I'm an Australian novelist and software engineer. My brother Matt is a law professor at Emory University in Atlanta, USA. I write science fiction. He's an expert in copyright and AI. He and I have obviously had many chats about creativity, intellectual property, and the increasingly fraught relationship between them. To quote a character in my own book, I'm no lawyer, but I know there are laws. I've read many of Matt's papers over the years. I was excited to see that he's recently assembled a lifetime of thinking about this stuff into his new book, Copyright Law in the Age of AI. It synthesises his thinking, and research, on the topic. He's very generously made it available to read online for free. Go Matt! It's very readable even for a not-American, not-lawyer like me.
That said it's a huge piece of work: 28 chapters examining US copyright law, its history, its underlying principles, and the challenges presented by artificial intelligence. It explores what copyright protects, who qualifies as an author, how ownership works, what constitutes infringement, and when using someone else's copyrighted material might be permissible. According to US law anyway, which, it turns out is philosophically different from other nations' copyright laws.
Matt goes deep about whether AI-generated works even can attract copyright protection, and he investigates the legality of training of AIs on copyrighted material. Matt talks a lot about nonexpressive use: using copyrighted material to extract information, discover patterns, or to develop knowledge without reproducing the original creative expression for an audience.
Ref: aicopyright.ai/chapters/15
It's an idea he's been developing for years, well before generative AI became a thing. Well before I started writing Daedalus. Okay there was always a lot of talk about AI in our household, even as kids. Mum was, and still is, albeit less often now, a science fiction reader. She thought Daedalus was "modern," and a little confusing.
All of this got me thinking: What is copyright actually for, and what is it actually there to protect?
I'm not a lawyer, and this isn't intended as a legal analysis, just so we are clear. I'm approaching the subject as someone who creates copyrighted works. I'm also someone who designs and writes software, both for fun and commercially, alone and in teams, with and without AI. I've read a thousand books. I've seen a lot of films and TV, and studied their composition and grammar. I have spent all of my life learning from others, often without their consent. Often even after they are dead!
Enjoying this essay? Consider subscribing for more.
What's copyright for?
Copyright is a bargain between writers, artists, filmmakers etc, and society as a whole. Society grants creators a limited monopoly over certain uses of their work. This monopoly lets them benefit economically from their creativity. In exchange, society benefits from the creation and eventual availability of literature, music, art, software and other cultural works. It's important to note here that copyright does not grant creators the right to benefit economically from their work, it only enables the ability for them to do that. Copyright law doesn’t guarantee an author an income or insulate them from ordinary competition.
Authorship implies moral rights, including recognising authorship and protecting the integrity of creative works. The fundamental economic arrangement is that a creator receives certain exclusive rights, for a limited period, in the expectation that doing so encourages creativity. It's also important to note that copyright doesn't protect ideas themselves. It protects human-made, original, expressions of ideas.
I can write [a novel about an artificial intelligence that becomes conscious without infringing the copyright of every author who has previously explored machine consciousness. "I've kenned Banks, Stross, and McLeod," to quote my own book again. I can learn from their stories, absorb their techniques, borrow broad concepts, and develop my own style. It's honestly hard to imagine how art or science would progress at all if we couldn't learn from others. Every creative work exists within cultural contexts built by an unknowable network of prior-creators. That's not to say there is nothing new, but there really are only 21 stories in the world.
So where do we draw the line between protecting creative expression on one hand, and controlling what can be learned from it on the other?
Why does the method of learning matter?
Consider these three scenarios.
In the first, I read a thousand science-fiction novels. I pay attention to how their authors construct plots, develop characters, describe imaginary technologies, and build convincing fictional worlds. In the fullness of time I write my own novel, influenced by everything I've ever read. This is entirely uncontroversial.
In scenario the second, I write some software that scans those same thousand novels. It identifies recurring themes, measures sentence lengths, examines vocabulary, and discovers statistical relationships between different narrative structures. I then use what the software has distilled to inform my own writing.
In the third scenario, I train a massive artificial neural network on those thousand novels. It develops a mathematical model of the relationships within the training material. I subsequently use outputs from that model to help produce new writing.
These activities are obviously different, both technically and in their potential consequences. They have something fundamental in common though. In each case, information is extracted from existing creative works and used, by a human, to create something new.
The legal treatment of these activities can differ substantially. Reading a book and learning from it doesn't infringe copyright. Computational analysis, however, may require making copies of entire works. Training neural networks generally involves similar acts of reproduction. The intermediate copies can be ensnared by copyright law even when the resulting analysis or model doesn't reproduce any recognisable portion of the original works.
As a software engineer, I find this distinction fascinating. A process that is unremarkable when performed by a human brain can become legally problematic when performed by a computer. The distinction you could argue is that computers can do things that humans can't. They outperform us in terms of scale and speed. Their activity has enormous commercial and societal consequences.
But, and it's a big but, ought the mechanism by which learning occurs determine whether the learning itself is permissible? Or should we be more concerned with what happens to the knowledge afterwards?
Meanwhile, Down-Under…
Alas Matt clearly hasn't had time to write a book about Australian copyright law. But I can tell you, Australian and US copyright law have some significant differences, and again IANAL, but I can read. The USA has a legal doctrine called fair use. Fair use allows courts to consider whether particular uses of copyrighted material are permissible, even without the copyright owner's permission. Courts consider several factors, including the purpose of the use, the nature of the original work, how much was copied and the effect on the market for the original. It's a flexible approach, capable of accommodating circumstances that legislators might never have anticipated.
Australia OTOH takes a different path. Our Copyright Act 1968 contains specific exceptions, including fair dealing for research or study, criticism or review, parody or satire, and news reporting. These exceptions are great, but they are very narrowly defined. There is no general Australian equivalent of US fair use. Consequently, a novel technological use of copyrighted material might be considered fair under US law while falling outside the exceptions available under Australian law.
Note: That doesn't mean every AI training exercise is lawful in the USA, or that every computational analysis is unlawful in Australia. The circumstances matter enormously. See: legislation.gov.au/C1968A00063
But the underlying approaches do differ. US law provides a fairly open-ended mechanism for asking whether a particular use is fair. Australian law requires us to establish that the use falls within an existing statutory exception. Personally I find myself wondering whether these differences reveal something deeper about how our two countries understand the purpose of copyright.
The US approach gives courts a lot of scope to balance competing interests as technologies evolve. Ours places more emphasis on boundaries established through legislation. There are swings and roundabouts. Both approaches have advantages, and both create uncertainties of their own.
Australia considered changing its rules
Australia did in fact examine this question. In 2013, the Australian Law Reform Commission completed a substantial inquiry into copyright and the digital economy. Its report recommended introducing a general fair-use exception into Australian copyright law. The Commission specifically considered text and data mining, including the kind of nonexpressive use that Matt discusses in his work.
The ALRC recognised that computational analysis frequently requires copying entire works, even when the purpose is simply to extract information or discover patterns. The Commission concluded that fair use would provide a flexible way to assess such activities, taking account of their purpose, their commercial implications, and their effects on copyright owners. Alas the recommendation was not implemented. Then, a year ago, in October 2025, the Australian government explicitly ruled out introducing a text and data mining exception for AI training. Their stated concern was protecting Australian creators' control over their work and ensuring they can receive appropriate compensation when that work is used to develop commercial AI systems.
They reaffirmed this position in July 2026. Ref: ministers.ag.gov.au/media-centre/transcripts/radio-interview-abc-radio-national-breakfast-sally-sara
And that's fair enough, surely. Authors, musicians, illustrators, and other creative professionals have legitimate interests in how their work is commercially exploited. And yes the emergence of generative AI has made those interests considerably more, um, well it's complicated. A company might use millions of copyrighted works to develop a product worth billions of dollars. It's hardly a shock that the creators of those works would want recognition and compensation. The AI industry haven't exactly behaved nicely here either. Anthropic buying up rare books and shredding them feels like a slap in the face to people who love books.
On the other hand, licensing requirements can create barriers for researchers, independent developers, and hobbyists and fans. They favour big companies with the financial clout to negotiate access to massive collections of copyrighted material. Like industrial fishing. These competing interests are quite hard to reconcile. The difficulty arises, I think, in part because we're trying to regulate a way of learning using legal concepts developed to regulate copying and distribution.
Then we taught sand how to learn
There's an important complication in my three examples above. The person reading a thousand novels might eventually write one, or a few novels of their own. A generative AI system trained on a million novels might help produce a hundred new novels per day. Even if none of those new books reproduces any protected expression from the training material, the economic consequences could be substantial. This is where the analogy between human learning and machine learning becomes a bit squirmy. Human creativity has always involved learning from others. But the scale of industrialised learning introduces possibilities that didn't previously exist. A machine can absorb extraordinary quantities of information, operate continuously, and generate outputs at negligible marginal cost. Those outputs compete directly with the works from which the system learned.
But should that competition even matter to copyright law? Remember, copyright law is not here to protect a creator's income, but their right to own a piece of property that they may or may not be able to use to generate an income. No one expects copyright law to protect an author against competition from another author who has learned from their work. I don't owe Charles Stross or Michael Moorcock a cent for shamelessly, um, honouring them in Daedalus.
A successful science-fiction novel might inspire hundreds of other writers to explore similar themes. That's just how culture works. But what if the competing works are generated by industrial systems trained on millions of existing books? Does the difference in scale justify treating the activity differently? If so, where do we draw the line? Would the same logic apply to software that analyses scientific papers, systems that identify patterns in medical research, or astronomy data, or tools that discover relationships across vast collections of historical documents?
There's a distinction between the act of learning and the subsequent commercial exploitation of what has been learned. I think those activities need different treatment. Maybe the economic consequences of industrial learning require entirely new mechanisms for compensating creators, or perhaps existing copyright principles can accommodate these developments, provided we remain clear-eyed about what copyright is actually meant to protect.
Unfortunately I don't have a settled answer, this is just stuff I've been thinking about since reading my brother's book.
Who owns the right to learn?
As a novelist, I want my creative work protected, and to be able to write new books, inspired by things I've learned. As a software engineer, I want to be free to analyse information, develop new technologies, and build upon existing knowledge. As a reader, I want authors to continue writing books worth reading, and not be inundated with AI slop. And I want my AIs to know and understand all the world's books, films, tv shows, plays, and other media. I want it to be able to help me understand the symbolism used in the song Mr Heartbreak, or settle an argument about the lyrics of Love Is a Stranger, or give me a short synopsis of the book White Noise, which I've heard is great but I just couldn't get into.
Copyright has always involved a deal between the rights of creators and the interests of the wider public. AI has pushed that deal into unfamiliar territory, on players with different agendas. The differences between Australian and US law offer two ways of approaching the problem, each reflecting choices about where to place the boundaries of copyright protection. Matt's distinction between using a work for its creative expression and using it to discover information, I find especially interesting. For me though, the most interesting question remains philosophical rather than legal. We have very good reasons why we protect creative expression. We also have good reasons to encourage learning, experimentation, and the development of new, high quality knowledge and cultural material. As machines become increasingly able to learn from the wider creative output of humanity, I'm most interested in knowing, Who owns the right to learn from what someone else has created?. Or should no one own that right?
If you liked this essay, consider subscribing for more.