This week marks the 50th Anniversary of the Copyright Act of 1976, which was passed into law on October 19, 1976. Like every copyright law before it, the ‘76 Act (as true nerds know it) was full of trade-offs. Exclusive rights and fair use, exclusive licenses and termination of transfers: in copyright, almost every quid has a quo. One of the oldest copyright bargains was compliance with copyright formalities in exchange for the long, strong federal monopoly of copyright itself. For the first 200 years or so, published works entered the public domain unless their owners took affirmative steps, often called “formalities,” to claim their federal rights. Content industries upended that bargain in 1976, taking longer, stronger rights without any obligation to register their works or attach copyright notices to published copies.
As the AI age takes hold, they might live to regret unwinding that particular deal. If formalities had persisted, notice and registration, suitably updated for modern technology, could offer solutions to some of the market challenges raised by artificial intelligence. For example, formalities might have enabled the training data “transparency” called for by some. But since formalities have been mostly dead for half a century, any fully comprehensive copyright transparency effort is doomed.
Not to be confused with transparency requirements that relate to safety, fair competition, and other broad social values, copyright transparency bills are only interested in transparency about the presence of copyrighted works in training data. These bills purport to serve copyright holders by requiring AI researchers to collect and disclose information about every single copyrighted work, and every copyright owner, represented in their training datasets. Once disclosed, these inventories would allow copyright holders to search the disclosures and fire off demand letters and complaints to AI labs, or solicitations to would-be class-action clients. That’s the bills’ avowed goal: to enable more legal actions against AI labs over training.
California’s AI copyright “transparency” bill AB 412 started out with two requirements for all AI model developers: to create a comprehensive list of all registered copyrighted works used to train any AI model, and to respond promptly to any rightsholder who asked whether their work was used. The bill’s sponsors eventually realized that the first requirement was impossible and retreated to requiring AI developers to respond swiftly to copyright owner inquiries about whether their works were in the devs’ training data. The bill died over the summer amid lingering doubts about its technical feasibility. Its federal siblings, the TRAIN and CLEAR Acts, take similar approaches: TRAIN requires disclosure on demand, while CLEAR requires disclosure of a comprehensive inventory.
AI copyright transparency bills have many flaws. They are redundant: courts already have tools to help would-be plaintiffs prove copying when they have an otherwise valid claim. They are premature: requiring disclosure in all cases assumes AI training is never a fair use, when two courts have already ruled that it is. Relatedly, state laws like AB 412 are preempted by the federal Copyright Act because they effectively create new state copyrights that exceed the bounds of federal law.
But the deepest, most fundamental problem, the one no amendment can fix, is that full compliance with these bills is impossible. “Transparency” bills assume that someone building a training dataset can reliably ascertain two basic facts about the billions or trillions of texts, images, videos, and audio files in their data: 1) what copyrighted works do they contain, and 2) who owns those works. To glean that information, you’d need to have a robust system of markings on copies of in-copyright works to indicate their identity and status, and a comprehensive registry of in-copyright works, with up-to-date information about ownership, that could be correlated to the markings on copies in your training data.
We don’t have those things – not even close. While some industries and sectors have codes and markers (and databases to match), like the ISBN number for books or the DOI for academic articles, these systems are the exceptions that prove the rule. Applying these mechanisms to some published AI training datasets is what enables the search tools you may have seen that tell you whether “your work was used to train AI.” The effectiveness of these tools depends on their narrower focus—they only track niche subsets of works. They also require authors and publishers to take affirmative steps to register their works using the system, and to keep that information up to date. There is no single, comprehensive system that reliably marks all copies of copyrighted works as such, nor is there a comprehensive registry that could connect all copies to copyright holders.
It wasn’t always this way. For nearly 200 years, copyright was treated as a valuable privilege that could only be secured by compliance with strict rules. Non-compliant publications would forfeit copyright and immediately enter the public domain. These rules, known as formalities, evolved over time. The first Copyright Act, passed in 1790, required an application be submitted to a federal court, a notice published for four weeks in a major newspaper, and a copy of the work mailed for deposit to the Secretary of State. Starting in 1831, “notice” of copyright was attached to copies of works rather than advertised in the newspaper, and in 1870 the deposit and registration system was centralized in the Library of Congress. For over a century after that, we had a centralized registry of all federally copyrighted works, and any work covered by federal copyright was required to bear a standard notice. There was even a mechanism to keep the records at least somewhat up-to-date: copyright registrations had to be renewed after an initial term, which provided fresh ownership information to the Copyright Office, where anyone interested could look and find out who owns what.
Formalities weren’t just regulatory hoops to discourage would-be monopolists. Compliance with formalities – placing accurate notice on published copies and filing timely registration with the Copyright Office – would allow the public to ascertain whether a particular work is in-copyright, who owns it, and how long it will be protected. Complying with formalities was seen as a reasonable burden to bear in exchange for the powerful monopoly privileges granted by federal copyright law, and it created useful public information. It was, in other words, a system that would enable exactly the kind of transparency some content industry groups now claim to want. The content industries lobbied for decades to scrap it, with remarkable success.
In the post-war boom years, the U.S. morphed decisively from a young nation thirsty for mostly foreign works of science, philosophy, and literature into a cultural powerhouse selling Mickey Mouse and Elvis Presley to the world. The copyright industries became eager for the US to join international frameworks that gave their works stronger protection in foreign markets. To do so, however, the U.S. would have to modify its copyright law to look more like Europe’s, which is based on a belief in the natural rights of the author, not the utilitarian benefits of a limited government monopoly. European copyrights were granted instantly and automatically at the moment a work was created; to Europeans, U.S.-style formalities were an insult to authorial dignity.
Formalities were put on life support in the Copyright Act of 1976 and Congress pulled the plug in 1989 when the U.S. fully joined the Berne Convention, the treaty embodying Europe’s vision of copyright. Industry insiders and academics had spent decades condemning mandatory notice as a “trap for the unwary,” pointing to victims like George Romero, whose Night of the Living Dead was thrust into the public domain in 1968 because a distributor accidentally left off the required copyright notice.
Authors would be better off, it was argued, if we “harmonize” U.S. law with the Europeans. Industry conglomerates would also gain longer, stronger copyrights abroad, and easier enforcement against piracy there. Today, millions of copyrights are awarded automatically each year to authors who likely have no idea of their rights. Trillions of emails, Reddit posts, memes, text messages, photos, and videos are encumbered by copyright until 70 years after the author’s death, and there’s no record of which work is which, and who owns what. Had the U.S. retained formalities, the universe of in-copyright works would be smaller and easier to track. In a post-formalities world, that universe is vast and untraceable.
AB 412’s slow death proves the point. Tellingly, the version of the bill that finally expired had retreated from covering all copyrighted works to covering only registered works, a quiet concession that the full universe of copyrighted material—all those blog posts, Instagram photos, and other forms of casual, everyday expression—is a hopelessly untraceable mess. But retreating to the registry doesn’t solve the problem, because registration was only ever half of the old system. Registration told you who owned a work; notice on the copy told you which work you were looking at. Strip out reliable notice and a developer sitting on a trillion files has no way to know which of them, if any, corresponds to which of the millions of registrations in the Copyright Office’s records. Crucially, those records contain no sample of the work at all, and are not reliably updated to reflect changes in ownership or control, so even matching a title or an author to a registration record doesn’t guarantee you’ve got the right work, or that you’ll identify the right owner. When critics pointed this out, AB 412’s sponsors suggested AI developers could just nerd harder to fill the information gap with some kind of technology, details TBD. The California Senate Appropriations Committee did the right thing by declining to pretend that was possible.
The system of formalities that the 1976 Copyright Act torched 50 years ago had its downsides, but it created an authoritative and comprehensive record of federally-protected works. Reformalizing copyright could restore one of copyright’s broken bargains, and rightsholders who have rejected formalities in the past may now see the upside of creating a comprehensive copyright registry in the AI age.