BitLaw

Summary of the Fair Use Principle

Fair use allows limited use of copyrighted works without permission, acting as an essential balance between creators' rights and the public’s interest in free expression, innovation, and education. U.S. courts analyze fair use claims by applying four statutory factors:

• the purpose and character of the use, including whether the use is commercial or transformative;
• the nature of the copyrighted work;
• the amount and substantiality of the portion used; and
• the effect of the use upon the market or potential market for the original work.

No single factor alone determines fair use, as courts carefully weigh all four factors together. In practice, however, courts often emphasize the transformative nature of the use and its potential impact on the original creator's market as the most significant considerations. A use is often considered transformative when it adds new meaning, message, or purpose to the original work rather than merely superseding or competing with it. Courts typically favor transformative uses, especially when they support innovation, scholarship, or commentary. However, even transformative uses can fail a fair use analysis if they significantly harm the original creator's commercial interests or potential licensing markets.

Importance of the Fair Use Argument for Generative AI

Generative artificial intelligence (AI) models, particularly large language models (LLMs) like ChatGPT, Gemini, and Claude, rely heavily on analyzing massive amounts of textual data. To be effective, these AI models must ingest billions of words to learn language patterns, style, factual content, and contextual information. Because the vast majority of text-based knowledge created in the last century—including books, journalism, source code, and artistic expression—is protected by copyright law, these AI systems inevitably encounter the challenge of training on copyrighted content. In fact, the scale of data required for effective AI training arguably makes the use of copyrighted works nearly unavoidable.

Under U.S. copyright law, any unauthorized reproduction or copying of a protected work constitutes copyright infringement unless a recognized defense applies. Because AI developers routinely make digital copies of complete works as part of their training process, copyright infringement claims have emerged against many AI companies for these types of "intermediate" copying. Without a valid defense, the act of copying books, articles, source code, or artwork for training purposes would be clear infringement.

The primary legal defense being advanced by AI developers is fair use, a doctrine courts have historically applied to new technologies that rely on intermediate copying or indexing of copyrighted materials. The most prominent examples include the fair use decisions concerning Google’s online search engines and Google Books. In the landmark Authors Guild v. Google case, for example, the court allowed Google to scan millions of books to create a searchable index, concluding that Google's use was highly transformative because it significantly expanded public access to information without substituting for the original books themselves. Similarly, courts have permitted Google's image search engine to create and display thumbnail images from copyrighted photographs, finding these uses to be transformative because they provided entirely new functionalities that benefited the public. See Perfect 10, Inc. v. Amazon.com, Inc.

AI developers cite these Google decisions as precedents, arguing their activities are comparably transformative. They assert that training LLMs is akin to creating sophisticated indexing and search tools, with AI systems analyzing large datasets to extract patterns and produce original responses. AI developers contend their copying is a necessary intermediate step, and that their AI outputs do not reproduce protected expression but instead create entirely new content. However, copyright holders strongly dispute these comparisons. They argue that generative AI models differ fundamentally from search and indexing tools because their outputs often compete directly with the original works. As explained below, authors have argued that AI-generated summaries of books can substitute for buying the originals, while news publishers claim AI-generated news digests and factual summaries can replace visits to their websites, thus directly undermining their revenues.

If the fair use defense fails in these cases, the implications for the generative AI industry could be profound. AI developers could face enormous statutory damages for infringement, potentially ranging up to $150,000 per work infringed. Considering that AI training typically involves millions of individual copyrighted works, the potential financial exposure is staggering and could threaten the financial viability of even the most prominent AI companies.

More critically, courts might order AI companies to cease the distribution of any models trained on infringing data, effectively shutting down many popular AI services. Plaintiffs in several major cases, including authors suing OpenAI and visual artists suing Stability AI, have explicitly requested injunctions that would force developers to disable or retrain their models from scratch using licensed or public-domain datasets. Such injunctions would significantly disrupt AI innovation, as obtaining proper licensing for billions of data points would be extraordinarily challenging, costly, and time-consuming.

Furthermore, losing these cases could reshape the AI landscape, prompting companies either to dramatically reduce the scale of their training data or attempt complex licensing schemes. These scenarios could disadvantage smaller companies and open-source AI initiatives, which lack the financial resources to obtain licenses or defend against lawsuits. Conversely, a victory for AI developers could effectively establish that large-scale copying of copyrighted content for AI training is permissible, dramatically altering the balance between copyright protection and technological innovation.

Given the stakes, these fair use disputes are among the most important intellectual property cases of our generation, as their outcomes will shape the future of AI development and define how copyright law applies in the age of artificial intelligence. In the sections below, we will therefore consider in depth the arguments made by both sides in this debate. Because the fair use statute requires the analysis of the four fair use factors specified in section 107 of the Copyright Act before determining whether a particular use is fair, the sections below will follow these same four factors.

Factor One: Purpose and Character of the Use (Transformative Use vs. Competing Use)

This factor asks why and how the copyrighted material was used. AI companies argue that using works to train a generative AI model is highly transformative – the AI “learns” functional patterns or gleans information, rather than simply republishing the work. They analogize AI training to a human researcher reading many sources to develop new ideas. For example, OpenAI and others claim that an AI model extracts knowledge (e.g. language patterns, style, factual information) from the training data and then uses that to generate original output, which is a different purpose than the original expressive purpose of a novel or image. Moreover, many AI outputs are not literal copies of any one work but rather amalgamations or new creations. This, they contend, is akin to Google’s transformative use of images to create searchable thumbnails (which courts found to be a fair use in Perfect 10) or Google’s scanning of books to enable text search (found to be a fair use in Authors Guild). In the code context, defendants note that merely functional uses of code (to learn programming patterns) should be seen as transformative, much like how a clean-room reimplementation of software interfaces was fair use in Google LLC v. Oracle America, Inc.

Plaintiffs counter that AI training is not truly transformative when the end result serves as a market substitute for the original. They emphasize that AI models often compete directly with the original works. Judge Bibas, in Thomson Reuters v. ROSS, underscored that ROSS’s use was not transformative because it used Westlaw headnotes for the same purpose they were created – to provide legal research answers. Similarly, authors and publishers argue that an AI’s use of their text isn’t like a scholar quoting it for commentary, but rather a wholesale commercial exploitation to produce a competing product (e.g., ChatGPT’s summaries competing with book sales, or AI news answers competing with visiting news sites). They contend that an AI telling a user the plot of a novel or facts from an article is essentially performing the same function as the original work (delivering the story or information) – just in a compressed form – which is not a transformative purpose. In short, the plaintiffs say these AI uses usurp the original market rather than add a new meaning or message. They also stress the blatantly commercial nature of AI training--these are for-profit companies leveraging creators’ work without compensation, which weighs against fair use. They will frequently note that the Supreme Court’s 2023 Warhol v. Goldsmith decision warned that even some new uses can fail fair use if they are commercial and too closely substitute for the original’s purpose.

A unique aspect of this debate is the “intermediate copying” argument. In some earlier cases (like Sega v. Accolade or Sony v. Connectix), courts allowed copying of copyrighted material as an intermediate step in a technological process (e.g., copying a console’s software to reverse-engineer it for compatibility) under fair use. AI developers latch onto this analogy, claiming that copying works into memory to train a model is an intermediate technical step to create a new tool (the AI model). They argue that, similar to those cases, the resulting AI model does not contain expressive chunks of the original work in a readily accessible form – it contains abstracted data (model weights), which is more like unprotectable facts or functional data. However, as Judge Bibas noted in Thomson Reuters v. ROSS, this analogy may be weaker when the material copied is expressive content (like prose or images) and the resulting tool can produce content in the same expressive domain. In Sega, copying code was fair use because it was necessary to access functional elements (unprotected by copyright). By contrast, copying thousands of novels may not be “necessary” if the AI’s purpose (producing more text) inherently overlaps with the original expression. Thus, whether AI training is “transformative” is a point of contention. Defendants will frame the AI model as fundamentally different from the inputted works, explaining that it learns the written language from those works but it does not contain those works and it creates new and original material. Plaintiffs, in contrast, will frame the AI’s function as superseding the inputted work by creating directly competitive outputs.

Factor Two: Nature of the Copyrighted Work

Under this factor, courts consider whether the original works are more factual or creative, and whether they are published or unpublished. In the present AI cases, many of the works (novels, code, artwork, songs) are highly creative and published. Plaintiffs emphasize this, noting that using creative expression (fiction, imagery, etc.) as raw input weighs against fair use – especially if it’s used wholesale. In Thomson Reuters v. ROSS, interestingly, the court found the headnotes to be only modestly creative (summaries of facts/law) which slightly favored fair use on factor two. But the court pointed out that this factor rarely overrides the others. In the AI context, defendants may concede that many inputs are creative, but will argue this factor is less important than the others. Some defendants argue that if the works were publicly available on the internet (even for free), then using them for training is less problematic than, for instance, using private manuscripts. However, the creative nature of the copied works means factor two often tilts for plaintiff. This could make fair use harder to establish.

Factor Three: Amount and Substantiality of Use

This looks at how much of the original was used and whether it was the “heart” of the work. Here, AI training typically involves using the entire work (the full text of a book, the entirety of an image) in the training process. Plaintiffs argue that copying works in their entirety weighs heavily against fair use. Indeed, they say AI models rely on mass copying of everything, far beyond any reasonably necessary excerpt or sample. In the authors’ suits, it’s undisputed that entire novels were ingested; in the image cases, millions of full-resolution images were stored and processed. This comprehensive taking, including the core creative elements, is traditionally a mark against fair use.

Defendants respond with two points: (a) copying the entirety was necessary for the transformative purpose – to effectively train a model on stylistic or linguistic patterns, you often need the whole work. They analogize to Google’s image search (where copying entire images to create thumbnails was allowed by Perfect 10, since a partial image wouldn’t serve the purpose) or to Google Books (scanning whole books was allowed by Authors Guild to enable full-text search). (b) They also argue that, although the AI internally ingests the whole work, it typically does not output the work in full. For instance, GitHub Copilot usually doesn’t regurgitate an entire source file from start to finish – it might output a function or a few lines. ChatGPT doesn’t output an entire novel verbatim. Rather, it might summarize the novel or answer questions about it. Thus, from the perspective of what is made available to the public, the amount of any one work shown to a user is small. In Thomson Reuters, the court noted that ROSS did not disclose the headnotes to users, which actually weighed in favor of fair use on factor three. Similarly, OpenAI will argue that ChatGPT never gives a user the full text of a copyrighted book – at most it provides a summary or a snippet, which is a partial and modified use. This interplay is tricky: some judges might focus on the training-stage copying (entire works taken = weighs against fair use), while others focus on the output-stage disclosure (if only small portions ever surface, that may favor fair use).

Plaintiffs counter that even if output is partial, the training itself is infringing – you can’t copy an entire book into your system without permission just because you promise to only output summaries. They stress that the law doesn’t have a blanket exception for “internal” or “intermediate” copies unless a clear transformative purpose justifies it. Moreover, if/when an AI does output a larger chunk (even occasionally), the entire work was available to it, creating a risk of near-total appropriation. Indeed, the news publishers cite statistics of significant overlap between AI output and original text, suggesting that a large amount of protected expression can and does leak through.

Another part of factor three is the analysis of the "substantiality" of the work, which is important when only a portion is copied of a work but that portion forms the “heart” of the work. See Harper & Row, Publishers, Inc. v. Nation Enterprises, where a magazine's quotation of only a portion of Gerald Ford's memoirs before publication was not considered a fair use because the quotations went to the heart of the book. Plaintiffs in the AI context might argue, for example, a novel’s key plot points or a signature artistic style can be considered the book's heart. Summarizing a book might seem like less text, but it gives away the core expression of the story (spoilers and all), potentially impacting the market for the book. Defendants would argue that a summary is not the “heart” in the copyright sense (it’s a different expression of the plot, not the original prose) and that style cannot be owned. This factor will likely hinge on how courts view the locus of the infringement – the training ingestion (whole works) or the typical output (partial or fragmentary).

Factor Four: Effect on the Market for the Original

This factor is often the most crucial factor in any court's fair use analysis. It is difficult for a defendant to win on fair use when their copying has a significant, detrimental impact on the market for the original work. Plaintiffs in these AI cases will strongly assert market harm, typically relying upon one of these theories:

• Direct Substitution: If an AI can produce something that satisfies a user’s need without accessing the original work, it replaces that original. Authors argue that readers might use ChatGPT to get a detailed synopsis of a novel instead of buying the book. News publishers argue that users might get answers from an AI chatbot instead of clicking their articles, cutting off web traffic and ad revenue. Artists fear that clients who might have commissioned them for an illustration could instead use an AI to generate a similar-style image. In essence, the AI output can serve as a competing product.
• Loss of Licensing Opportunities: Even if the AI outputs aren’t a perfect substitute, the very use of the works for AI training is a market that copyright holders are entitled to participate in. For instance, Getty Images offers licenses for AI training – Stability AI didn’t pay for this license but rather used the images for training without permission. Book publishers or authors might have charged a fee to include their books in a training dataset. In fact, there are existing proposals for the collective licensing of literature for AI training purposes. By using these works for training without paying a license fee, AI firms harm the potential market to license works for AI training. Courts have recognized lost licensing opportunities as a market harm.
• Quality and Reputation Harm: Some plaintiffs (like the New York Times) argue that AI outputs mixing or misusing their content can hurt their brand or market. For example, if ChatGPT produces an erroneous summary or a fake news piece and attributes it to the NY Times, that could erode trust in the NY Times’ actual content. Similarly, artists worry that a flood of AI-generated art “in their style” could saturate the market, diluting the uniqueness (and thus market value) of their style. While this is more intangible, it ties into potential market for derivative works – an artist could license their style or do official collaborations, which might be less in demand if knockoffs abound.

Defendants dispute the extent of market harm. They argue that many of these claimed harms are speculative or indirect. For example, OpenAI might argue that there’s no evidence book sales have declined due to ChatGPT – a summary might actually intrigue some readers to experience the full story (similar to how Google Books search was argued to help readers find books, not replace them). Stability AI might point out that an AI-generated image is usually of lower quality or different in character from a commissioned human artwork, so it serves a different segment of the market (someone who’d never pay an artist anyway). Similarly, a ChatGPT summary is not the same experience as reading a novel, and many who use the summary might never have bought the book regardless.

On licensing markets, defendants caution against a theory of harm that is too broad: almost any innovative use could be said to “create a new unlicensed use” of content, but not all such uses are compensable markets under copyright. Fair use often tolerates some loss of potential licensing (e.g., parodies don’t pay license fees even though the original owner would gladly license). The question is whether the AI use undermines a traditional or reasonable market for the original. AI companies might argue that an author has no traditional market to license her novel’s text for inclusion in a general AI model (as opposed to licensing film rights or audiobook rights, which remain intact). Plaintiffs respond that given the scale of AI, licensing for training is a reasonable new market – one that some parties (like Getty) are already pursuing – and that giving AI a free pass would “destroy the seedcorn” by letting tech firms exploit creative content without returning value to creators.

However, certain uses present clearer harm. In Thomson Reuters v. ROSS, ROSS aimed to compete directly against Westlaw, which is Thomson Reuters’ core product in the legal research field. The court found this factor strongly favored Thomson Reuters, noting ROSS intended to “develop a market substitute” and that if such AI tools proliferate, it could undercut the market for curated legal databases.

Artificial Intelligence (AI) Copyright Attorney

Please see Dan Tysver's bio and contact information if you need any AI-related legal assistance. Dan is a Minnesota-based attorney providing AI advice on intellectual property and litigation issues to clients across the country.