HomeAdvocacyArtificial IntelligenceAI Licensing: What Authors Should Know

AI Licensing: What Authors Should Know

As you know, artificial intelligence companies illegally copied authors’ books from pirate ebook sites to develop their large language models (LLMs). They did this without obtaining consent from authors or publishers, much less providing them compensation. This means that authors and publishers have no means of controlling uses of their books for training or in outputs that incorporate their works. That has led to the unrestricted use of LLMs to generate copycat books that mimic or incorporate authors’ work. It is crucial that we move AI companies away from the current reliance on fair use to licensing regimes where rightsholders have the ability to say no to use of their works or to engage in licensing on terms and conditions they believe are fair, including restrictions on output use.

For the last two years, the Authors Guild has been fighting the flagrant theft of books and journalism by AI companies on all fronts. We have thus far:

  • Litigated against AI companies for infringing authors’ copyrights.
  • Lobbied for amendments to the copyright law, including laws that would make it clear that AI training on copyrighted works is not fair use and legislation for AI transparency so that copyright owners know when their works are used to train AI and consumers know when material was generated by AI.
  • Developed the soon-to-be launched “Human Authored” certification that authors can use for texts that are human written (save for the use of grammar and minor editing AI apps).
  • Recommended clauses for authors’ book contracts to prevent publishers from using AI or licensing books for AI without permission.
  • Worked with licensing agencies, lawmakers, and publishers to try to steer licensing for AI so that authors’ interests are respected—where authors’ get to decide whether and how to engage in licensing and are fairly paid.

With hundreds of billions of dollars already invested in generative AI, and the U.S. government working to ensure its leadership in AI development (and other countries trying to lure AI developers to their shores), it is clear that we will not stop the further development of generative AI and LLMs. We also know that LLMs need—and will continue to need—high-quality human authored copyrighted books and journalism.

What we can stop, however, is the unlicensed use of copyrighted books and journalism and the subsequent unrestricted use of LLMs to create copycat books. How? By licensing—by enforcing copyright and bringing control over uses back to the rightsholders so that all professional copyrighted works used to train generative AI are licensed and thus controlled. This is true for past use (which is not excused just because they went ahead and did it) as well as future uses. Unlicensed use needs to be converted to controlled, licensed use.

There is naturally some confusion about what licensing would entail, especially as the market develops. A few points to bear in mind:

  • Licensing is a way to enforce copyright. Licenses come with limitations and restrictions, as well as compensation. Just as your publishing agreements are limited to the rights expressly granted and you can expressly withhold or forbid certain uses, so are AI licenses.
  • Replacing reliance on fair use with licensing systems gives authors the choice to not license at all. By establishing licensing as the norm, it will be even riskier for AI companies to use the works that authors have chosen not to make available for licensed uses.
  • When a work is used to train an LLM, it helps to teach the LLM how to write and provides the LLM with the information and expression  in the work. We understand that use is difficult to disgorge today. But LLM owners do have the ability to prevent the LLM from generating outputs that include or compete with your work, for instance by preventing responses to prompts for the actual text of a work or for the use of an author’s style.
  • Different authors have different interests when it comes to AI licensing, and we need to respect and support them all. In our first survey on AI, 65 percent of Authors Guild members said they support a collective licensing system that would pay authors a fee for use of their works in training AI, while 27 percent were unsure. Only 9 percent said they did not support a collective licensing system, with many of those opposed to it not wanting AI to use their works at all. In a follow-up survey, 38 percent of Authors Guild members said that they would be willing to license their work to AI companies to train AI if the author could control whether the company allowed use of the work in outputs and the company offered additional compensation for any substantial, identifiable use of the work in outputs. 33 percent said they would not be willing to license such uses, and 29 percent said they did not know.
  • The markets developing for AI licensing are multifaceted. A work can be licensed without feeding the big LLMs or allowing outputs that use your work. When you license a work for “training-only” that means that the LLM owner must block any uses of your work in outputs—and the license will say that those rights are reserved.
  • Authors can separately license rights just for research and reference uses, such as specific Retrieval-Augmented Generation (RAG) uses (where the AI system queries a database in real-time for information to generate more accurate responses without incorporating the material into the model).

A great deal of licensing has already occurred—dozens of news organizations as well as some academic book publishers have entered into licensing deals with AI companies. Meanwhile startups and established licensing organizations are developing collective licensing solutions. Over the last year, the Authors Guild has consulted with several of them.

Our goal is to foster a licensing ecosystem where authors remain at the center with the power to authorize and negotiate fair compensation for AI uses. We want to get out ahead and ensure that the terms of the licensing deals are fair for authors and that authors are given the tools to determine whether and on what conditions they want to allow use of their works (including the compensation), as well as tools to enforce their rights against those who are using their works without licenses.

Once AI companies are forced to adopt licensing models, authors and publishers can prevent the use of their work by simply not engaging in licensing. And soon they will have tools to detect if any AI models are using their work without licenses and enforce their rights.

Down the road, as AI becomes more sophisticated, it is inevitable that licensing will become standard, and it will become easier to enforce your rights when there are licensed models.  

The Authors Guild recently announced a partnership with Created by Humans (CbH), a company building a licensing platform that will give authors control over licensing and an opportunity to earn revenue. We have taken a position as an official advisor to make certain that terms are fair for authors. In the coming weeks and months, we will announce further details about our work to steer the path for AI licensing, including partnerships with other platforms aligned with our values that are offering services to assist in moving AI to licensed uses and to enforce rights.

This and any future partnerships are part of a multi-pronged approach to ensuring authors’ rights and the future of human created literature are protected. We are working on this from all fronts.

FAQs on AI Licensing  

These FAQs provide more information on AI licensing. We will continue to update them, given how fast AI is evolving.

1)

Absolutely not. Whether or not an author wants to license their works is their own choice. The Authors Guild has stridently advocated for authors to have the final say over whether or not they want their works included in any licensing deals, even where the rights might belong to the publisher. We want authors to have the option to license works on terms they decide if they decide to do so.

2)

In our AI best practices, we ask authors to be mindful of the fact that the current major LLMs were trained illegally on unlicensed content and that many allow users to generate outputs without sufficient restrictions. By converting unauthorized, unlicensed training to licensed uses—as should have occurred to begin with—authors and publishers can impose necessary restrictions on use and obtain fair payment. Authors and publishers can also decide not to engage—to not permit their works to be used at all.

Licensing mitigates if not blunts harms caused by generative AI in this manner. Technological measures exist that AI companies can use to block or track outputs to certain types of prompts, and those need to be adopted. For instance, licenses can be conditioned on the AI company preventing its users from prompting its LLM to generate outputs that compete with the authors’ works, such as book summaries, use of the author’s style or voice, outlines for sequels, and other uses that infringe the author’s copyrights.

The licensing solutions we are exploring and promoting, including through our partnership with Created by Humans (CbH), give authors and publishers control over whether or not they want to permit certain output uses, and if they do, to earn adequate revenue from them.

3)

No. Without licensing, there is rampant unrestricted use of LLMs to mimic authors and their work. That is now occurring. Licensing will curtail unrestricted unlocked output uses because it will give rightsholders the contractual ability to restrict and police output uses. Authors and publishers know best how to monetize (or not) their works, and they will not permit uses that will undermine the market for their own works.

Copyright history has shown us that when legitimate licensed technologies come into the marketplace, they replace the unrestricted, infringing uses that preceded them, filling a market gap. And the sooner the licensed uses are offered, the more likely they are to retain value. For instance, when the record companies delayed in making digital music available to consumers in the 1990s, many music listeners moved to illegal downloading (e.g., Napster and its progeny). By the time the record companies got into the game and started licensing digital music, illegal downloads and streams had become acceptable in certain circles—legitimizing illegal behavior. It resulted in the industry downsizing by more than half. Through a multi-pronged approach of litigation and subsequent licensing, the record industry was eventually able to move users to licensed services (though a little late to make users pay what they used to for music).

4)

The AG believes that authors should have the choice to determine whether they want their books used for AI on a title-by-title basis and should have the ability to decide what uses they wish to license.

An author might determine that for some works, licensing is worth the additional income, and for others, perhaps not. Some authors might also be interested in allowing their works to be used for training with no permitted output uses. Or, they might wish to only allow specific types of research and reference uses, such as specific Retrieval-Augmented Generation or RAG uses (where the AI system queries a database in real-time for information to generate more accurate responses without incorporating the material into the model). Down the road, marketplaces might also develop for certain types of derivate work uses, such as fan fiction apps; we imagine those will be one-off deals with relatively high rates of compensation. Whether to enter into any of those types of licenses or not are assessments every author should have the right to make. When authors decide to license their works, we believe that they should have control over the terms.

5)

AI training use was never contemplated in any publishing agreements until very recently, and rights to AI uses were not expressly granted to publishers. Because AI is capable of taking so much more from an author than just the text of the book, we believe that licensing authors’ works for AI training without their consent is unethical and unfair. Each author should have the right to decide whether their publisher may license out their work for AI use. We believe this is true even where the author entered into a journalism, academic, or textbook contract that provided for the publisher to acquire the entire copyright—because as authors have recently expressed, they may well have decided not to enter into the agreement had they known. That said, some publishers do claim that they possess the rights to license for AI training.

Publishers do however have rights to some types of output uses under most contracts, such as book summaries, abridgments, or personalization (e.g., allowing a user to pay to change the name of a character in a children’s book to their child), and have an interest in protecting against those uses or being compensated for them. Moreover, the publisher often possesses the final copy used to print the book, which may be the preferred copy for the AI licensee.

As such, we recommend consulting with your agent (if you have one) and having an open conversation with your publisher prior to engaging in licensing. For training-only licenses, which are not addressed in past publishing agreements, if the publisher is involved and provides the final text file, we recommend a split of 80/20 or 75/25 (with the higher amount going to the author)—similar to translation and foreign rights. Other splits are addressed by the contract.

The best licensing platforms also will accommodate authors and publishers signing on as rightsholders—recognizing that in some or even many cases, it might be necessary for both the author and publisher to sign off on AI use of a traditionally published, in-print book.

Authors Guild members may consult with the Guild’s lawyers by submitting a legal help request with any questions about their rights. You may also refer to our Model Contract for advice on rights and our model AI clauses.

6)

Currently, AI companies are licensing rights to literary works to train their LLMs, and more recently have also started to license trustworthy literary materials for specific Retrieval-Augmented Generation (RAG) uses for reference purposes to ensure factually accurate outputs. The latter might include the ability to summarize texts or quote verbatim from it. (Such RAG uses implicate rights that fall more squarely within the publisher’s subsidiary rights.) In addition, there is a burgeoning market for using books to create applications for information, explore books interactively, or create fan fiction, as well as to convert text into other formats. Many of these latter derivate work uses may be best negotiated individually, but some may be amenable to collective licensing.

7)

No. Not participating in a license does not give AI companies the right to use the works. That is still infringement.

8)

To date, more than a dozen lawsuits have been filed by authors against AI companies for training their LLMs without permission. The Authors Guild is a named plaintiff in a class action lawsuit against OpenAI and Microsoft (and our lawyers have also filed class action lawsuits against Anthropic and Meta). The basis for the lawsuits is that the books used were not licensed—the AI companies did not bother to get permission. The Authors Guild strenuously disagrees with their claims that the training is fair use. Unlicensed, unrestricted AI training could destroy the ecosystem for books in the long run, and copyright exceptions do not extend to undermining the very purpose of copyright law.

Compensation for uses covered by the class action cases will be provided through the class action process—assuming the cases are successful.

9)

The fees will be set on a per-license basis and there are no standard fees to date. The platforms that we have seen incorporate different ways to set prices, which authors will then be able to choose or reject. Rather than basing the pricing on what AI companies have offered to pay, the licensing platforms that we have spoken to are using econometrics to recommend reasonable bandwidths for pricing, including pricing based on usage in AI prompts. From any income generated, the licensing platforms will take a small fee to cover operations. Your agent may also be entitled to a commission and your publisher a portion of the revenue.

10)

We are advocating for authors’ prerogatives in all cases to allow or withdraw from licensing deals, regardless of the terms of their prior publishing agreements—which in no way contemplated AI use. For all new contracts, we recommend that authors include express language that they have the right to approve AI uses and to receive a fair share of any licensing deals. We continue, as we have for the last 122 years, to work to ensure that authors have control and secure fair compensation when their works are used.

11)

If you are an Authors Guild member, our legal team would be happy to review your contract and assist you in understanding and negotiating an AI licensing deal. You may submit a legal request for assistance.

Revised October 21, 2024