Recent technological advances in generative AI have transformed the way in which we listen to and create music. Many artists and industry players have readily adopted generative AI, viewing it as a creative force to be harnessed, while others remain wary.
The use of increasingly sophisticated AI tools to generate musical works poses a number of challenges from a legal perspective, particularly in relation to how they may impact intellectual property rights.
In this blogpost we explore some of these legal considerations in further detail.
Input Considerations – Training AI Models
Generative or algorithmic art is not particularly new; the application of mathematical or rule-based systems and processes in the creation of art is well established. More recently, by utilising machine learning techniques, AI tools can craft impressive, creative works based on user inputs or prompts in real time.
Generative AI tools such as MuseNet, Jukedeck, AVIA and Udio have made it possible to generate entire musical compositions based on input parameters such as genre, tempo, mood and training data. In some cases, this may include catalogues of existing music being used by such models. However, large generative AI models require vast amounts of training data. This raises concerns for rightsholders whose works may be used in training these models. Using unauthorised or unlicensed works for training AI models carries significant risk of copyright infringement.
Potential copyright infringement issues associated with using tools which rely on existing music for its training data could be avoided by using only works in the public domain, or through the use of formal licensing arrangements. However, a single song may have many rightsholders (i.e. the writers, performer, publisher, record label, etc.) and clearing use for AI training by way of licensing will require a number of different licences for the same song.
Outside of the music industry, we have already seen strategic partnerships forming between technology companies and rightsholders, particularly news organisations and other media companies, as they negotiate formal licensing arrangements which benefit both parties. For example, OpenAI is currently engaging in licensing agreements with a number of media publishers such as the Financial Times and Associated Press to use their content as AI training data. We expect to see a similar trend emerging in the context of use of existing musical works as AI training data, potentially in the form of a blanket licensing regime being adopted by key music collecting societies.
Various stakeholders across the music industry have been grappling with the use of unauthorised AI-generated content on online platforms. Recently, Universal Music Group (UMG) and Meta Platforms announced an expansion of their existing partnership, extending their licensing arrangement to cover use of UMG-owned music across a number of Meta’s platforms, including Facebook, Instagram, Messenger, Horizon, Threads and WhatsApp. While the terms of the partnership agreement have not been published, when announcing this partnership, Meta indicated that the partnership seeks to address the use of unauthorised AI-generated content, recognising this could affect artists and songwriters. Similar collaborations between publishers/labels and online platforms may become more prevalent in future, with stakeholders taking an increasingly formalised approach to use of existing music to create AI-generated content.
Exceptions to Copyright Infringement
There are various statutory exceptions which may allow for the use of works protected by copyright to train generative AI models without the risk of infringement:
- Text and Data Mining (TDM)
- Research for non-commercial purposes (s.53A)
A person can make a copy of a copyright protected work where: (1) they have lawful access to the work, (2) the copy is made in order to carry out a computational analysis of anything in the work for the sole purpose of research for a non-commercial purpose, and (3) it is accompanied by a sufficient acknowledgement.
This exception reflects Article 3 of the EU Directive 2019/790 on copyright and related rights in the Digital Single Market (EU Copyright Directive).
The Irish exception under s.53A only applies in cases which do not conflict with the normal exploitation of the work and do not unreasonably prejudice the legitimate interests of the rightsholder.
In practice, this exception for non-commercial research is unlikely to be useful for training AI models using copyright protected works. This is because the vast majority of companies developing large generative AI models intend to commercialise them and/or their outputs.
- Express opt-out (s.53B)
A person can reproduce a copyright-protected work for the purposes of TDM where: (1) they have lawful access to the work but do not fall within the scope of section 53A (i.e. TDM for non-commercial research purposes), and (2) the author has not expressly reserved in an appropriate manner the use of the work for the purposes of TDM.
This express reservation must be machine-readable in the case of content made publicly available online or, in the case of content not made publicly available online, clearly communicated to all those with lawful access to it.
This exception reflects Article 4(3) of the EU Copyright Directive.
The EU Artificial Intelligence Act (AI Act), which came into force on 1 August 2024, requires providers of general-purpose AI models to put in place a policy to comply with EU law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, reservations of rights expressed pursuant to Article 4(3) of the EU Copyright Directive. This obligation will apply 12 months after the AI Act comes into force.
We have already seen some industry stakeholders make ‘opt-out’ declarations against the use of their content, including through TDM, for AI training. For example, Sony Music Group (SMG) recently published its ‘Declaration of AI Training Opt Out’, in which it clearly prohibits and opts out of any text or data mining, web scraping or similar reproductions, extractions or uses of any of its content (or its affiliates’ content) for any purpose without its authorisation. SMG has also reportedly reached out directly to hundreds of technology companies, warning them not to use its content to train AI.
In Ireland, the Irish Music Rights Organisation (IMRO) similarly published a statement on 5 June 2024 which warns that any use of copyright protected musical works and associated literary works or data that it administers, for AI training, without a valid licence infringes the copyright in those works. IMRO goes on to expressly reserve the right to engage in text and data mining in accordance with the CRRA and Article 4 of the EU Copyright Directive.
Notably, both of the Irish law TDM exceptions considered above are conditional on the copier having ‘lawful access’ to the relevant work. What this means in practice is that circumventing paywalls in order to access content, or otherwise using content in a way that violates a provider’s terms of service, will preclude reliance on these exceptions.
- Fair Dealing
Irish copyright law provides for a broad ‘fair dealing’ exception, which permits the use of copyright protected works for certain activities (e.g. for the purpose of research or private study, caricature, parody or pastiche, criticism or review, etc.).
While most of the activities permitted under this exception are unlikely to be relevant for those seeking to train AI models, some may seek to rely on the exception for ‘research or private study’. However, this exception is unlikely to be helpful to most AI technology companies as it is very narrow and in order to rely on it the use in question must not unreasonably prejudice the interests of the relevant rightsholder(s).
- Temporary Acts of Reproduction
Under Irish law, it is not an infringement of copyright to undertake or conduct temporary acts of reproduction which are transient or incidental and which are an integral and essential part of a technological process. This exception only applies where the relevant act has no independent economic significance and where its sole purpose is to enable (a) a transmission of a work in a network between third parties by an intermediary, or (b) a lawful use of a work.
This exception is quite narrow and is intended to cover caching and acts that enable browsing of content over the internet, and is unlikely to cover the use of copyright protected works for the training of AI models.
Irish law also contains specific provisions in relation to the authorship of computer-generated works. A computer-generated work is defined under the CRRA as a work that is “generated by computer in circumstances where the author of the work is not an individual” and the author of such a work is “the person by whom the arrangements necessary for the creation of the work are undertaken”. This definition is vague and could potentially capture a number of different people involved in the creation of AI-generated works, including the developer of the AI model itself, the individuals responsible for training the model, the company that owns the product, the individuals whose work is used to train the model or the individuals who input prompts into the tool.
Copyright protection in Ireland is automatic and arises upon the creation of an original work, giving authors exclusive rights to use and distribute their creations. The requirement of originality necessitates the work containing some element of the author’s own intellectual creation. However, where an individual’s creative process involves the use of AI systems, determining whether the test for originality is satisfied (and, indeed, whether it ever really can be where AI tools are relied on) is complex and will be highly dependent on the facts of each individual case, including the level of input an individual has and their ability to make creative choices as well as the capabilities of the AI tool being used.
Copyright may subsist in certain elements of a musical work, but not others. For example, if a user inputs their own lyrics into an AI tool with a prompt to create an upbeat pop song using those lyrics, the user will (unless otherwise agreed by contract) own the copyright in the lyrics regardless of the position in respect of the generated music.
Some providers of AI tools seek to allocate copyright ownership through their terms and conditions with users. However, the extent to which this is actually effective as a matter of law remains to be seen. If no copyright subsists in the output of an AI tool then contractual provisions purporting to assign or otherwise deal with it will have no legal effect on the ownership of that output.
Conclusion
With the implementation of the AI Act on the not-so-distant horizon, developers and providers of AI tools will need to ensure that the methods they adopt to train their AI models are compliant with applicable copyright law, and their own policies as required under the AI Act. They must respect express reservations by rightsholders (like Sony and IMRO) in respect of the use of their works for the purposes of TDM, or face potential copyright infringement claims.
In terms of output, ownership of copyright in works created using complex generative AI tools is yet to be tested before the Irish courts, and it remains to be seen whether it is possible for copyright to subsist in such works and, if so, who is the author and first owner of that copyright.