<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Nina Brown on Medium]]></title>
        <description><![CDATA[Stories by Nina Brown on Medium]]></description>
        <link>https://medium.com/@nmibrown?source=rss-2bb7449896b1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*NiE5Mq36JD3fwgPfSkc9Ug.jpeg</url>
            <title>Stories by Nina Brown on Medium</title>
            <link>https://medium.com/@nmibrown?source=rss-2bb7449896b1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:09:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nmibrown/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Copyright Battle between Authors and Generative AI Models]]></title>
            <link>https://generative-ai-newsroom.com/the-copyright-battle-between-authors-and-generative-ai-models-b02fcc89aff8?source=rss-2bb7449896b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/b02fcc89aff8</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[legal]]></category>
            <category><![CDATA[fair-use]]></category>
            <category><![CDATA[copyright]]></category>
            <category><![CDATA[copyright-infringement]]></category>
            <dc:creator><![CDATA[Nina Brown]]></dc:creator>
            <pubDate>Tue, 01 Oct 2024 11:01:37 GMT</pubDate>
            <atom:updated>2024-10-01T11:01:37.923Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong><em>An overview of legal arguments from the idea/expression dichotomy to the fair use defense</em></strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tv6WMZsHaOXIoUa6cs3q6A.jpeg" /><figcaption>Image from Unsplash.</figcaption></figure><p>“<em>Rain beat against the courthouse windows as attorney Rachel Summers stood in front of the jury. The air in the courtroom was tense, thick with anticipation. Her client, Paul Jefferson, a mild-mannered software engineer, sat silently at the defense table, his hands clasped together, the weight of the trial etched deeply into his face</em>.”</p><p>This is the opening paragraph from <em>The Silent Witness</em>, a short story about an attorney who defends a man accused of murder, only to uncover a deeper conspiracy that threatens to silence the truth. This gripping legal thriller was written by ChatGPT. And truly, it wasn’t half bad. The characters were compelling, there was a strong plot with an intriguing conflict, and satisfying conclusion. It might have been a story written by a student of John Grisham’s work, or maybe someone who’s read a lot of David Baldacci or Michael Connelly.</p><p>And in a sense, that’s exactly what this story is: a product written by a student of great legal writers. By this point, much has been written about how Large Language Models (LLMs) like ChatGPT <a href="https://www.wired.com/story/how-chatgpt-works-large-language-model/">work</a>. Central to that understanding is the premise that generative algorithms are only as good as their datasets. A large and robust dataset enhances an LLM’s ability to understand and generate human-like text, making it more accurate, versatile, and capable of handling a wide range of queries and topics. So it probably isn’t surprising that books by John Grisham, David Baldacci, and Michael Connelly — among other writers of legal fiction — are <a href="https://admin.bakerlaw.com/wp-content/uploads/2023/12/ECF-1-Complaint.pdf">among the works</a> in the training corpus for ChatGPT. That corpus allowed the LLM to learn the sequence of words and structure of language used in legal stories so that it could generate one itself. Thus, the inclusion of these works and others like them may be one of the reasons that ChatGPT was able to write such a compelling courtroom drama. Unfortunately, it has also led to a real-world courtroom drama.</p><h3><strong>The Scraping Claims</strong></h3><p>John Grisham, David Baldacci, and Michael Connelly are among 18 plaintiffs suing OpenAI (maker of ChatGPT) for copyright infringement. Their claim is simple enough: OpenAI “scraped” the web for data to train ChatGPT and copied works of the plaintiffs for inclusion in this set. “Scraping” is the term used to describe the process of automatically extracting data from websites to create or augment datasets used for training generative AI models. Because the right to copy is an exclusive right of the copyright owner, the plaintiffs (all authors and the Authors Guild) claim that OpenAI violated their rights when it copied their protected works during the training process.</p><p>These authors are not alone. This lawsuit is one of several in the same vein — the rise in generative AI programs has given way to <a href="https://www.bakerlaw.com/services/artificial-intelligence-ai/case-tracker-artificial-intelligence-copyrights-and-class-actions/">more than a dozen lawsuits</a> alleging copyright infringement so far in the past two years on the same basis. News organizations, music publishers, authors, and other content creators have all made essentially the same claim against OpenAI, Stability AI, Microsoft, Google and other AI developers: those companies have scraped large amounts of the web, and in doing so made copies of the plaintiffs’ works, for the purposes of training their models. They assert that this violated their copyrights.<a href="#_ftn1">[1]</a></p><p>Is there merit to these lawsuits? Absolutely. Is it as clear cut as the plaintiffs’ claim? Not one bit. Sometimes copyright infringement is pretty clear cut. A musician uses another artist’s melody as the basis of their song, or a company uses downloaded images from the internet for promotional purposes. But often the determination of whether there has been infringement is more complicated. Not all copying is infringement, even when a defendant <a href="https://www.oyez.org/cases/1993/92-1292">makes a profit</a> from their use of a plaintiff’s work, or <a href="https://law.justia.com/cases/federal/appellate-courts/ca2/13-4829/13-4829-2015-10-16.html">uses a large quantity</a> of it. Whether there has been a copyright infringement is a case and fact-specific inquiry, and rapid innovation in this field has created uncertainties in how copyright law applies in this context in the first instance. In fact, the United States Copyright Office is undertaking a <a href="https://www.copyright.gov/policy/artificial-intelligence/">study</a> of the copyright law and policy issues raised by artificial intelligence which includes an examination of the legal implications of training AI models on copyrighted works.</p><p>This won’t be the <a href="https://blogs.loc.gov/copyright/2017/05/copyright-law-and-new-technologies-a-long-and-complex-relationship/">first time</a> that copyright law has had to respond to a <a href="https://generative-ai-newsroom.com/the-long-history-of-technological-media-innovation-and-copyright-legal-battles-1f283c97c6f7">new technology</a>, and we’ll get a sense of how the law will adapt to this area as the cases make their way through the federal court system. Understanding some of the nuances of copyright law will help both understand and predict how these cases are likely to pan out.</p><h3><strong>The Battle Over What Is Copied</strong></h3><p>The plaintiffs paint this as an open and shut case of copyright infringement. The <a href="https://www.bakerlaw.com/services/artificial-intelligence-ai/case-tracker-artificial-intelligence-copyrights-and-class-actions/">defendants</a> so far have denied that their uses of plaintiffs’ works constitute infringement and instead claim that they are fair uses. The fair use argument is without question the strongest for the defendants, and one that we will return to shortly. But there’s another defensive position that OpenAI has taken that also merits mentioning here.</p><p>In both its court filings and in response to the United States Copyright Office’s request for public comments about the copyright law and policy issues raised by artificial intelligence, OpenAI has suggested that the idea/expression dichotomy should bar plaintiff’s claims. In copyright law, only the expressions of ideas are protected. Copyright law <a href="https://copyrightalliance.org/faqs/whats-not-protected-by-copyright-law/#:~:text=Ideas%2C%20facts%2C%20and%20concepts%20are,as%20a%20database%20of%20facts.">doesn’t protect facts or concepts</a>. This is understood as the idea/expression dichotomy: the expression of an idea can be protected by copyright law, but <a href="https://www.copyright.gov/title17/92chap1.html">not the idea itself</a>. This explains why there can be 20 (or 200) books about the same exact topic. The way the ideas and facts within the book are communicated can be protected, but not those underlying facts and ideas themselves.</p><p>OpenAI has argued that in fact it <em>hasn’t </em>made duplicate copies of plaintiffs’ protected works. Instead, it argues that the model stores only the information from copyrighted works that is necessary for it to “learn” how to respond to prompts with either text or an image. Specifically, <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf">it has argued</a> that the LLM underlying ChatGPT stores only the structure of the language itself, such as the syntax and grammatical function of words (e.g. pronouns, verbs, adjectives, etc.) as well as the probabilistic data about the chronology of words so that its model can accurately predict the next set of words in a given sequence. In other words, OpenAI has argued that it <em>copies only the information from the expression and not the copyrightable expression itself</em>. It has been <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf">clear</a> that “[d]espite a common and unfortunate misperception of the technology, the models do not store copies of the information that they learn from,” but instead as the models “learn” during the training process and become better at predicting the next word, that information is incorporated into the model itself. OpenAI compares it to a person who has read a book and set it down, <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf">explaining</a> that its “models do not have access to training information after they have learned from it.”</p><p>It’s true that generative AI programs don’t copy protected works during the training process in order to reproduce those works. (At least not intentionally.) That’s <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf">not how these generative AI products work</a>. Instead, they are copying the data from the work. How words are arranged to form poems. How pixels are arranged in relation to one another to form an image of a horse. And the argument that copyright does not protect that information — the way words or pixels are arranged — is on the money. But what this argument misses is that <em>at some point</em> during the scraping process, the copyrighted work must be copied at least temporarily in order to process it. Even if only the unprotectable components of the work are retained thereafter, the entire work must be copied in the first instance in order for the LLM training process to be carried out. And that copying is prohibited under copyright law.</p><p>There may be valid normative arguments that this <em>type</em> of copying is very different from what copyright law has traditionally prohibited and thus should be permitted under the law. For example, when someone duplicates a photograph from the internet to a printed piece of paper, the medium changes but the photograph remains the photograph. When a musician uses a melody someone else wrote, it may be accompanied by different words and instruments, but it’s the same melody. When a work is copied as part of the training process for an LLM, the output of the LLM is not essentially the same as the original. Depending on the prompt, the response may bear no relation at all to the works used during training.</p><p>This argument appears in early court filings by defendants in generative AI cases, and it’s something the courts will need to address in the near term. Either way, OpenAI and the other defendants have a much better defense.</p><h3><strong>The Fair Use Defense</strong></h3><p>Even if the courts agreed with plaintiffs that scraping their content constituted making a copy, the defendants can still win based on the fair use defense. Fair use allows limited use of copyrighted material without permissions under certain conditions in an effort to balance the rights of creators with the public’s interest in accessing and using creative works.</p><p>The fair use doctrine is <a href="https://www.law.cornell.edu/uscode/text/17/107">codified</a> in the Copyright Act, and directs courts to consider four factors in making the determination of whether a defendant’s use is a fair use. It is a balancing test that is designed to be fact and context specific and doesn’t give way to bright line rules. And ultimately, only a federal court can make the determination of whether a particular use is a fair use.</p><p>This means that even though legal experts can opine on whether something will or won’t be a fair use based on the four-factor test and relevant precedent, it’s only just a guess until a federal court issues a determination on that specific question. In the context of LLMs, each LLM may be different in the ways it scrapes and incorporates bits of data for training. Those differences may be significant in a fair use inquiry, meaning that scraping or training as a general process won’t necessarily be seen as a fair use (or not) but rather the court will determine whether a particular defendant’s scraping <em>process</em> constituted a fair use.</p><p>OpenAI has <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf">argued</a> that its practice of copying a work to extract its unprotectable aspects constitutes fair use. In support, it has pointed to cases<a href="#_ftn2">[2]</a> where courts have permitted the intermediate copying of a protected work under this doctrine. However, the copying in these cases is distinguishable from scraping other forms of content in important ways.</p><p>Each case cited by OpenAI involved the copying of computer programs. While computer programs themselves are protected by copyright, they often contain many functional elements and ideas that are not protected. (This is not dissimilar from books, where the individual words used are not protected by copyright, but the selection and arrangement of those words may be.) What distinguishes computer programs from books, however, is that they have what one court described as “unique problems” because in order for someone to gain access to the unprotectable elements in the computer programs at issue, they need to copy the <em>entire</em> program. In other words, in the cases cited by OpenAI, there was <em>no other way</em> for the defendants to gain access to the ideas and functional elements embodied in the copyrighted computer program other than through intermediate copying.</p><p>This isn’t the case with scraping content such as books or news articles. In the cases against OpenAI and other generative AI companies, the plaintiffs are creators of works of literature and art that one can easily access without first making a copy of the entire work. This makes them fundamentally different than computer programs that store unprotectable data.</p><p>The critical question is whether courts are willing to treat scraping the same way despite these differences. There is a policy argument that the generative AI defendants will likely make that supports such an outcome: in order for an algorithm to parse the unprotectable elements of these works, it must make a copy of the entire thing. (The fact that it can access the work without copying the work is not, in and of itself, enough to allow it to meaningfully use the unprotectable elements.)</p><p>It’s possible that Congress or the Copyright Office could buy this argument and, in an effort to further innovation, determine as a matter of policy that scraping for model training does not constitute an infringement so long as a copy of the work isn’t retained in the model. (This would track with the aim of the United States copyright laws, which is to stimulate artistic activity and innovation for the greater good.) However, it’s just as likely that these government entities could establish a licensing system for scraping works not in the public domain.</p><p>Either way, there is a stronger fair use argument the generative AI defendants will make. There are certain uses, known as “transformative uses,” that have a dramatic impact on the fair use analysis. A use is considered transformative if it adds to or alters the original work with a different purpose, meaning, or message, rather than simply replicating or adding onto it. Transformative works often provide commentary, criticism, or parody, or offer new insights or understanding that are beneficial to the public. And importantly, highly transformative uses are more likely to be deemed fair.</p><p>Arguably, even if a court agrees with plaintiffs in these cases against LLM’s that scraping includes copying their work in order for the LLM to learn information, the defendant’s use is undeniably transformative. In the case against OpenAI brought by authors referenced above, for example, the defendant trained its LLM not only on the words written by the plaintiffs, but <a href="https://downloads.regulations.gov/COLC-2023-0006-8906/attachment_1.pdf"><em>trillions</em></a><em> </em>of other words. And the LLM used all of those works — including plaintiffs’ — to help it understand how words fit together and work to form higher-level ideas. In other words, it learned from the facts, not the protected expression of those ideas.</p><p>And thus, in this defense of transformative fair use lies a neatly packaged argument: the purpose of an LLM is so transformative that it should not be possible to recognize the plaintiff’s work in the LLM’s output. Almost like they never copied it at all.</p><p>—</p><p><em>Nina Brown is an attorney and professor who teaches at the Newhouse School of Communications at Syracuse University. Her research focuses on the intersection of media law and technology.</em></p><p><a href="#_ftnref1">[1]</a> It is important to note that some plaintiffs have made additional claims of copyright infringement based on the <em>output</em> of generative AI products. This article focuses only on the scraping claims and not on those that depend on AI-generated content.</p><p><a href="#_ftnref2">[2]</a> Assessment Technologies of Wi, LLC v. Wiredata, 350 F. 3d 640, 644–45 (7th Cir. 2003); <em>Sony Computer Entertainment, Inc. v. Connectix Corp.,</em> 203 F.3d 596, 602–08 (9th Cir.2000); <em>Sega Enterprises Ltd. v. Accolade, Inc.,</em> 977 F.2d 1510, 1520–28 (9th Cir.1992).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b02fcc89aff8" width="1" height="1" alt=""><hr><p><a href="https://generative-ai-newsroom.com/the-copyright-battle-between-authors-and-generative-ai-models-b02fcc89aff8">The Copyright Battle between Authors and Generative AI Models</a> was originally published in <a href="https://generative-ai-newsroom.com/">Generative AI in the Newsroom</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>