Artificial intelligence (AI) has been variously characterized as the greatest boon and the greatest threat to humankind. Clearly, it’s too early to tell whether the enthusiasts or the doomsayers are right. And the same is true about the impact of AI on law in general and publishing law in particular.

Here’s a summary of where we stand and where we are heading in terms of the law of AI. But stay tuned—the law is being written and rewritten even now.

Generative Artificial Intelligence

AI can do a great many things, but the focus of this article is “generative” AI, that is, a computer-based digital technology that can be trained to generate what the law recognizes as “expressive” content—text, static and moving images, animation, sounds, etc.—in response to a request from the user. By way of example, and at least in theory, a publisher can ask an AI program to write and illustrate a book about a specified subject by entering a “prompt,” and the AI program will deliver a finished manuscript.

At least three different legal issues arise whenever generative AI is used to create content for a book or other publication.

First, is it lawful to train an AI program with content that contains copyright-protected works of authorship that belong to someone else?

Second, is it lawful to use AI to create content if a work of authorship by a third party can be detected in the end product? Third, and if so, who owns the end product?

The answers to these questions are still incomplete and will change over time, but we already know a few fundamentals.

AI and Fair Use

“Fair use” is a fundamental element of US copyright law. Under the so-called fair use doctrine, the use of limited portions of a copyrighted work without permission of the owner is a defense to a copyright infringement claim. Fair use creates a kind of safe harbor for authors and publishers, and virtually every book is likely to contain at least some third-party content that is used without permission but is defensible on the fair use doctrine.

Fair use is not an affirmative right. Rather, it is an “affirmative defense” to an infringement claim, and whether any particular use is fair use can only be determined by a court on the facts of a specific case. The fair use doctrine itself consists of several factors, none of which is dispositive. However, so much case law on fair use has accumulated since the doctrine was introduced in an 1841 appellate opinion that authors and publishers (and their attorneys) feel comfortable in making practical decisions on what is and what is not likely to qualify as fair use. Such decisions are made every day in the publishing industry precisely because fair use is the lubricant that allows the machinery of publication to operate.

No such decisions can yet be made about AI-generated content. Although there are pending cases that will test the applicability of the fair use doctrine to generative AI, none of them have reached the status of settled law. All we can say now is that the courts and perhaps Congress may provide an answer to the question of whether, how, and to what extent training data and AI-generated content can be used without permission of copyright owners whose works of authorship can be discerned in the input and output of AI technology.

The answer, when it comes, will almost certainly be based on the fair use doctrine. For that reason, the answer will not consist of a set of fixed rules. Rather, as with every decision based on fair use, the decision to rely on the fair use doctrine in selecting the training data for AI or using the content generated by AI will include a degree of legal risk.

Ownership of Generative AI Content

The earliest legal ruling about AI was issued in 2018 by the US Copyright Office, which refused to register the copyright in a work of authorship that was generated entirely by an “a computer algorithm running on a machine.” The ruling was based on the fundamental proposition that copyright protects works of authorship by human authors. Since the AI-generated work was created “without any creative contribution from a human actor,” it was not eligible for a copyright and belonged to no one. To put it another way, anyone can use AI-generated content because it is outside the protection of copyright.

The Copyright Office later modified the rule by making a distinction between works that are authored in their entirety by AI and works that are co-authored by AI and a human author. In February 2023, the Copyright Office registered the copyright in a graphic novel that consisted of text that was authored by a human being and images generated by AI. Even so, only the text—and not the images—was protected under copyright.

So far, the legal position that is taken by the Copyright Office is only a “statement of policy” and not a law. Eventually, the policy will be tested in the courts, and only then will know whether AI-generated content automatically falls into the public domain. However, as a practical matter, the policy currently imposes a duty on applicants for copyright registration to “disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author’s contributions to the work.”

For now, that’s enough guidance for authors and publishers to seek registration of works that include some human-created content and some AI-created content.

Contract Clauses

There are no standard contract clauses that apply to AI in common use, but many of us who practice publishing law have begun to draft such clauses at the request of our clients. Here are a few of the many approaches that are available.

No Training Use – The first AI contract clause that I drafted was requested by an author who did not want her work of authorship to be used by the publisher as training data for AI. After some negotiation, I drafted a clause that prevented both the author and the publisher from doing so. However, the language needed to be carefully crafted because training data is typically vacuumed up by search engines without bothering to ask or even to inform the copyright owners whose work is copied. We used a clause that prohibited the author and the publisher from “knowingly, willfully, and affirmatively granting permission to third parties” for such uses. If a third party uses the work without permission, neither the author nor the publisher will be in breach of contract.

No Use of AI in the Author’s Work – AI clauses in publishing contracts will sometimes require the author to “represent and warrant” that no artificial intelligence has been used in the preparation of the work that is submitted to the publisher. In my experience, some authors will push back on a “no-AI” clause because some form of AI may have been used in course of research or because the author consulted but did not use AI-generated content in the final manuscript. Also, an author who signs the contract without disclosing the use of AI puts the publisher at legal risk without any warning of the problem. For these reasons, I prefer a more nuanced approach as discussed below.

Disclosure of Use of AI in the Author’s Work – A more serviceable AI clause is one that a) requires the author to disclose in writing whether and how any form of AI was used in the preparation of the work; b) requires the author to identify any specific content in the work that is AI-generated; c) entitles the publisher to remove any AI-generated content from the work; and d) obliges the author to replace the AI-generated content with content that has been authored by a human, whether it is the author or a third party from whom the author has secured the right to use the content. On the assumption that the author is forthcoming in making the required disclosures, both the author and the publisher are better positioned to assess and address the legal risk.

Representations, Warranties, and Indemnities – The standard contract clauses that set forth the author’s representations, warranties, and indemnities should always be revised to include an explicit representation that the author’s work contains no content that was generated by artificial intelligence. If the publisher agrees on a case-to-case basis to permit the use of specified item of AI-generated content in a particular work, the contract should identify exactly what items of AI content are permitted and characterize the permission as an exception to the representations, warranties, and indemnities.

Low-Risk Solutions

Unless and until the law of generative AI is fleshed out by the courts and Congress, the risk-averse publisher may choose to adopt one or more of the following approaches: a) contact the copyright owner of AI-generated content that can be discerned in the work and ask for permission in writing to use it; and/or b) if the copyright owner cannot be identified, or if the owner refuses to give permission, remove the problematic content from the work and replace it with content that originates with the author or with a third party who has granted permission for its use.

---