The overlap between Singapore copyright law and AI governance

Yuen Kit Kuan

Drew & Napier, Singapore

kitkuan.yuen@drewnapier.com

Introduction

The widespread adoption of artificial intelligence (AI) has resulted in questions about copyright infringement issues in the context of generative AI-related use of third-parties’ works that enjoy copyright protection.

This article will examine key aspects of the Singapore Copyright Act 2021 (the ‘CA 2021’) that permit otherwise infringing uses of copyrighted works in Singapore in the context of AI.[1] It will then analyse how AI governance compliance can help in regard to the application of CA 2021 in this area.

The computational data analysis exception

Section 243 of CA 2021[2] defines computational data analysis (CDA) to include: (1) ‘using a computer program to identify, extract and analyse information or data from the work or recording’; and (2) ‘using the work or recording as an example of a type of information or data to improve the functioning of a computer program in relation to that type of information or data’.

Section 244 of CA 2021[3] provides that it is not copyright infringement if a person makes a copy of a work or a recording of a protected performance (the ‘copier’) where the following key conditions are met (the ‘CDA exception’):

• The copy must be solely made for the purpose of CDA or preparing the work or recording for CDA.
• The copy must not be supplied to any person other than for the purpose of: (1) verifying the results of the CDA carried out by the copier; or (2) collaborative research or study relating to the purpose of the CDA carried out by the copier.
• The copy must be made from lawfully accessed material, such as not circumventing paywalls or acting in breach of the terms of use of a database. 
• The copy must not be made from material that is an infringing copy and, if it is an infringing copy, it must satisfy other requirements like the copier’s knowledge of the said infringement.

Section 187(1)(c) of CA 2021[4] provides that any contract term is void to the extent that it purports, directly or indirectly, to exclude or restrict any permitted use under the CDA exception.

Currently, these are the only AI-specific provisions in CA 2021.

Due to the specific nature of the CDA exception, it is likely to only apply in situations involving input-related infringements, weights-related infringements and, depending on the facts of the case, dealings-related infringement.[5]

The fair use provision

Section 190 of CA 2021[6] provides that it is not copyright infringement to make fair use of the copyrighted work, the performance or a recording of the performance (the ‘fair use provision’).

Section 191 of CA 2021[7] highlights the following non-exhaustive list of factors that should be considered in determining whether the fair use provision applies:

• the purpose and character of the use, including whether the use is of a commercial nature or is for non‑profit educational purposes (the ‘purpose factor’);
• the nature of the work or performance (the ‘nature factor’);
• the amount and substantiality of the portion used in relation to the whole work or performance (the ‘substantiality factor’); and
• the effect of the use upon the potential market for, or value of, the work or performance (the ‘market factor’).

The open-ended fair use provision is likely to be applicable in: (1) situations involving input-related infringements, weights-related infringements and, depending on the facts of the case, dealings-related infringements, whereby the CDA exception does not apply; (2) situations involving output-related infringements; and (3), depending on the facts of the case, dealings-related infringements.[8]

Uncertainty in application

To date, there is no case law in Singapore involving AI-related copyright infringement issues or Singapore cases that consider the fair use provision in any context since it was introduced in CA 2021.

As such, there is a high level of uncertainty as to the application of CA 2021 in situations involving AI-related copyright infringement.

In relation to the CDA exception, this uncertainty is exacerbated as key terms defining the conditions for relying on it, like ‘supply’, ‘collaborative research or study’ and ‘lawful access’ are not defined.[9] Further, it is unclear whether Section 187(1)(c) of CA 2021 voids a contractual term that merely regulates access to a work (as opposed to excluding or restricting such access),[10] and the extent to which this provision can be circumvented by relying on technical access control measures like machine-readable directives.[11]

There are also territorial issues depending on where the training data is sourced from, hosted and deployed in training the AI model. Given Singapore’s small population size, there is a relatively small corpus of data that is available to train and develop AI models in Singapore. Further, given the labour-intensive nature of training and developing AI models, much of the work is likely to be carried out in low-income and middle-income countries.[12]

In the UK case of Getty Images v Stability AI,[13] Getty Images had abandoned the claim of primary copyright infringement prior to the giving of closing submissions as there was no evidence that the training and development of Stability AI’s AI model took place in the UK.[14] Similarly, much of the training and development of AI models, which are deployed in Singapore, is unlikely to take place in Singapore.[15]

The narrow nature of the CDA exception, as well as the numerous conditions that need to be satisfied to successfully rely on it, means that the CDA exception would rarely be invoked, much less successfully relied on. As such, the uncertainty surrounding the scope of its application is likely to remain unsettled for a long time.

The fair use factors in the fair use provision heavily mirror the factors to be considered in the fair use provision in the US.[16] In analysing the fair dealing provision,[17] which is the predecessor of the fair use provision under the previous Singapore Copyright Act, the Singapore court had specifically commented that US jurisprudence would be helpful in shaping Singapore’s law on fair dealing.[18] As such, US case law involving AI-related copyright infringement issues is likely to be helpful in analysing Singapore’s fair dealing provision.

There are some existing US cases involving AI-related copyright infringement issues like Bartz v Anthropic;[19] Kadrey v Meta Platforms;[20] and Thomson Reuters v ROSS.[21] However, there is still little certainty on how the fair use provision in the US would be applied in such situations. Fair use analysis is highly dependent on factors like the nature of the copied work, how the AI model uses the work and evidence of the market effect.[22]

For example, in regard to the purpose factor, the fair use provision is more likely to be successfully invoked in a Bartz/Kadrey‑type scenario, which involves the building of a generative AI tool that performs text analysis and generates content in response to prompts, than a ROSS-type scenario, which involves the building of a competing AI search engine tool that generates judicial opinions in responses to prompts. However, there is a whole spectrum of use to be considered, which depends on the functionality and deployment of the AI model, as well as the commerciality of its use.[23]

In relation to the market factor, aside from a ROSS-type scenario, which involves a direct market substitute tool for conducting research, uncertainty remains as to the extent to which evidence of indirect market harm, like market dilution and lost revenue in licensing markets, should be considered.[24]

AI governance

Since 2019, Singapore has established three non-legally binding AI governance frameworks (collectively referred to as ‘AIGFs’): (1) the Model AI Governance Framework; (2) the Model AI Governance Framework for Generative AI (‘AIGF 2024’); and (3) the Model AI Governance Framework for Agentic AI. 

All three AIGFs explain how AI governance principles, like transparency, explainability and accountability, should be adhered to when developing and deploying AI systems. However, AIGF 2024 is the only AIGF which expressly refers to copyright infringement issues (albeit briefly and focused on the context of balancing copyright against data accessibility in the ‘data’ dimension of AIGF 2024).[25]

AI Verify is an AI governance testing framework and software toolkit that helps to operationalise the AIGFs.[26] It enables organisations to assess the implementation of their AI systems against 11 AI governance principles and international governance standards, like the International Organisation for Standardisation (ISO) ISO/IEC 42001:2023.

There is no law, guidelines or other authoritative statements that state: (1) AI governance compliance must be considered when determining issues involving AI-related copyright infringement; (2) AI governance compliance would enable one to successfully invoke the CDA exception or the fair use provision; or (3) AI governance non-compliance would bar one from relying on the CDA exception or fair use provision in an AI-related copyright infringement case.

However, evidence of AI governance compliance may be able to aid in the application of CA 2021 provisions in situations involving AI-related copyright infringement. This would be similar to the transparency and technical documentation-related requirements under Article 53 of the European Union’s AI Act[27] and the Copyright Chapter of the EU Code of Practice for General-Purpose AI Models,[28] which expressly links copyright infringement with AI governance compliance. This situation is also echoed by recent AI-related copyright infringement cases in China, which seem to suggest that there is an emerging copyright alignment obligation for AI service providers to take AI governance-related measures to mitigate copyright infringement risks.[29]

For example, in regard to the ‘data’ and ‘content provenance’ dimensions in AIGF 2024, clear documentation on the scope, lineage and provenance of AI training data can be relevant to the analysis of the nature factor and the substantiality factor if the fair use provision is invoked, as well as in the assessment of whether the ‘lawful access’ requirement of the CDA exception is satisfied.

Regarding the ‘trusted development and deployment’ dimension in AIGF 2024, evidence of the use of guardrails like filters to block the generation of outputs that are too similar to copyrighted works and training techniques designed to make copyright-infringing outputs less likely[30] can be relevant to the analysis of the substantiality factor and the market factor if the fair use provision is invoked.

Conclusion

There is legal uncertainty around how the CDA exception and the fair use provision would be applied in a case involving AI-related copyright infringement. In the absence of any legislative change to the copyright regime in Singapore, this uncertainty is likely to persist.

Nevertheless, AI governance compliance provides operational context to the management of copyright infringement risks arising from the development and deployment of AI systems. This can inform the legal interpretation of the factors to be considered in AI-related copyright infringement cases.