Home

Donate

AI Training and Copyright Infringement: Solutions from Asia

Seth Hays / Oct 30, 2024

Seth Hays is managing director and co-founder of APAC GATES, a non-profit management and rights advocacy consultancy based in Taipei.

Tokyo, Japan. Justin Hendrix/Tech Policy Press

Strategic litigation in the US and Europe between copyright owners and generative AI companies about the use of copyrighted works to train AI models is ongoing. While major litigation has not come to the shores of countries in Asia vying for the AI industry, more can be done to protect the rights of creatives while proactively supporting an emerging AI industry in this diverse region. Examining the current state of the play sheds light on the core economic and political interests of the debate and points to innovative solutions to prevent disputes between these two important industries in the future.

Singapore and Japan: Global AI Training Heaven or Haven?

Singapore and Japan have some of the most liberal copyright rules in relation to AI training in the world. Both countries’ copyright laws appear to allow copyrighted material to be used for training AI systems. Unlike the UK and EU copyright acts, which also have text and data mining (TDM) exceptions to infringement, however, Singapore and Japan’s rules allow for a commercial-use exception to infringement.

Singapore’s Copyright Act Amendment of 2021 introduced Section 244, which allows the use of copyrighted works for the purposes of “computational data analysis” (CDA). Of course, copyright exceptions usually have qualifications, and the Singapore CDA exception stipulates that the user of the works must have “lawful access” – or if access was unlawful, the user must not have knowledge of the unlawful access (244(2)e).

Singapore has made no secret of its support for the AI industry, and the country’s National AI Strategy indicates that it wants to be a regional, if not global, leader in the industry. It plans to train 10,000 AI engineers and has launched world-leading AI governance initiatives such as the AI Verify Foundation, which is an AI governance testing framework and software toolkit for benchmarking AI systems against a set of internationally recognized principles offered to AI developers.

Recent guidance published by Singapore on generative AI is insightful on the government’s views of the relationship between copyright and AI. The guidance proposes that access to data needs to be “balanced” with interests in copyrighted works under the pillar of fostering a trusted AI ecosystem. Given the large amounts of data needed to train state-of-the-art generative AI models, the guide suggests developing “approaches” to resolve conflicts in a “clear and efficient manner.”

Implicit in the notion of “balance” is that both sides come to the table as equals in terms of rights and liabilities. It appears that the vagueness of the guidance indicates that parties will be left to fend for themselves to resolve any issues – perhaps before the courts at some point.

Singapore’s Economic Interest at Heart

Singapore’s copyright CDA exception was introduced to support the country’s technology industry according to the legislative history of the copyright amendment. At the time of the second reading of the law in 2021, Second Minister for Law Edwin Tong said, “[The CDA exception] supports our Smart Nation initiatives, our push towards data-driven innovation, and Singapore’s efforts to grow our Artificial Intelligence and technology sectors.”

The economic data support such a decision. According to research by the World Intellectual Property Organization (WIPO), the economic contribution to the GDP of copyright-related industries in Singapore stood at around 5%. However, according to the InfoCom Media Development Authority, the digital economy represents over 17% of the GDP in Singapore – of which AI is a growing part.

Even the small island nation’s community of writers and artists is self-deprecating on the debate between AI and copyright. Recently, the Singapore government surveyed the community about having their works used to train local language LLMs that reflect Singapore’s unique culture and the region’s linguistic and cultural identity. “There is very little we can do,” said one author about policies to allow for data training, according to a recent article.

Japan’s long-ruling but recently imperiled Liberal Democratic Party (LDP) issued white papers on its AI policy in 2023 and 2024. Notably, the two papers show a clear shift from balancing AI and copyright to clearly favoring the AI industry.

For example, the first white paper notes: “Regarding the discussions over the interpretation of intellectual property laws [of which copyright is a subset] in relation to generative AI, consider establishing guidelines, etc., to promote the progress of AI technology while preventing its abusive use, and further develop the content industry, which is one of Japan's strengths.”

But without mincing words – the title of the second white paper clearly shows the direction policymakers want to move: “Towards the most AI-friendly country.”

When it comes to protecting copyright owners from infringement, the LDP policy is rather meek, suggesting the government promote “appropriate AI utilization” and “awareness” of recommendations by a study group on intellectual property (IP), which encourage contractual remuneration between the AI industry and copyright owners.

In 2019, Japan amended its copyright act to include the AI-friendly copyright exception for data training:

Article 30-4: It is permissible to exploit a work, in any way and to the extent considered necessary, in any of the following cases, or in any other case in which it is not a person's purpose to personally enjoy or cause another person to enjoy the thoughts or sentiments expressed in that work…[emphasis added]

Unlike Singapore’s CDA or the UK and EU’s TDM exceptions, Japan’s copyright act focuses on the human experience – the enjoyment of the expression embodied in the work – rather than on a technical carve out for computers and data along the lines of the Singapore and European rules.

Of course, the statute also qualifies the exception saying the exploitation must not “unreasonably prejudice the interests of the copyright owner.” This is where any future litigation may focus.

Japan’s Copyright Office issued guidance on AI and copyright earlier this year. The report emphasized that the 30-4 exception for “non-enjoyment use” generally does not harm the economic interests of the author of the copyrighted work used in training. The guidance also noted that “non-enjoyment use” does have limits, such as where generative AI models “overlearn” – or simply output copyrighted material.

Concerns from Japan’s Creative Industry

The Nippon Anime and Film Culture Association published comments in response to the Japanese government’s “IP in the AI Era” consultation paper earlier this year. The association urged the use of watermarking to identify AI-generated material and wants stronger protection for neighboring rights, such as the right of publicity, to protect recognizable voice actors from exploitation by generative AI.

Politicians are trying to thread the needle between the creative industry and the AI industry. Recently, a manga artist-turned-lawmaker in Japan voiced support for generative AI companies to compensate artists whose works were used in training models by setting aside a percentage of their revenues to support the arts.

While domestically, it may seem that the AI industry is beating out copyright owners, Japan has taken a leading role in the global conversation on AI governance and highlighting AI’s intellectual property issues. For example, Japan led the G7’s adoption of the Hiroshima Process – an 11-point framework for governments and companies to adhere to for ethical AI development.

In the final point of the framework, the Hiroshima Process specifically raises the need to protect copyright. Compare this to other international governance initiatives, such as the UK’s Bletchley Declaration, which does not mention IP at all, or UNESCO’s Recommendation on the Ethics of Artificial Intelligence, which only raises IP once among 144 recommendations and only calls on UN member states to “assess” the impact of AI on IP rights holders.

Japan has taken the Hiroshima Process’s IP concerns to heart. The Ministry of Economy Trade and Industry’s AI Guidelines For Businesses published this year advises Japanese firms to “Implement appropriate data input measures…to honor rights related to privacy and intellectual property, including copyright-protected content.”

While Japan’s copyright rules and the ruling party’s white papers indicate a pro-AI policy, the situation on the ground in Japan may respond to actual disputes that arise between copyright owners and AI companies and the political blowback, if any, from such disputes.

Japanese policymakers will want to see real returns from the AI industry for the economy, and to assess any real damage done to the creative industry before it becomes more proactive in the debate. It is likely no accident that OpenAI’s first office in Asia was opened in Japan earlier this year.

The copyright rules elsewhere in Asia do not have Japan and Singapore’s pro-AI exceptions to infringement, but some jurisdictions are seeing legislative changes on the horizon, while others see the courts address copyright’s related rights issues.

China’s courts have addressed generative AI issues more quickly than others in the region – and, indeed, globally. The IP docket in China is the largest in the world – and courts dispose of cases quickly, making it an interesting testbed for the redress of harms between IP and AI. Most recently, Chinese courts found a generative AI platform infringing on the rights of a voice actor when it trained its model on the actor’s voice without compensation.

In general, China’s regulations around AI have included mention of respect for IP. For example, China’s “Interim Administrative Measures for Generative Artificial Intelligence Services” raises the need to follow IP rules, although not with much regulatory clarity. Specifically, Article 7.2 stipulates that AI providers should not infringe others’ intellectual property rights.

In Hong Kong, the IP Department recently issued a consultation paper on proposed changes to the city’s copyright law. Namely, the changes will create an infringement exception similar to those found in Japan and Singapore – including for commercial use. The favorable view for amendment is summarized by the consultation paper:

Given the overall benefits brought by the Proposed TDM Exception, in particular in driving and boosting the development of AI technology and industry, and considering that most of the possible drawbacks can be balanced out by providing appropriate safeguards to copyright owners, the Government is of the view that it is justifiable to introduce the Proposed TDM Exception to the CO.

South Korea Values the Creative Industry but Wants More AI

South Korea presents an interesting case study in balancing copyright and AI for policymakers. Unlike Singapore, Korea is a copyright heavyweight. According to a WIPO study, copyright-related industries contribute nearly 10% to Korea’s GDP. It goes without mentioning the economic importance and global impact of Hallyu – or the Korean Wave – including K-pop musical acts or K-drama movies and TV.

But the country is also trying to attract the full lifecycle of AI to its shores, recently promising to be one of the “top 3” countries in AI. Part of the policy mix has seen efforts to amend the copyright act and bring it into more AI-friendly territory. In 2021, a bill was introduced to add a TDM exception to the Korean copyright act. It has yet to pass.

In the meantime, the Korean Copyright Commission has taken a proactive stance on mediating the interests of copyright owners and the generative AI industry. In 2023, the government issued a Guide on Generative AI and Copyright. The guide encourages AI businesses to obtain the rights to works before using them for training by paying adequate compensation, and to proactively prevent infringement in genAI outputs. It encourages rightsholders to opt out of training if they desire, and the guide encourages users of genAI tools not to use prompts that induce the output of infringing material. However, it also notes that the current copyright act’s fair use provisions have not been interpreted as to the extent copyrighted works can be used to train AI without permission.

The Philippines has taken a similar approach to raising awareness between both sides of the copyright and AI debate. Earlier this year, the Philippines IP Office published a guide on fair use – which can be seen as potentially addressing the AI training issue, although the guide did not mention AI training specifically. Diplomatically, one official at the launch of the Guidelines said, “If, despite the guidance offered by the Guidelines, one is still in doubt, err on the side of caution. Obtain permission.”

Industries Lobby in Australia and India

Australia’s Attorney General’s Office established a Copyright and AI Reference Group in 2023. The group's goal was to prepare copyright rules for challenges presented by AI. It specifically addresses the question of AI training on copyrighted works. Australia also recently published a Voluntary AI Safety Standard, which includes 10 specific guardrails. The guardrails mention IP under data governance but only call for documenting “data usage rights.”

At the same time, the Australian Senate has conducted a series of inquiries on “Adopting Artificial Intelligence.” At the most recent inquiry, Amazon and Google both lobbied for amendments to the copyright law easing the way for AI training on copyrighted material.

While the AI industry is lobbying for copyright amendments in Australia, India is seeing the copyright industry lobby for their interests.

Earlier this year, a group of publishers in India, the Digital News Publishers Association, requested the government to amend rules to specifically give content publishers royalties when their works are used for data training.

How to Get Rights-Respecting Innovation Policy in Asia

Asia’s diverse economic and cultural landscape is mirrored in the diversity of the region’s policy approaches to AI and copyright rules. While litigation continues outside the region, more can be done to protect the rights of creatives in Asia while still promoting innovation for the AI industry.

At the heart of the copyright infringement question in AI training is the value of labor and receiving just recognition for the product of human effort.

To better address AI-related harms to the creative industry, organizations such as copyright management organizations, individual creators, and civil society organizations representing the rights of artists should organize more intentionally on a cross-border level and share on-the-ground instances of infringement, compensation, negotiation, and agreement when it comes to training AI models on copyrighted material.

Public awareness needs to be maintained about the nature of the harms that artists are suffering from infringement and any economic harms that are inflicted. Additionally, a well-functioning copyright system in the era of AI will require that judges be trained further on the technology of AI in anticipation of cases that will eventually come before them.

Copyright agencies around Asia should be proactive and provide specific guidance about copyright infringement and AI training to creators and the AI industry, as they already have in Japan and South Korea, in order to prepare both industries for interacting in an ethical and mutually beneficial way, according to the standards set by local IP rules and case law.

In the meantime, policymakers should look to peers, foster interoperability of rules and guidelines, and ensure that respect for copyright is emphasized in government guidance.

Regional organizations and platforms such as the APEC IP Rights Experts Group, the ASEAN Working Group on IP, or discussions at trade groupings, such as the CPTPP, offer ample opportunity to continue talking specifics about best practices in AI and copyright while also keeping track of the exact harms being experienced and the current state of the science and technology of AI.

More broadly, solutions to help prevent harm to creatives should be applied to other types of AI harm, such as human rights, environmental, consumer safety, or other rights infringement. The Asia-Pacific AI Harms Remedy Network is an example of a civil society effort to assist those whose rights have been infringed and to hold the AI industry accountable across borders by monitoring, cataloging, and sharing specific cases - not just copyright infringement cases - in order to raise public awareness and identify what works and what does not work to foster a practical, rights-enhancing AI governance framework in Asia, and globally.

Authors

Seth Hays
Seth Hays is Managing Director and Co-founder of APAC GATES, a non-profit management and rights advocacy consultancy based in Taipei. He also serves on the board of the non-profit organization Digital Governance Asia. Seth has worked in the public interest and non-profit sector in Asia for over two ...

Topics