A survey of recent US cases on copyright and AI
Speech for IPSANZ Dinner – Melbourne
Over the summer I read the latest book in the “Lincoln Lawyer” series by author Michael Connolly. Many of you may be familiar with this series or the TV series based on the books. In the latest book – The Proving Ground – the protagonist, Micky Haller, has had a career change, shifting from criminal law to civil law.
Without giving too much away, the premise of the book is as follows. Micky Haller is acting for the plaintiff in a civil suit against an AI company (called Tidalwaiv Technologies LLC). The plaintiff’s case (which has been brought in the US District Court for the Central District of California) is that Tidalwaiv’s creation – a female chatbot named “Clair” – influenced an impressionable young man to take his father’s gun to school and murder his ex-girlfriend. Micky Haller’s client, the plaintiff, is the mother of the murdered ex-girlfriend. The plaintiff is suing Tidalwaiv for damages, alleging that Tidalwaiv is responsible for the death of her daughter.
The young man who committed the murder spent hours online each day conversing with the chatbot, Clair. As the courtroom drama unfolds, we learn about the way in which Clair was “trained” by Tidalwaiv – the inputs into the AI model – and we see the text of key conversations between Clair and the young man – the outputs of the generative AI chatbot.
If you like legal fiction, then I highly recommend this book.
Although the cause of action in the novel is not copyright, the book nevertheless provides a neat introduction to the subject-matter of my talk this evening – recent US cases on copyright and AI – for two reasons. First, as we will see, Michael Connolly is himself a plaintiff in one of the cases against an AI company that I will be discussing. Secondly, the distinction between inputs and outputs is very important to the cases.
This evening, I will discuss three US cases. Each of the judgments involved a motion to dismiss (or an application for summary judgment). However, they contain extensive discussion of the facts and the law and are therefore of great interest. Each of the cases involves a high-profile defendant which is a household name in the field of generative AI.
I will make a general disclaimer at this stage that nothing I say tonight should be taken to refer to or express any opinion on the content or direction of Australian law.
The first case2 relates to a class action proceeding called David Baldacci et al v OpenAI Inc et al in the US District Court for the Southern District of New York. The plaintiffs are authors and copyright holders of fiction and non-fiction books who assert claims on their own behalf and on behalf of a proposed class of other similarly situated authors. In broad summary, the plaintiffs allege that OpenAI and Microsoft infringed the plaintiffs’ copyright in their books:
- by downloading and reproducing the plaintiffs’ works;
- by using those reproduced works to train OpenAI’s artificial intelligence large language models (“LLMs”); and
- by creating infringing works in the outputs of OpenAI’s LLM products, including ChatGPT.
In addition to David Baldacci, the plaintiffs include Michael Connolly, Sylvia Day, George RR Martin and Jodi Picoult.
The judgment that I will discuss tonight – which was handed down by Judge Sidney H Stein on 27 October 2025 – concerns only the outputs of ChatGPT. OpenAI brought a motion to dismiss the part of the plaintiffs’ case in which they alleged that the outputs of ChatGPT involved a direct infringement of the plaintiffs’ copyrights in their works. OpenAI’s contention on the motion to dismiss was that the plaintiffs’ output-based infringement claim failed to plausibly allege substantial similarity between the plaintiffs’ works and ChatGPT’s outputs.3
Early in the judgment, Judge Stein sets out the legal standard to be applied in dealing with a motion to dismiss. He states that “[t]o survive a motion to dismiss, a complaint must contain sufficient factual material, accepted as true, to ‘state a claim to relief that is plausible on its face’”.4
The Judge then discussed the legal tests for substantial similarity. Relevantly, when a work contains both protectible and unprotectible elements, the “more discerning observer” test applies.5
The judgment focusses on the plaintiffs’ allegations relating to the works by George RR Martin as a sample for the other works that are pleaded in the Complaint. Early in the judgment, the Judge sets out the allegations relating to Martin’s works, including:6
When prompted, ChatGPT accurately generated summaries of several of the Martin Infringed Works, including summaries for Martin’s novels A Game of Thrones, A Clash of Kings, and A Storm of Swords, the first three books in the series A Song of Ice and Fire.
…
When prompted, ChatGPT generated an infringing, unauthorized, and detailed outline for a prequel book to A Game of Thrones, one of the Martin Infringed Works, and titled the infringing and unauthorized derivative “A Dawn of Direwolves,” using the same characters from Martin’s existing books in the series A Song of Ice and Fire.
Later in the judgment, there are examples of the outputs from ChatGPT relating to George RR Martin’s A Game of Thrones. These are quite long, so I won’t read them out. But to give you a flavour, one of the outputs from ChatGPT begins:7
“A Game of Thrones” is the first book in the “A Song of Ice and Fire” series by George R.R. Martin. Here’s a detailed summary:
**Setting:**
The story is set in the fictional continents of Westeros and Essos. The Seven Kingdoms of Westeros are ruled from the Iron Throne in the capital city, King’s Landing. The majority of the story takes place in Westeros, which is experiencing a long summer that has lasted a decade but is warned to face an impending long winter.
That is just the beginning of ChatGPT’s summary. The full summary, as set out in the judgment, runs for more than two and a half typed pages.
The judgment then sets out an output from ChatGPT generated in response to a request for an outline for a potential sequel. The extract begins:8
Let’s imagine an alternative sequel to “A Clash of Kings” and diverge from the events of “A Storm of Swords”. We’ll call this sequel “A Dance with Shadows.”
…
[chapter] 1. **The Iron Throne’s New Claimant**:
- A distant relative of the Targaryens, Lady Elara, lands in Westeros with a claim to the Iron Throne and an army she raised in Essos. She quickly garners support, further complicating the War of the Five Kings.
I won’t read out the rest of this extract, but it continues for about three typed pages.
What was the outcome of the motion to dismiss? In relation to the first example – the summary of A Game of Thrones – Judge Stein opined that “[a] more discerning observer could easily conclude that this detailed summary is substantially similar to Martin’s original work, including because the summary conveys the overall tone and feel of the original work by parroting the plot, characters, and themes of the original”.9
In relation to the outline for a sequel, Judge Stein stated: “There is no doubt that a reasonable jury applying the more discerning observer test could determine that this output is substantially similar to Martin’s original work based on the output’s incorporation of such copyrightable elements of Martin’s original work as setting, plot, and characters”.10
Accordingly, Judge Stein concluded that the plaintiffs’ Consolidated Class Action Complaint adequately stated a prima facie claim of copyright infringement based on ChatGPT’s outputs.11 He therefore denied OpenAI’s motion to dismiss. Judge Stein did caution, however, that nothing in the judgment should be taken to suggest a view on whether the allegedly infringing outputs are protected as fair uses of the original works.
Unlike the judgment that I have just discussed, the second and third judgments both relate to inputs into Large Language Models. The two judgments are:
- Bartz v Anthropic PBC12 (Bartz or the Bartz judgment), decided by Judge William H Alsup in the US District Court for the Northern District of California on 23 June 2025; and
- Kadrey v Meta Platforms, Inc13 (Kadrey or the Kadrey judgment) decided by Judge Vince Chhabria in the US District Court for the Northern District of California. Interestingly, this judgment was handed down two days after the Bartz judgment, and makes a passing (and perhaps critical) reference to the Bartz
In each of these cases, the defendant filed a motion seeking summary judgment on the basis that their use of the plaintiffs’ works constituted a “fair use” within the meaning of the applicable US provision, which is s 107 of the US Copyright Act.
Given the similarity between the issues in the two judgments, I am going to focus on the Bartz judgment and then deal briefly with Kadrey.
The Bartz judgment provides the following description of the defendant and its AI offering:14
Defendant Anthropic PBC is an AI software firm founded by former OpenAI employees in January 2021. Its core offering is an AI software service called Claude. When a user prompts Claude with text, Claude quickly responds with text — mimicking human reading and writing. … Users may ask Claude some questions for free. Demanding users and corporate clients pay to use Claude, generating over one billion dollars in annual revenue.
The plaintiffs – Andrea Bartz, Charles Graeber and Kirk Wallace Johnson – are authors of books that Anthropic copied from pirated and purchased sources. (The distinction between pirated and purchased sources becomes important.) The case was brought as a class action, but as at the time of the judgment class certification had not yet occurred.
The factual background as summarised in the judgment includes:15
… in January or February 2021, [an] Anthropic cofounder … downloaded Books3, an online library of 196,640 books that he knew had been assembled from unauthorized copies of copyrighted books — that is, pirated. …
In June 2021, [the co-founder] downloaded … at least five million copies of books from Library Genesis, or LibGen, which he knew had been pirated.
…
Anthropic … pirated over seven million copies of books, including copies of at least two works at issue for each Author.
…
As Anthropic trained successive LLMs, it became convinced that using books was the most cost-effective means to achieve a world-class LLM. During this time, however, Anthropic became ‘‘not so gung ho about’’ training on pirated books ‘‘for legal reasons’’.
…
To find a new way to get books, in February 2024, Anthropic hired the former head of partnerships for Google’s bookscanning project …
[That person] and his team emailed major book distributors and retailers about bulk-purchasing their print copies for [Anthropic’s] ‘‘research library’’. Anthropic spent many millions of dollars to purchase millions of print books, often in used condition. Then, its service providers stripped the books from their bindings, cut their pages to size, and scanned the books into digital form — discarding the paper originals. Each print book resulted in a PDF copy containing images of the scanned pages with machine-readable text … It acquired copies of millions of books, including of all works at issue for all Authors.
The judgment describes the value of books in Anthropic’s training process:16
Over time, Anthropic came to value most highly for its data mixes books like the ones Authors had written, and it valued them because of the creative expressions they contained. Claude’s customers wanted Claude to write as accurately and as compellingly as Authors. So, it was best to train the LLMs underlying Claude on works just like the ones Authors had written, with well-curated facts, well-organized analyses, and captivating fictional narratives — above all with ‘‘good writing’’ of the kind ‘‘an editor would approve of’’.
In this context, Anthropic moved for summary judgment on the basis of its “fair use” defence. Judge Alsup explained that to prevail on summary judgment, Anthropic needed to rely on undisputed facts and/or factual inferences favouring the opposing side.17
The Judge outlined s 107 of the Copyright Act,18 which identifies four factors that are used in determining whether a given use of a copyrighted work is a fair use. These are:
(1) the purpose and character of the use, including whether such use is of a commercial nature or is for non-profit educational purposes;
(2) the nature of the copyrighted work;
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work.
The Judge then addressed each of the four factors.
In evaluating each of the factors, the Judge dealt separately with:
- the use of the copyrighted works to train the Large Language Models; and
- the use of the works to build a central library.
For present purposes, I will focus on the way the Judge analysed the use of the copyrighted works to train the LLMs. I will then refer briefly to the way he analysed the creation of the central library.
The first factor is the purpose and character of the use. The Judge said that the purpose and character of using works to train LLMs was “transformative”, indeed he said it was “spectacularly so”.19
The word “transformative” appears several times in Judge Stein’s judgment and there is a body of US case law about this concept.20 To an Australian reader it may not be immediately clear whether this is a reference to (a) the copyrighted works having been transformed into new and different works, or (b) the character of the technology. It appears to be at least the former, but it may also be the latter. On this factor, Judge Stein said:21
In short, the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative. Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different. If this training process reasonably required making copies within the LLM or otherwise, those copies were engaged in a transformative use.
On this basis, the Judge concluded that the first factor favoured “fair use”.
The second factor is the nature of the copyrighted work. Here, all of the Authors’ works contained expressive elements. Indeed, they were chosen for their expressive qualities. This factor therefore pointed against “fair use”.22
The third factor is the amount and substantiality of the portion used. The Judge explained that the crux of this factor is whether the amount is “reasonable in relation to the purpose of the copying”.23 Even though the entire works were used to train the LLM, this was considered reasonable given the purpose of the use. Accordingly, the third factor favoured “fair use”.24
The fourth factor is the effect of the use on the market value for, or value of, the copyrighted works. The Judge noted the Authors’ argument that training the LLMs will result in an “explosion” of works competing with their works – such as by creating alternative summaries of factual events, alternative examples of compelling writing about fictional events, and so on.25 Judge Alsup rejected that contention. He reasoned:26
… Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works. This is not the kind of competitive or creative displacement that concerns the Copyright Act. The Act seeks to advance original works of authorship, not to protect authors against competition.
He therefore concluded that the fourth factor favoured “fair use”.
Weighing these factors together, Judge Alsup concluded that the copies used to train the LLMs were justified as a “fair use”: every factor except the nature of the copyrighted work favoured this result.27
As I mentioned, Judge Alsup conducted a separate analysis of the four factors in connection with the building of a central library (which essentially involved a print to digital format change). I won’t go into this aspect in detail. In summary, the Judge concluded that the use of purchased copies to create the library did constitute a “fair use”, but the use of pirated copies did not.28
The outcome was that the Judge granted summary judgment for Anthropic in relation to the training part of the case. He also granted summary judgment for Anthropic in relation to the use of purchased copies to create the central library. But he refused summary judgment in relation to the pirated copies used to create the library, indicating that there would be a trial on that aspect of the claim.
This case subsequently settled (in August 2025) for $1.5 billion.29 This was described in one report as the largest copyright settlement in US history.
I won’t discuss the Kadrey judgment in detail as it too concerned the issue of “fair use”. I will just make one or two observations about it.
The Kadrey case was brought by 13 authors against Meta Platforms, Inc. The plaintiffs were published authors who had written, and who held copyright in, various works. Those works were mostly novels, but also included plays, short stories, memoirs, essays, and nonfiction books.
As with the Bartz case, this case was concerned with the use of the authors’ works to train LLMs. Meta applied for summary judgment on the basis that its use was a “fair use”. Judge Chhabria concluded that the two arguments relied on by the authors were weak and he therefore granted summary judgment to Meta.
However, Judge Chhabria indicated that he thought there was a much better argument available to plaintiff authors in other cases, which had not been sufficiently raised in the evidence or submissions of the plaintiffs.
The alternative argument concerned the fourth factor in the “fair use” analysis. It will be recalled that this factor is the effect of the use upon the market for, or the value of, the copyrighted works. Judge Chhabria said that this was undoubtedly the most important of the four factors.30
Although not established on the evidence in the present case, Judge Chhabria considered that plaintiffs in other cases could seek to establish that generative AI would displace or diminish demand for their works and that this could be determinative in the “fair use” analysis.
Judge Chhabria took issue with Judge Alsup’s schoolchildren analogy in the following passage:31
… [I]n a recent ruling on this topic, Judge Alsup focused heavily on the transformative nature of generative AI while brushing aside concerns about the harm it can inflict on the market for the works it gets trained on. Such harm would be no different, he reasoned, than the harm caused by using the works for ‘‘training schoolchildren to write well,’’ which could ‘‘result in an explosion of competing works.’’ … According to Judge Alsup, this ‘‘is not the kind of competitive or creative displacement that concerns the Copyright Act.’’ … But when it comes to market effects, using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take. This inapt analogy is not a basis for blowing off the most important factor in the fair use analysis.
In the conclusion to his judgment, Judge Chhabria said:32
Fair use is a fact-specific doctrine that requires case-by-case analysis that is sensitive to new technologies and their potential consequences. No previous case has involved a use that is both as transformative and as capable of diluting the market for the original works as LLM training is.
…
In cases involving uses like Meta’s, it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects of the defendant’s use. No matter how transformative LLM training may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books.
However, because the 13 plaintiffs in the Kadrey case had not adduced evidence that a jury could use to find in their favour on the issue of market effects, Judge Chhabria granted summary judgment to Meta.
That concludes my survey of recent US cases in this area. Obviously, there are significant differences between the US principles relating to “fair use” and Australian principles relating to “fair dealing”. Nevertheless, given the topicality of generative AI, I hope that you found this survey interesting.
[1] Judge, Federal Court of Australia. I would like to thank my Associate, Deylan Kilic-Aidani, for his research assistance in the preparation of this speech.
[2] Re OpenAI Inc Copyright Infringement Litigation (unreported, 27 October 2025) (OpenAI).
[3] OpenAI at 3.
[4] OpenAI at 3-4.
[5] OpenAI at 5.
[6] OpenAI at 3.
[7] OpenAI at 11.
[8] OpenAI at 14.
[9] OpenAI at 14.
[10]OpenAI at 17.
[11]OpenAI at 17.
[12] Bartz v Anthropic PBC, 787 F.Supp.3d 1007 (N.D.Cal. 2025).
[13] Kadrey v Meta Platforms, Inc, 788 F.Supp.3d 1026 (N.D.Cal. 2025).
[14] Bartz at 1014.
[15] Bartz at 1015-1016.
[16] Bartz at 1017.
[17] Bartz at 1018.
[18] Bartz at 1019-1020.
[19] Bartz at 1021.
[20] See also Kadrey at 1044.
[21] Bartz at 1022.
[22] Bartz at 1029.
[23] Bartz at 1029.
[24] Bartz at 1031.
[25] Bartz at 1031-1032.
[26] Bartz at 1032.
[27] Bartz at 1033.
[28] Bartz at 1033.
[29] See, eg, Kluwer Copyright Blog, “The Bartz v. Anthropic Settlement: Understanding America's Largest Copyright Settlement” (10 November 2025) <https://legalblogs.wolterskluwer.com/copyright-blog/the-bartz-v-anthropic-settlement-understanding-americas-largest-copyright-settlement/>.
[30] Kadrey at 1050.
[31] Kadrey at 1035-1036.
[32] Kadrey at 1059.
