Since 2023 in particular, United States courts have entertained a series of copyright class actions against AI developers. The most prominent, Bartz v Anthropic, produced a settlement of US$1.5b that received final approval in July 2026 and stands as the largest copyright class action settlement in history. No equivalent proceeding has (ever) been commenced in Australia. Why?
Copyright is the cause (of action)
Four features of Australian copyright law make it hard to consolidate claims into a representative proceeding.
First, ownership. Copyright is unregistered in Australia and arises automatically for eligible authors. Unlike trade marks and patents, no database exists to search, and no certificate exists to produce. The lack of a central ownership register means that any person who wishes to sue must individually establish ownership, tracing title through assignments, employment agreements, or publishing deals with the author. Each ‘work’ or ‘subject-matter’ in issue demands its own analysis.
In copyright law, the ‘author’ is the natural person who creates the work. That means the human whose skill and labour produces the original expression. The author is not necessarily the owner. Copyright in a work created by an employee in the course of employment vests in the employer, not the employee. And authors routinely assign their rights to publishers, record labels, production companies, or other parties. The result is that the ‘author’ and the ‘owner’ of a work are often different people, and the current owner may be several steps removed from the original creator.
Second, infringement. To establish infringement by reproduction, a claimant must prove that the respondent copied their work, and that what was copied constitutes a ‘substantial part’ of it.
Both inquiries are claimant-specific:
|
Did they copy from you?
|
Did they copy enough?
|
|
Even if the respondent copied thousands of works through the same course of conduct, each claimant must prove copying. Did the respondent copy this claimant’s work? Copyright complainants must establish a ‘causal connection’ between their work and the defendant’s conduct, and the answer may differ for each group member. |
Copyright does not protect ideas, facts, or style, only original expression. Even if copying is proved, the Court must assess whether what was taken is a ‘substantial part’ of that claimant’s work. The inquiry is qualitative. A small but highly original passage may be substantial; a large and commonplace passage may not. |
Consider an AI claim as an example. Two authors whose novels appeared in the same training dataset may have quite different infringement claims. One author’s prose may have been reproduced verbatim in model outputs, while another’s may not have surfaced at all. One author’s work may be highly original literary fiction, while another’s may be formulaic genre writing with less protectable expression. The infringement analysis must be repeated work by work, claimant by claimant.
Copyright does not protect a book, an album, or a portfolio as a single unit. It protects each discrete ‘work’ or ‘other subject-matter’, meaning each novel, each song, each photograph, each article. The Copyright Act 1968 (Cth) distinguishes between ‘works’ (literary, dramatic, musical, and artistic works) and ‘subject-matter other than works’ (sound recordings, films, broadcasts, and published editions). Each attracts its own bundle of rights and its own authorship and ownership rules.
Third, damages. Even if a claimant establishes ownership and infringement, they must still prove loss. Unlike the United States, Australia does not permit copyright claimants to elect fixed statutory damages without proving actual loss. Under s115 of the Copyright Act 1968 (Cth), a successful claimant may recover compensatory damages or an account of profits, but must prove what those damages are. A bestselling novelist and an obscure blogger may both appear in the same AI training dataset, but the hypothetical licence fee for the novelist’s work may vastly exceed the blogger’s. The Court must assess, for each claimant, what a reasonable licensee would have paid for the right to use that work in that way.
Where a respondent has used a claimant’s property without permission but the claimant cannot prove actual financial loss, Courts may award ‘negotiating damages’, sometimes called the ‘user principle’. The measure is the fee that a willing licensor and willing licensee would hypothetically have agreed for the use in question. The concept originated in patent and property cases but applies equally to copyright. It asks: what would the parties have agreed, bargaining in the shadow of the claimant’s legal right to refuse?
Compare the position in the United States. Under 17 USC § 504(c), a copyright owner may elect statutory damages of between US$750 and US$30,000 per work infringed, or up to US$150,000 if the infringement is wilful. This obviates the need to prove actual loss and allows complainants to aggregate claims across thousands of works with ease. It is therefore no coincidence that the wave of AI-related copyright litigation has concentrated in US Courts, where statutory damages make class-wide recovery commercially viable.
Australian claimants have no equivalent mechanism, and no shortcut to an overall damages figure exists. Litigation funders typically demand a credible model of aggregate recovery before committing capital. Each claimant must prove damages separately, and the returns remain uncertain, which may make it difficult for litigation funders to justify their investment.
The same individualisation problems that complicate the merits may also invite a de-classing application under s33N of the Federal Court of Australia Act 1976 (Cth) (and cognate provisions in other jurisdictions), compounding the applicant’s and funder’s risk that the Court will unwind the proceeding before trial.
Fourth, jurisdiction. Australia has no text and data mining exception, and no equivalent to the US fair use doctrine, the EU’s DSM Directive, or even the UK’s narrow research exception that could provide defences. This might make Australia an attractive forum for copyright claims. But large-scale AI training occurs where infrastructure and legal conditions permit, and there is little evidence suggesting substantial training activity occurs onshore in Australia. In Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch), the UK High Court confronted this issue. Because training occurred on US servers, Getty abandoned its primary infringement claims, on the basis that there was no evidence to demonstrate that the relevant conduct occurred in the UK. The Court held that the model weights did not ‘store or reproduce’ the training images. Accordingly, those copyright claims failed.
Text and data mining (TDM) refers to the automated analysis of large volumes of text, data, or other content to identify patterns, trends, or relationships. In the AI context, it describes the process by which models train on datasets of text, images, or other works, extracting statistical patterns rather than ‘reading’ in the human sense. The Australian Law Reform Commission considered TDM in its 2013 report, Copyright and the Digital Economy, noting that ‘there is no exception in the Copyright Act that covers data and text mining’ and recommending a fair use exception that would accommodate non-expressive uses. The Government did not implement that recommendation. In October 2025, the Attorney-General confirmed that Australia would not introduce a TDM exception.
Case study: The Bartz v Anthropic Settlement
The Bartz v Anthropic settlement, approved by the US District Court for the Northern District of California in July 2026, provided a US$1.5b non-reversionary fund with an estimated per-work payment of approximately US$3,000, much higher than the US statutory minimum for wilful infringement (US$750) or for innocent infringement (US$200). The settlement obviated the need for individual proof of damages. Class members submitted claims through an administrative process, and funds were distributed pro rata based on the number of works claimed.
The settlement benefited claimants by reference to a ‘Works List’ of 482,460 books with ISBNs or ASINs that were registered with the US Copyright Office and appeared in pirated datasets (LibGen and PiLiMi) downloaded by Anthropic. The Works List served as ‘a defining record of which works are covered by the Settlement and which works are not’. Works not on the list were excluded from the release embodied in the settlement, preserving those claims for future litigation. The claims rate reached 91.3% of the Works List.
The settlement administrator sent notice to 506,194 potential class members identified through submissions from the Authors’ Guild, the Science Fiction Writers of America, the Authors’ Registry, more than 170 publishers, and third-party sources including Bowker ISBN Services and the US Copyright Office. Only 350 class members submitted valid opt outs, spanning 1,802 works.
While Anthropic agreed to destroy all original files of works from the pirated datasets, the settlement preserved claims about past outputs and all future conduct.
What to watch
On 15 July 2026, the Federal Labor Government announced that it would establish a new Office of AI within the Department of the Prime Minister and Cabinet, and confirmed that it would legislate to ensure Australian creators retain ownership and control of their work. The Prime Minister stated that ‘no company should use Australian books, music, art or news to build or train AI without the artist’s control. That includes the artist’s control of the price and value of their work.’ The detail remains unclear.
The Attorney-General has previously flagged 3 priority areas for reform:
- First, a new collective licensing framework, whether statutory or voluntary, to govern the use of Australian works in AI training.
- Second, legislative clarification of how copyright applies to material generated through the use of AI.
- Third, a small claims forum to provide a cheaper avenue for lower-value copyright disputes.
Without seeing the form of any proposed legislation, fundamental questions remain unanswered. Would a licensing scheme operate extraterritorially, and if so, how would it bind AI developers training models on foreign servers? Would it require developers to identify and compensate individual rightsholders, or would it function as a blanket levy distributed through collecting societies? And how would enforcement work when the allegedly infringing conduct occurs entirely offshore?
This jurisdictional paradox will not be solved by domestic licensing reform alone. Unless Australian law can reach conduct occurring on foreign servers, or unless AI developers voluntarily submit to Australian licensing requirements, the structural barriers to class-wide recovery will remain.
