The legality of generative AI, from an intellectual property perspective, has been a hotly contested issue ever since the release of ChatGPT in November 2022. 

In this article, we take a look at the current state of play, and explore the 'text and data mining' (or TDM) exception to copyright infringement, which is receiving increased attention as governments consider how to support and grow their AI industries. The merits of a TDM exception are also being debated close to home, as Australia’s Productivity Commission has recently sought feedback on whether a TDM exception would facilitate technological innovation and productivity growth

Generative AI’s never-ending day in court

With great power comes great… liability? ChatGPT’s sudden meteoric rise has ushered in a new wave of generative AI applications, and AI investment and development at a pace the world has not seen before. However, as the effects of generative AI have been felt across different industries, especially the creative industries, AI developers have found themselves the target of numerous court cases alleging copyright infringement in the development and use of AI applications.

There are two primary types of claim a copyright owner might have in respect of a generative AI model:

1. That infringement has occurred in the process of training an AI model. Generative AI is typically trained on large datasets which can be comprised of existing copyright works, or data generated from existing works. The creation of these datasets will naturally involve some degree of interaction with copyright works, which might amount to infringing copying of those works.
2. That infringement has occurred in the creation and/or dissemination of AI-generated outputs which are themselves copies of copyright works.

Many developers have designed or trained their generative AI models to ‘learn’ from existing copyright works, but to not produce outputs that are an obvious copy of those works. As a result, the focus of recent court action has been on the use of copyright works during the AI model training process.

In recent months, a couple of US District Court rulings have indicated that a ‘fair use’ defence may be available to developers faced with copyright lawsuits relating to AI model training. This includes Bartz v Anthropic, in which a number of authors brought a class action alleging copyright infringement in the training of AI models on copyrighted books. A US federal judge found that the use of lawfully obtained copies of the plaintiffs’ works for the purposes of training AI models was a ‘fair use’, but that the case was to proceed in respect of any pirated copies included in the relevant datasets, to which a fair use defence could not be justified. A US$1.5 billion settlement of the remainder of the case has recently received preliminary court approval. 

High profile cases which are still before the courts include Getty v Stability AI and New York Times v OpenAI, both of which involve allegations around the use of copyright works (images and news articles) to train generative AI models. There are also a number of other pending lawsuits involving generative AI which have not proceeded to judgment, so we do not yet have a definitive answer to the question of whether the development of this technology actually infringes copyright, although it is possible the Anthropic approach will be followed in other jurisdictions where a ‘fair use’ defence is available. It is also highly likely that any judgments that do come will be appealed, given the high stakes, so it may be some time before we have a clearly established precedent on this issue. Further, any precedents relating to 'fair use' will not necessarily have a bearing on jurisdictions that do not recognise this doctrine.

This article focuses on the issue of exceptions and defences to copyright infringement, which assumes an initial 'copying' or other infringing act has occurred at the outset in respect of a copyright work. However, there is still a live issue in each case as to whether an existing work has actually been interacted with in a manner that constitutes 'copying' or another infringing act. This issue may not be a straightforward one, given the manner in which some AI models interact with their training data, and for the purposes of this article we have not considered it in detail.

Overcoming the infringement risk

It is a safe assumption that court actions are proving challenging for copyright owners and AI developers alike. For copyright owners, litigation is expensive, there may be difficulties in proving some of the factual and legal elements of a claim, and bringing a claim does little to solve the immediate problem of owners not being remunerated for the use of their creative works. For developers, litigation is also expensive, and the threat of successful litigation may disincentivise external investment and the further development of AI technologies.

While it is possible that none of the infringement cases before the courts will succeed, in the absence of a well-established legal precedent, a theoretical risk remains for developers and customers of AI tools. This risk may also pose a challenge for governments wishing to take advantage of the 'AI boom'. The UK government has raised concerns that uncertainty over the application of copyright law to AI models may hinder innovation, and prevent countries with less permissive copyright laws from being competitive in the AI space. At the same time, it has acknowledged that this uncertainty may also be making it difficult for copyright owners to assert their rights and seek remuneration – so the creative sectors may equally be suffering under the status quo.

In light of this uncertainty, some governments have been considering how to provide a clear pathway for future AI development. One option that has been mooted is a 'text and data mining' exception to copyright infringement. Text and data mining is in the spotlight at the moment, as both the Australian and UK governments have indicated that a new or more expansive TDM exception is under consideration. Other jurisdictions also appear to be considering whether clarity is required regarding the application of their existing TDM exceptions to AI training.

Another option that has been considered is to encourage, or mandate, the use of collective licensing arrangements to enable developers to obtain (and pay for) licences from copyright owners on a collective basis. Collective licensing could be equally attractive to copyright owners and developers, given its relative simplicity and the certainty it can offer to both sides as to their legal position. Collective licences have already been established in some jurisdictions, and a number of large media companies have successfully entered into licensing agreements which permit OpenAI to use their content for training purposes, which suggests there might be some merit to the licensing option.

What is the TDM exception?

Text and data mining is “the process of deriving information from machine-read material” which involves “copying large quantities of material, extracting the data, and recombining it to identify patterns” (reference). TDM is typically carried out through software which deploys automated analytical techniques to ‘mine’ information sources for patterns, trends and high level data – essentially, the process of training most generative AI models.

TDM exceptions under copyright legislation provide that the use of copyright works for TDM purposes will not amount to infringement, even though there may be a technical copying of those works. In 2013, the Australian Law Reform Commission commented that there was “growing recognition that [TDM] should not be infringement because it is a ‘non-expressive’ use” and “copyright law protects the expression of ideas and information and not the information or data itself”.

Australia does not have a TDM exception currently, but an August 2025 report by the Productivity Commission recommends this is considered as a means of boosting Australia’s AI industry. The Australian government has neither explicitly confirmed nor denied that this option is on the table, although the Arts Minister has stated there is no intention to 'weaken' Australia’s copyright laws.

While Australia is still considering its options, a TDM exception does currently exist (in some form) in a number of jurisdictions:

• In the EU, text and data mining of copyright works is permitted pursuant to the 2019 Directive on Copyright in the Digital Single Market. Article 4 of the Directive requires member states to implement an exception to infringement for “reproductions and extractions of lawfully accessible works and other subject matter for the purposes of text and data mining”. Unlike some other jurisdictions, this exception is not limited to research or non-commercial applications, so on its face it is very permissive. However, this exception applies to a copyright work only if the relevant rights holder has not expressly opted out of it “in an appropriate manner, such as machine-readable means”. It appears there is little guidance on what amounts to an effective ‘opt-out’, nor how companies engaging in TDM (including AI developers) can effectively identify and track any ‘opt-outs’ in the (often vast) databases they are working with. Accordingly, there is dissatisfaction with this model on both sides of the fence.
• In the UK, the Copyright, Designs and Patents Act 1988 provides that “a person who has lawful access to the work may carry out a computational analysis of anything recorded in the work for the sole purpose of research for a non-commercial purpose”. While there is some ambiguity around what constitutes “research for a non-commercial purpose”, there seems to be a general consensus that using TDM to train AI models, where those models will be distributed as commercial products, will not fall within this exception.
• Singapore and Japan have fairly broad TDM exceptions which are not limited to non-commercial applications and do not provide for an opt-out regime. However, Japan’s Copyright Act 1970 requires that the person undertaking TDM must only be doing so for 'non-enjoyment' purposes – that is, they must not personally 'enjoy' or cause another person to enjoy the relevant work and it must solely be used for data extraction purposes. Further, the use of the work must not unreasonably prejudice the interests of the copyright owner, given the nature or purpose of the work or the circumstances of its exploitation.
• There is no specific TDM defence in the US, but our understanding is that TDM has been viewed in some cases as falling within the scope of the broad 'fair use' defence to copyright infringement. (Note that 'fair use' is different to the concept of ‘fair dealing’ in New Zealand, which we discuss below.)

Beyond AI model training, TDM is also relevant to the wider technology sector, and the existence (or not) of a TDM exception under copyright law will no doubt be of interest to members of that sector. TDM can be used for the purposes of information retrieval and extraction, market research and social media analysis, customer service optimisation, and any other task requiring the analysis of unstructured and/or high volume data.

TDM in NZ: Where do we sit?

There is no specific TDM exception or defence to copyright infringement in New Zealand. Nor does New Zealand have a general 'fair use' defence which could be invoked in a TDM context.

New Zealand’s Copyright Act 1994 contains a series of specific 'fair dealing' defences, as well as other specific exceptions to infringement. Of these, there are two which could potentially provide an argument in defence of text and data mining of copyright works, although there are challenges in both cases:

• Section 43 permits fair dealing with a work for the purposes of research or private study. It is debatable whether mining a work for data amounts to 'research'. Factors relevant to whether any dealing has in fact been 'fair' include whether the work could have been obtained within a reasonable time at an ordinary commercial price, and the effect of the copying on the potential market for, or value of, the work. Again, it is debatable whether this defence could be utilised in connection with the development of commercial AI models which may disrupt the market for the mined copyright works.
• Section 43A provides that a 'transient or incidental' reproduction of a work does not infringe copyright, provided the reproduction is a necessary part of a technological process for making or receiving a non-infringing communication or enabling the lawful use of the work, and it has no independent economic significance. Whether this exception might technically apply in instances of TDM is, in part, a technical question – although we understand similar exceptions have been discussed in connection with TDM by commentators in other jurisdictions. Further, as discussed above, it is debatable whether reproductions for the purpose of developing commercial AI models have 'no independent economic significance'.

There is no indication that a TDM exception to copyright infringement is on the New Zealand government’s radar at all, let alone a priority policy area. The AI Strategy released in July 2025 made no mention of including TDM in the Copyright Act review that has stagnated since 2018, and the government has consistently maintained the position that existing New Zealand laws are fit for purpose from an AI perspective. The AI Strategy also “deliberately emphasises AI adoption and application rather than foundational AI development, reflecting both economic reality and strategic opportunity for New Zealand’s context”, which suggests that (at least at this stage) there is no strong desire to grow a competitive AI development sector and the focus remains on AI use. As long as this remains the case, we would be surprised to see a TDM exception being floated in New Zealand.

However, we will be watching our cousins across the ditch and further afield with interest, to see how their laws evolve, are amended, and are enforced. Things are changing at an incredibly rapid pace, and New Zealand will not be immune from the effects of global economic and legal trends.

Talk to the experts

Get in touch with our technology experts Campbell Featherstone, Hayley Miller, Ashleigh Ooi or Güneş Haksever for advice on the procurement, development, commercialisation and use of AI and other technologies. Their team also regularly advise on the application of New Zealand laws, including privacy laws, to new and emerging technologies. For advice on New Zealand’s intellectual property laws, or support with IP filings or disputes in New Zealand, reach out to Jenni Rutter.