Image: Shutterstock
For those of you who may have missed it, this is “Fair Use and Fair Dealing” week, sponsored once again this year by the Association of Research Libraries in the US and the Canadian Association of Research Libraries in Canada. I have written blog posts on fair use/fair dealing for the past couple of years, (here, and here), trying to provide some perspective on the topic given the annual “celebration” of this feature of copyright law by the library community. Yes, it is an important feature of copyright, a defence to infringement in the US and declared to be a “user’s right” by the Supreme Court in Canada. It is an essential part of the copyright landscape, a limitation on the exclusive rights an author enjoys over their work, and is designed to allow reasonable, permissionless and uncompensated (i.e. unlicensed) use of copyrighted works in order to achieve and preserve certain public policy objectives, such as the encouragement of learning and dissemination of knowledge.
Fair use and fair dealing also enable some essential features of a modern democracy, for example by allowing news reporting and artistic criticism to function effectively as well as by facilitating some other specified uses, such as parody and satire. All this has been baked into copyright law for a century or more in Canada and the UK and for longer than that in the US through jurisprudence. However, while fair use and fair dealing permit a relatively wide range of uses of copyrighted materials without licensing, there are limits set by legislation and the courts with regard to the amount of content used and the purpose to which it is put. Each time a new technology comes along, those limits are tested. And AI is severely testing those limits today.
AI is not the only challenge faced by rights-holders as a result of the ambiguities of fair use and fair dealing. The continued uncompensated and unauthorized widespread use of copyrighted and licensable materials by Canada’s education sector, particularly post-secondary institutions, under the pretext of education fair dealing is one current example. The unlicensed and unauthorized digital scanning of copyrighted published works under the invented, contentious and ultimately legally unsustainable theory of “controlled digital lending”, which I wrote about here and here, is another. AI, however, is the unlicensed use where the financial stakes are the highest, running into billions, if not tens of billions of dollars. It is remarkable that the US high tech industry and generative AI developers have tied their substantial investments to a very shaky interpretation of fair use. In effect, they have admitted to permissionless (i.e. unauthorized/unlicensed) reproduction of copyrighted works by resorting to a fair use defence. If there is no potential infringement through unlicensed reproduction and distribution (i.e. no copy has been made, or the copying is de minimis or does not reproduce the protectable expression of a work), there is no need to invoke fair use, yet that is what OpenAI, StabilityAI, Midjourney and other generative AI developers that are being sued by rights-holders (ranging from individual authors and artists to the New York Times, Getty Images and Universal Music) have done. What is the basis of this fair use defence? From the perspective of the generative AI developers, it is that the infringing uses are “transformational”, producing something new that is not a derivative nor a commercial substitute for the original.
This is a line of legal argument that has some traction in the US. While not enshrined in any US legislation, the “transformation doctrine” has nonetheless been relied on by US courts in recent years to justify fair use rulings, (when combined with consideration of the other three factors used in the US to determine fair use). One of the most notable of such transformative use decisions was the Google Books case (Authors Guild, Inc v Google Books, Inc), where a US court found that Google’s unauthorized scanning of the plaintiff’s copyrighted works was a fair use because the sampling index it produced was different from the original works and did not directly compete with them commercially. This is the line of argumentation the AI developers are counting on to justify their unauthorized ingestion, through reproduction, of hundreds of thousands of copyrighted works. But will this carry the day? The situation today looks a lot less certain than when the Google Books case was decided in 2015.
The Association of Research Libraries (ARL), the sponsor of fair use week in the US, has gone on record to declare that, in its view, training generative AI models on copyrighted works is a fair use. (The librarians should be careful what they wish for as AI has the capacity to do as much damage to the library sector as it threatens to do to authors.) The ARL cites the Google Books case as well as others in support of its position. However, there are other more recent cases that make it far from obvious that this transformational use argument will prevail in the case of AI, or at least in some AI cases. Two recent cases in particular have cast doubt on the “open season” on creators that overly broad interpretations of the transformation doctrine have brought about in recent years.
The first was the Internet Archive v Hachette case in 2023 (now under appeal) that dismissed the arguments of the Internet Archive that its unauthorized scanning (copying) of a number of copyrighted works (127 in total, represented by Hachette and three other publishers), and the subsequent lending of those works in digital format, was a transformational fair use. While the court’s decision in favour of the publishers was not based exclusively on a rejection of transformative use for the works in question, this was a big part of the decision. The second case, also decided in 2023, (on appeal to the US Supreme Court), Warhol Foundation v Goldsmith, resulted in the Andy Warhol Foundation being held responsible for infringing the copyright of photographer Lynn Goldsmith by licensing one of Warhol’s remakes of her iconic photograph of musician Prince as a magazine cover. The Warhol Foundation had sought a ruling that Warhol’s use of Goldsmith’s photo to create a series of coloured silkscreen artworks, including the “Orange Prince” print that was licensed as a magazine cover, was transformational and thus a fair use. While the district court had originally found in favour of Warhol, this was overturned on appeal and sustained by the Supreme Court, which held that the licensing of Warhol’s “Orange Prince” work (based directly on Goldsmith’s photograph of Prince) was not transformational because it substituted for the original in the commercial market.
The same could be said of an AI generated image or work that is substantially similar to an original work it has been trained on, (i.e. it incorporates protected elements of a copyrighted work), and which then substitutes for or competes with the original commercially.
The Supreme Court’s Warhol decision seems to have sent a chill through the world of appropriation art, as I noted in a recent blog regarding the willingness of appropriation artist Richard Prince (no relation to the musician, who died in 2016) to reach what amounted to a public settlement (final judgment) over a longstanding lawsuit with two photographers whose works he had appropriated to produce what amounted to derivative works. Artist Prince agreed to conditions that prevent him from “reproducing, modifying, preparing derivative works from, displaying, selling, offering to sell or otherwise distributing” the works based on the plaintiff’s photographs, while paying the photographers five times what he had earned from selling the derivative works, plus covering the plaintiff’s legal costs. Total payments amounted to almost a million dollars. This case was significant, and could well have ramifications for pending AI lawsuits.
The AI industry has staked a lot on a very thin reed, particularly given the welcome swing of the fair use pendulum back to a more balanced position by US courts. Uttering the words, “transformative use” is no longer an automatic get-out-of-jail-free card when works are copied holus-bolus. While it is true that not all generative AI outputs reflect or are substantially similar to the works they were trained on, sometimes they are, as we have seen in the New York Times v OpenAI case. If the AI output is very similar to the original work, and substitutes in the market for it (e.g. artwork promoting….in the style of….), what is transformational about that?
A lot can depend on what prompts are entered, as OpenAI is arguing. But that is putting the responsibility for the infringement on the user, i.e. the human creating the prompts, rather than on entity that did the infringing in the first place by reproducing the copyrighted work without permission. Moreover, it is not enough to say, as OpenAI has argued, that there was a “bug” in their system that they will fix in future iterations. Past infringement is still infringement. You can’t just say, “Sorry, I won’t do it again”. If you infringed, you infringed and there is no guarantee that it won’t happen in future. If the product wasn’t ready for prime time, it should not have been launched on the public.
While neither of the previous cases I have referred to are dispositive of all the issues, and each case must be decided on its own merits, there is nonetheless a compelling trend suggesting that the “transformative use” free ride is coming to an end. If I were an AI developer who had staked hundreds of millions of dollars on the shaky premise that my unlicensed appropriation of copyright protected expression would be found to be a fair use, I would be worried, very worried.
The key message is that while fair use is an important facet of copyright law, it has its limits. This is an important message worth repeating during fair use week. Those limits are being tested by AI developers who have been following a “ask forgiveness after rather than permission before” approach. They may get an unpleasant surprise.
In sum, when it comes to fair use and fair dealing, let us by all means acknowledge its importance, but also remember the other side of the coin. Its limits. What exactly fair use means in the age of AI remains to be determined by the courts, but it is important to remember that it is not a blank cheque based on some mystique of transformation. Along with fair dealing, it is a set of exceptions that gives users plenty of scope for unlicensed uses–but is not a licence for unlimited free-riding and unfair competition with the copyrighted output of creators. That’s a good takeaway for Fair Use and Fair Dealing Week.
© Hugh Stephens 2024. All Rights Reserved.
This post has been updated to clarify that invocation of fair use is a de facto admission of unlicensed reproduction and distribution of a copyright protected work, and thus a potential infringement.
