the volume and types of uses that copyright critics believe should be excepted under the doctrine.

I cannot prove this assumption and would not claim that, if true, it necessarily simplifies the technological challenge at hand. But it if we are going to take the matter seriously at all, it is important to know which definition of “fair use” is being applied—one grounded in case law, or one which the copyright critics would like to revise as they see fit?

As a practical matter, if the average user of a protected work on a social platform makes a fair use of that work, it is more likely the result of dumb luck than a well-informed and carefully considered decision. This is a common-sense assumption based on the low probability that the average user knows anything about the fair use doctrine. And if that is not correct, then perhaps the entire foundation for the liability shield codified in Section 512 should be reconsidered. Because the premise for this whole conversation was, and remains, that the average user of the internet does not know much of anything about copyright law.

Moreover, we forget that the presumed neutrality of the service provider is in contradiction with the idea that a fair use analysis should be a component of technical measures in the first place. The role of platform management was anticipated by the DMCA to be somewhat deaf and dumb in the process. Infringing material would be removed upon receipt of a valid notice, and if the uploader of the material believed the use to be a fair use, he could file a counter-notice to that effect. The human actors on the two sides of this equation were always anticipated to play active roles, and although The Network insists that the low rate of counter-notice filing is predicated on fear alone, it is also plausible that it is the result of many uses of works that are not defensibly fair uses.

What know for sure is that tens of millions of infringing uses occur every day and that only the large, corporate copyright owners have anything close to the resources necessary to mitigate the scope of piracy online. Large platforms like YouTube have deployed their own technical measures (e. g., Content ID and Copyright Match) insofar as they serve the platform’s bottom line. But these systems do little or nothing for independent and small business creators, and leaving this class of professionals in the digital dust was not the intent of the DMCA. Whether some version of the SMART Act can address the problem remains to be seen. But these rhetorical arguments against even trying are as tedious as they are hollow.

[1] In case you don’t, the Electronic Frontier Foundation, PublicKnowledge, Re:Create Coalition, Fight for the Future, Library Copyright Alliance, Authors Alliance, and a host of legal academics.

Image source by: idaakerblom

Piracy on Legal Platforms is Still Killing Creative Careers

Once again, the question arises whether there is any hope of addressing mass online copyright infringement on otherwise legal platforms?  It’s an exhausting problem, more than two decades old, and it isn’t getting better. A recent article by Annie Levin for Observer describes a new campaign by Music Workers Alliance (MWA), in which she sums up the heart of the problem thus:

Because of sites like YouTube, where all music can be accessed for free, streaming services like Spotify can get away with paying musicians a starvation average wage of $0.0038 per stream. A musician must have their song streamed almost half a million times a month to make minimum wage. Far from making a profit, musicians often end up in debt after making an album.   

As the article describes, MWA is focused on lobbying for changes to Section 512 of the DMCA, which shields online platforms from liability for copyright infringement committed by their users. But with regard to major platforms like YouTube, Meta’s properties, and Twitter, it is hard to wonder if we are not past the point of seeking meaningful amendment of the DMCA to better support smaller creators. As noted in my recent post during Fair Use Week, there is significant evidence suggesting that the major platforms are not in compliance with DMCA §512 anyway.

If that is the case, what can legislative reform accomplish? §512 is a voluntary provision under which compliance provides a “safe harbor” liability shield against litigation. In order to prove non-compliance, some entity would have to sue, for instance, YouTube, and spend the next decade or more in a fight with Google’s tobacco-industry scale legal counsel. And ain’t nobody got the resources for that. Thus, if the actual litigation shield today is that YouTube et al are simply too big to sue, then §512 is little more than enforcement theater, and amending it could be little more than legislative theater—at least as a means to address chronic infringement on the biggest platforms.

Then There’s the Standard Technical Measures Debate

On a related topic, the U.S. Copyright Office, in late December, sought public comments regarding the agency’s future consultation for the development of more robust Standard Technical Measures (STM) to identify and mitigate infringing uses of works on legal platforms. Under DMCA §512(i), Congress intended that OSPs and copyright owners would collaborate to develop STM, and at the time the DMCA was written, it was the OSPs who promoted the efficacy of STM as a basis for establishing the “safe harbor” provisions in the first place. But as all creators are painfully aware, in the 24 years since the law was passed, collaboration to achieve the Standard in STM has never happened.

Instead, a hodgepodge of technical measures are used, several of these developed by the OSPs themselves to address infringement only insofar as it serves the platform’s interests. For instance, as the members of MWA are well aware, YouTube’s ContentID system is provided to major labels and sound recordings made by megastars but is not available to many thousands of other owners of music rights—even works produced by fairly well-known composers and songwriters. As the STM comments submitted to the USCO by Copyright Alliance state:

… these technologies are usually not voluntarily made available to all types of copyright owners and OSPs have refused to come to the table with other stakeholders to have them formally adopted as widely recognized standards under section 512(i). This has led to a lack of uniformity among and access to existing technical measures that makes it difficult for those copyright owners who do not have access to combat infringement. On the other hand, OSPs prefer the status quo because it allows them to avoid adopting and implementing standard technologies.

Copyright owners would like to see the Copyright Office play a more integrated, regulatory role to achieve standardization of technical measures to better protect the works of a much broader spectrum of copyright owners. Meanwhile, the Electronic Frontier Foundation and similar organizations cite flaws in existing technical measures as a basis to argue against any expansion of these technologies. “Despite years of financial and technical investment, filtering technologies continue to do a poor job of sorting legal expression from infringement,” state the comments submitted by EFF to the Copyright Office. More specifically, the comments state the following:

The core problem is this:  distinguishing lawful from unlawful uses often requires context. For example, the “amount and substantiality” factor in fair use analysis depends on the purpose of the use. So while the use may be a few seconds, as for some kind of music criticism, it can also be the whole piece, such as in a music parody. Humans can flag these differences, automated systems cannot.

There is a measure of truth in this refrain, which is played every time the topic of STM is on the table. But its relevance should be considered in terms of evidence rather than theoretical debates. At internet scale, with infringements occurring in the tens of millions every month, and Big Tech still saying, “We can’t police it all,” I think the fair question asks why creators like those in MWA should continue to bear the cost of allowing the perfect to be the enemy of the good?

If the argument is that the good (i.e. better STM) can never be achieved, then the EFF and their fellowship should be required to cite more than anecdotal evidence of potentially harmful error. Because after twenty years, the system in place has a 100% failure rate for many stakeholders on the enforcement side of the copyright equation. That we should continue to allow smaller creators to watch their careers dissolve simply because STM will produce some error is an immoral argument for the status quo. And when that argument is paired with the fact that the major OSPs are now barely incentivized by the DMCA liability shield, the resulting “wage theft” described in Levin’s article is downright criminal.