Australia's next AI bottleneck: Why copyright matters

If a major AI company was preparing a checklist to decide where to build its next AI training facility, you would expect to see access to land, power and water at the top of the list. It may be surprising to also see copyright featured prominently. This sleeper issue presents a major constraint on AI companies investing in Australia. It may even be a decisive factor.

Much of the material used to train AI – including books, news, images, social media posts, music, and software code – is protected by copyright. Use of that material is generally subject to the permission of the copyright owner. Training AI models in Australia without permission would expose an AI operator to a claim for compensation by the copyright owner and potentially an injunction to prevent use of AI models trained on infringing material. For AI training conducted at scale, there is a risk of litigation, potentially including class action lawsuits like those in the US (more on that later) and additional damages if the copyright infringement is considered flagrant.

On its face, that seems fair enough: creators should be, and generally are, entitled to control how their creation is used. If any AI operator wants to use someone else’s creative work, then they should just ask for permission and pay a licence fee. Right?

The theory may be simple; the reality is not. In practice, it is not feasible for AI operators to negotiate voluntary licences on a bilateral basis with millions of individual copyright owners. Even if AI operators concentrate efforts on a few owners of large bodies of copyright works – say, major media companies – that still leaves a large amount of material on the table, and an AI model trained on incomplete material is an inferior AI model. Technological limitations on how material is identified and used for training also limit the workability of voluntary or individual licensing models.

Of course, copyright has never been an absolute right. Australian law has always sought to strike an appropriate balance between the interests of copyright owners and broader society. A variety of existing exceptions permit use of copyright material without an express or voluntary licence from the owner.

Finding the right balance is a challenging task. Over time, changes have been made to accommodate technological developments. Australia’s current Copyright Act was made in 1968. As then-Attorney General Nigel Bowen put it on reading the 1968 bill for the second time:

[T]here are changes in the use of copyright material which have been brought about by changes in technology and the Government has been concerned to see that authors receive due payment for the use of their material. At the same time the Government recognises that existing practices and existing relationships in industries which depend upon copyright material cannot be ignored.

The 1968 Act replaced an earlier iteration introduced in 1912. It had 56 years of technological change to catch up on. Australians now had radio, cinema, TV, new printing methods, and more. It was obviously imperative for the law to respond, and it did. Since 1968, there has been further tinkering with copyright legislation to address technological developments. From the introduction of photocopiers in libraries, to the wide availability of blank VHS and audio cassette tapes enabling home TV and radio recording, through to the advent of internet piracy, it has been a constant battle for copyright law to respond to what technology has made possible. AI represents the next major development in this evolutionary chain.

There are many possible models which might provide AI operators with certainty over their ability to train AI models in Australia on copyright material without needing to strike millions of individual licensing deals. There is the ‘fair use’ doctrine in the US - a more open and flexible version of the fair dealing exemptions already recognised in Australian law. Uncertainty remains, however, over the scope of what constitutes a fair use in the context of AI training, as reflected in the current swathe of active AI-related copyright lawsuits in the US. A further alternative would be to introduce a dedicated ‘text and data mining’ exception to allow copying and analysis of large digital data sets. Such exceptions are recognised in some jurisdictions, such as the EU, subject to opt-out rights for copyright owners. However, the Albanese Government has ruled out introducing such an exception in Australia.

A brief look at other jurisdictions shows that there are many other solutions to explore. For example, Japan and Singapore - Australia’s direct competitors for AI workloads in the region - have introduced broad statutory exceptions to support data analysis and machine-learning under which no remuneration is payable to copyright owners, though each within limits (Japan does not allow unreasonable prejudice to owners, while Singapore requires that material is accessed by lawful means). Other novel approaches include an output-side tax and collective-management model (along with authorisation and labelling requirements) that has been proposed in France and a negotiated collective licensing framework that has been enacted in Denmark. In Brazil, a statutory remuneration framework has been proposed with AI operators needing to disclose protected material used in training and rightsholders being able to negotiate licence fees collectively or directly by reference to criteria including company size, frequency of use and competitive effects.