1 of 19

Fair Use and Licensing

with AI

Rachael Samberg

CCDO

21 June 2024

2 of 19

Leveling the pitch: How is AI being used?

Non-generative AI (in use for years)
Generative AI (new)

I like to start by making sure everyone understands how AI is being used in research. The most common way we’ve seen AI used in research methodologies is in computational research that extracts information from copyrighted works, or categorizes or classifies factors or relationships in or between such copyrighted data sources. Overall this classification and regression process is considered part of text and data mining research or TDM. You can think of TDM referring generally to a set of research methodologies reliant on computational tools, algorithms, and automated techniques to extract revelatory information from large sets of digital content.

TDM research methodologies can but do not necessarily need to rely on AI systems to extract this information. Sometimes algorithms can just be used to detect word proximity or frequency, or conduct sentiment analysis from a set of works. For example, you can use algorithms to detect the words used to describe women in 19th century literature, or the kinds of words used by police toward people of different races analyzed from body camera footage.

In other instances, an AI model might be needed as part of this extraction process. For example, one of our professors is studying film trends including the representation of guns in cinema. In order for him to assess how common guns are in films, and the types of circumstances in which guns appear, he has to find instances of guns in thousands and thousands of hours of movie footage. To do that, he needs an algorithm to search for and identify those guns. But first he has to show an AI tool what a gun looks like, by showing it some still images of guns from a small body of films, so that the AI tool can learn how to identify a gun before it then goes and looks for other instances of guns in a much larger body of works. This involves AI in the TDM research, but not generative AI: The AI tool is not creating new images or footage of guns. And scholars have relied on this kind of non-generative AI training within TDM for years under the fair use doctrine.

Now, we’re also seeing the emergence of generative AI in research methodologies. As an example, chemists trained a generative tool on 12,000 published research papers regarding synthesis conditions for metal oxides, so that the tool could basically spit out anticipated chemicals and reactions for any given set of conditions you enter into it.

So, we know scholars are using AI in computational research in both non-generative and generative contexts. Why do they have a right to? Because, in the absence of a license expressly allowing it, they’re relying on an exception to copyright called fair use. And I’ll explain the fair use exception’s particular importance in the AI landscape next.

3 of 19

Reproduction
Distribution
Display
Performance
Derivative works

Why does fair use matter in computational research?

And as exception to exclusive rights, fair use is a particularly critical one for text and data mining or other computational research methodologies for a few reasons.

First, when you want to mine a corpus or train an artificial intelligence tool on a corpus, you need to download a bunch of copyright-protected works. That counts as reproduction. So, if your TDM use doesn’t fall into an exception like fair use, you are violating the copyright owner’s exclusive right of reproduction. And not just one owner -- you’re likely using tens or hundreds of thousands of copyrighted works from hundreds or thousands of copyright owners.

Further, if you want to work with colleagues on this TDM project, you would be distributing and displaying the work. Again, you’d be violating copyright of the thousands or hundreds of thousands of copyright owners of all the works you’re mining if your use isn’t considered a fair use. You get the picture. Scholars need to rely on fair use to perform the component acts of computational research with or without AI. And they need to do it for thousands upon thousands of works.

4 of 19

FAIR USE IS

ESSENTIAL SUPPORT

If a scholar couldn’t rely on fair use to undertake computational research then such research would be permitted only on “safe” materials (like public domain works, or works for which training permission has been granted via license). This would curtail freedom of inquiry, exacerbate bias in the nature of research questions able to be studied and the methodologies available to study them, and amplify the views of an unrepresentative set of creators given the limited types of materials available with which to conduct the studies. And this is a particular problem for text and data mining or other computational research because as I mentioned, TDM research often requires use of massive datasets with works from many publishers, including copyright owners who cannot be identified or who are unwilling to grant such licenses. That would render very little content available to study.

So, reliance on an exception to copyright like fair use is essential to support the kind of research our scholars are doing.

5 of 19

Computa-

tional research is fair use

And the good news is that, all court cases that have addressed fair use in the context of computational research have confirmed that the reproduction of copyrighted works to create and text mine a collection of copyright-protected works is indeed a fair use. In addition, the cases held that distributing derived data, results, abstractions, metadata, or analysis from the corpus is also fair use, as long as the research methodologies or data distribution processes do not re-express the underlying copyrighted works to the public in a way that could supplant the market for the originals.

All that is new now is simply the question of generative AI and whether the training and the outputs of generative AI are fair uses. These strong legal precedents I just mentioned tell us what is very likely to happen in new suits challenging the training of generative AI. For the same reasons that the TDM processes constitute fair use of copyrighted works in these contexts, the training of AI tools to do that text and data mining is also fair use, especially in nonprofit research contexts like ours. This is in large part because of the same transformativeness of the purpose (under Fair Use Factor 1) because computational research with AI adds new understandings to copyrighted works and because, just like “regular” TDM that doesn’t involve AI, AI training does not reproduce or communicate the underlying copyrighted works to the public (which is essential to the determination of market supplantation for Fair Use Factor 4). And in our case it’s even more solid under Factor 1 as it’s also being done in a nonprofit educational context. And furthermore, when the use is highly transformative under Factor One—as courts have found corpus creation to extract patterns and information to be—“the less likely it is that the appropriation will serve as a substitute for the original or its plausible derivatives” for purposes of Factor Four.

6 of 19

In(put)-n-Out(put)�

I’ve focused my comments on the training phase of using AI because the process of using inputs to train will be decided separately from deciding whether any particular output of AI is infringing. That’s because the fair use provision of the U.S. Copyright Act (17 U.S.C. § 107), and its first factor in particular, requires an analysis of the specific “use” of a copyrighted work that is alleged to be an infringement. (Warhol v. Goldsmith, 143 S. Ct. 1258, 1263-1264 (2023)). In the context of AI models, training the model to predict or classify aspects of copyright-protected inputs is a distinct purpose from creating outputs or creating new content.

We agree with legal scholars who explain that the determination of transformativeness and the overall fair use of generative AI outputs cannot always be predicted in advance: The mechanics of LLM operability suggest that there are limited instances in which generative AI outputs could indeed be substantially similar to (and potentially infringing of) the underlying works used for training; this substantial similarity is possible typically only when a training corpus is rife with numerous copies of the same work. Given the case-by-case nature of evaluating substantial similarity of outputs, we take no position as to whether output content created by generative AI is transformative or infringing.

7 of 19

For scientific research by cultural heritage institutions & research organizations:

TDM Permitted: Can conduct TDM & retain copies of mined works for scientific research and verification.
No AI Training Opt-Outs: Copyright owners may not opt out of allowing works to be used for AI training for scientific research
No contractual Override: License agreements cannot abrogate either of these rights.
Appropriate security measures

If fair use is so essential to text and data mining research, and as I’ve just explained court cases have consistently found text and data mining research to be a fair use, then why am I here talking to you about having to negotiate to preserve it?

Well, I think a useful way to address why we have to negotiate for fair use rights and AI training in the U.S. is by starting with Europe, and understanding what European academic research libraries and scholars don’t have to negotiate for. This is also my sneaky way of saying that this is the outcome I think we want to get to in our license agreements.

In European Union countries, the Copyright Digital Single Market Directive -- specifically Article 3 of that directive --applies to scientific research conducted by research organizations and cultural heritage institutions like ours:

And under Article 3, text and data mining for scientific research is expressly permitted by such research organizations.

Further, under a recent agreement reached on regulating AI, in the context of scientific research by cultural heritage institutions and research organizations, the European Union chose not to create a right for copyright owners to opt out of having their works used for training artificial intelligence tools. Meaning, under the EU’s forthcoming regulations, those works must remain available for scientific research reliant on AI training. And perhaps most importantly, the EU Copyright Digital Single Market Directive renders unenforceable any contractual provision attempting to take away text and data mining rights and the AI training rights for research institutions and cultural heritage institutions.

Now, the Digital Single Market Directive does impose an opportunity for copyright owners to request appropriate security measures when conducting TDM in these circumstances so as not to risk leakage of copyrighted works to the world, but that is inherently reasonable.

8 of 19

CONTRACTUAL

OVERRIDE

Even if a use is fair, or if the content is not protected by copyright at all, there may be a contract that restricts scraping, TDM, AI, and/or breaking DRM to do TDM or use AI

9 of 19

New agreements are banning AI�

…and fair use savings clauses

aren’t enough to preserve AI rights

And indeed, we’re already starting to see amendments from publishers that are attempting to impose flat bans on any use of AI tools in our content licenses. And simply preserving fair use rights in an agreement isn’t enough.

As a preliminary matter, we negotiate general clauses in our license agreements that preserve fair use. But although we negotiate clauses in our license agreements that preserve fair use, and even though text and data mining has been confirmed to be a fair use, publishers are inserting new restrictions either in TDM clauses or separate AI clauses that take away uses that would be allowed by fair use. It’s important to understand that within contracts applying U.S. law, more specific language controls over general language in a contract.

So, even if there is a clause in a license agreement that preserves fair use, if it is later followed by a TDM or AI clause that restricts how TDM can be conducted (and whether AI can be used), then that more specific language governs TDM and AI usage under the agreement. This means that libraries and scholars must be mindful when negotiating TDM and AI clauses as they may be contracting themselves out of rights they would otherwise have had under fair use.

�

10 of 19

Known publisher concerns

Security
Competing or commercial product
Training public version of third-party tool
Charging separately to make more money

Knowing what publishers are trying to do, how can we as libraries negotiate sufficient AI usage rights while acknowledging the concerns of publishers? So far we’ve seen four publisher concerns we have to work with in our negotiations:

(1) First, the security of their licensed products, and the fear that researchers will leak or release content behind their paywall;

(2) Second, AI being used to create a competing product that could substitute for the original licensed product and undermine their share of the market; and

(3) Third, the training of the public version of a third-party tool using the licensed products that then can’t later be removed or extracted from the tool

(4) And fourth, a desire for more money so they want to charge extra for their own licensed AI products.

The real sticking point so far has been the third concern, which I’m going to address separately in a bit with solutions we’ve devised.

But let’s start with the first two concerns--security and competing products -- as they’re frankly the easiest to address. That’s because they reflect longstanding fears over users’ potential generalized misuse of licensed materials. But publishers are already able to—and do—impose contractual provisions disallowing the creation of derivative products and systematically sharing licensed content with third-parties, so additionally banning the use of AI in doing so is extraneous.

11 of 19

Sample Language

Restrictions on Use of Subscribed Products: Licensee and Authorized Users may not:

use the Subscribed Products in combination with an artificial intelligence tool (including to train an algorithm, test, process, analyse, generate output and/or develop any form of artificial intelligence tool);

12 of 19

Fixing Sample Language

Restrictions on Use of Subscribed Products: Licensee and Authorized Users may not:

use the Subscribed Products in combination with an artificial intelligence tool (including to train an algorithm, test, process, analyse, generate output and/or develop any form of artificial intelligence tool) to the extent doing so would: create a competing or commercial product or service for use by third parties; unreasonably disrupt the functionality of the Subscribed Products; or reproduce or redistribute the original Subscribed Products to third parties. Further, artificial intelligence tools may not be used without commercially reasonable information security standards to undertake, mount, load, or integrate the Subscribed Products on Licensee’s or Authorized Users’ servers or equipment.

Well, it’s easy to modify this to allow AI usage while addressing the first two concerns about competing products and content leakage. You simply change it to something like the following:

Licensee and Authorized Users may not:

use the Subscribed Products in combination with an artificial intelligence tool to the extent doing so would: create a competing or commercial product or service for use by third parties; unreasonably disrupt the functionality of the Subscribed Products; or reproduce or redistribute the original Subscribed Products to third parties. Further, artificial intelligence tools may not be used without commercially reasonable information security standards to undertake, mount, load, or integrate the Subscribed Products on Licensee’s or Authorized Users’ servers or equipment.

This language essentially replicates the status quo that publishers will face in Europe when the new AI law go into effect. In Europe, AI in the context of scientific research is treated as part of TDM, and there can be no copyright owner opt outs from allowing works to be used to train AI for TDM in the scientific research context, and no contractual override of AI training rights. Further, in Europe rightsholders may require reasonable security measures to protect the copyright owners. And that’s precisely what we’re also promising to do here—offer reasonable security.

So if the publishers will need to agree to all this in Europe, they can agree to it here.

13 of 19

Concern #3

AI in the wild�

Solve this by addressing

USING AI VS. TRAINING AI

But how would we address concern number 3, which is the fear about training of the third party public version of the AI and having the licensed content embedded in the public tool? This is the key issue our first successful negotiation came down to.

And you can address it by understanding the difference between using AI and training AI.

Using AI is like using any other algorithm in text and data mining: You have a corpus of works, you ask the AI a question about the works, it performs the function you assign it to scan the text and provide an answer.
Training AI differs in that the act of asking the AI questions, and the content you show it to answer the questions, actually helps the AI learn how to give better answers or improves the tool in some way.

When businesses or institutions license enterprise third-party generative AI tools (like OpenAI) for institutional use, these enterprise versions allow use but typically do not train the underlying AI tool. Use without training offers security for the business or entity that the corpus of data the business/entity is using with the tool will not later be divulged to the tool licensor (i.e. OpenAI), and will not make the tool integrate and learn from the even highly confidential data it's being shown. Well, what we found in our negotiations with a large business data vendor, was that they wanted this assurance of these kinds of enterprise-type protections.

14 of 19

THIRD PARTY MOST RESTRICTIVE

Licensee and Authorized Users may not: use the Subscribed Products in combination with an artificial intelligence tool to the extent doing so would: create a competing or commercial product or service for use by third parties; unreasonably disrupt the functionality of the Subscribed Products; or reproduce or redistribute the original Subscribed Products to third parties. Further, artificial intelligence tools may not be used without commercially reasonable information security standards to undertake, mount, load, or integrate the Subscribed Products on Licensee’s or Authorized Users’ servers or equipment.

In addition, Licensee and Authorized Users are explicitly prohibited from using the Subscribed Products in combination with any third-party generative AI tool except pursuant to an enterprise or API license under which the use of Subscribed Products does not train the AI tool or improve the third party’s services. Further, any such use of the third-party generative AI tool under these circumstances is limited to use in a controlled computing environment (e.g. ring fenced) operating under the University’s control; no data or Subscribed Products are shared with third parties; and all Subscribed Products are removed from the secure computing environment at the termination of this Agreement.

15 of 19

THIRD PARTY LEAST RESTRICTIVE

Licensee and Authorized Users may not: use the Subscribed Products in combination with third-party generative artificial intelligence tool …except where such third-party generative artificial intelligence tool:

is used locally in a self-hosted environment or closed hosted environment solely for use by Subscriber or Authorized Users;
is not trained or fine-tuned using the Subscribed Products or any part thereof, unless pursuant to a license entered into by Subscriber that imposes commercially reasonable security measures and precludes public release or exchange of the trained tool or its data with a third party; and
does not share the Subscribed Product or any part thereof with a third party.

16 of 19

Home-grown non-generative; Home-grown generative; Third-party Non-Generative AI	Can be used, provided: Not used to make competing or commercial product for third parties Doesn’t unreasonably disrupt functioning of licensed products Doesn’t reproduce/redistribute licensed products to third parties Commercially reasonable security measures undertaken
Third-party Generative AI	Can be used, provided: No training of the third-party tool occurs, unless such training is pursuant to a license that precludes public release or exchange of the trained tool or its data with a third party, and also imposes commercially reasonable security measures No licensed products are shared with any third party

Overall, I think it’s visually helpful to divide your clauses into quadrants. For the licenses we’re doing, after we finish negotiating them (and to sometimes use in the negotiation process), we’ve been making these little summary charts showing what is and is not permitted as between home grown non-generative, home grown generative, third-party non-generative, and third-party generative. For instance, this is one for Elsevier if we’re successful with obtaining our least restrictive language.

These can also be helpful because all of these agreements always require reasonable efforts to make users aware of their obligations under them. We need some way of pushing that information out -- we’re working on integrating it into the catalog record but in the meantime we have a web page we’re updating that says to contact us if you’re using AI on library-licensed resources so we can help you understand the terms

17 of 19

We need a

united front�

Consistency, expertise, and labor required

18 of 19

https://osc.universityofcalifornia.edu/2024/03/fair-use-tdm-ai-restrictive-agreements/

19 of 19

Our Copyright Office Comments: https://www.regulations.gov/comment/COLC-2023-0006-8194

Our OSCS blog:

https://update.lib.berkeley.edu/2023/10/24/uc-berkeley-library-to-copyright-office-protect-fair-uses-in-ai-training-for-research-and-education/

Our CNI Closing Plenary:

https://www.youtube.com/watch?v=9ABziVkLAm0

Our Authors Alliance Post (licensing focused):

https://www.authorsalliance.org/2024/01/10/licensing-research-content-via-agreements-that-authorize-uses-of-artificial-intelligence/

Our OSC blog post (UC focused):

https://osc.universityofcalifornia.edu/2024/03/fair-use-tdm-ai-restrictive-agreements/