Asia’s current approaches to TDM exceptions

By Qi Jun Kwong (Affiliated Research Fellow, RCLIP; Hokkaido University)

The first snow has just arrived in Sapporo! With the year seemingly hurrying to a close, I thought it’d be a good continuation from Dr Saliltorn’s previous post to just highlight how some other Asian jurisdictions have been handling the use of copyrighted inputs. As of now, Japan and Viet Nam are the two Asian jurisdictions with an express exception for text and data mining (TDM) activities, while Singapore’s computational data analysis exception and fair use provision also permit such use within limits. Most others, in the absence of such exemption, would likely rely on the prior consent of the copyright holder, or on narrowly worded scope under fair dealing provisions.

1. Japan’s recent soft law approaches

Japan’s TDM exception, as provided under Article 30-4 of the Copyright Act, permits exploitation of a work for testing for the development or practical use technology, data analysis, or computer data processing as long as the purpose is not to enjoy, or have another person enjoy, the thoughts or sentiments expressed in the work; and provided that it does not unreasonably prejudice the interests of the owner (see also Art. 47-5 on minor exploitation for computerised information processing). Unlike the UK’s exception, for example, it neither requires lawful access nor is limited to just non-commercial purposes.

Since the 2018 amendment, Japan has initiated a series of soft-law approaches on broader AI governance, particularly after the AI Promotion Act of 2025 (Act No. 53 of 2025). More recently, in August 2026 the IP Strategy Promotion Secretariat of the Cabinet Office (the secretariat to the IP Strategy Headquarters set up under the 2002 IP Basic Act) published a Principles-Code aimed at generative AI providers (provisional translation here and official summary here).

Under the Principles-Code, AI providers are expected to publish on their websites an outline of their AI models, training data, relevant internal decision-making processes and the measures they take to protect IP rights, and to respond to disclosure requests from rightsholders preparing legal proceedings (including ADR) and from users who want to know whether something resembling their output is in the training data. Providers that do not comply are expected to explain why, a comply-or-explain approach the code expressly borrows from the “Stewardship Codes in the field of corporate governance.” Japan’s Stewardship Code asks institutional investors to act as responsible stewards of their clients’ assets, so I suppose AI providers are now expected to act as responsible stewards of other people’s works.

As for the ongoing lawsuit brought by Asahi and Nikkei against Perplexity in August last year – no updates as of yet.

2. Viet Nam’s TDM exception and Singapore’s computational data analysis exception

Viet Nam’s amended IP Law, in force since April 2026 (see implementing Decree 134), introduces an exception for the use of texts and data in scientific research, testing and the training of AI systems. The use must be non-commercial; source materials must be lawfully published and lawfully accessed; and the use must not conflict with the normal exploitation of the work, unreasonably prejudice the rightsholder’s legitimate interests, or produce outputs that substitute for the original in the market or create unfair competition. While it is termed a TDM exception, what Viet Nam has is in some respects closer to Singapore’s computational data analysis exception, particularly on the lawful access requirement. But it is also narrower since Singapore’s exception carries no non-commercial limitation.

On Singapore’s exception, the copy must be made for computational data analysis and not used for any other purpose, may be supplied to others only for verifying results or for collaborative research, the user must have lawful access to the material, and the source must not be an infringing copy. This is a stand-alone defence separate from the general fair use provision (which replaced the former fair dealing provision), though how either defence would play out in an infringement dispute remains untested. Singapore has also opened a public consultation on AI and Singapore’s IP regime. On the copyright side, it asks whether the scope of the computational data analysis exception is sufficiently clear and how lawful access and rightsholder safeguards should work in practice; on the patent side, it covers inventorship and prior art.

3. Malaysia: Looking East or South?

Malaysia is now weighing both the Japanese and Singaporean models. MyIPO’s public consultation on amending the Copyright Act 1987 closed in August 2026, and an AI Governance Bill is in the works. The consultation paper sets Japan’s exception (framed as a “purpose and market harm” approach) against Singapore’s (framed as a “lawful access” test).

Malaysian copyright law traces its roots to UK law, with one of the earliest protections dating back to the 1902 Telegram Copyright Ordinance. On amendments to existing statutes Malaysia has tended to look to the UK or its common law cousins, so looking to Japan cuts across the usual common law–civil law divide (and some of the feedback cites, of course, Waseda’s very own Professor Tatsuhiro Ueno). Would we see a case of Japanese legal transplantation in Malaysia? Malaysia’s own MyIPO has adopted many of JPO’s approaches in streamlining the local patent registration system, but referring specifically to a Japanese statute might be a first.

4. Indonesia: emphasis on royalties  

In Indonesia, a public consultation on the Copyright Bill has held in May. There is no TDM exception proposed; instead, the emphasis is on royalty arrangements. According to Indonesia’s National Collective Management Institute, the House of Representative’s committee-stage deliberation of the draft law is targeted to finish by 17 December 2026, so we may know more by the year’s end.

5. Elsewhere in the region

Hong Kong launched a public consultation in 2024 on a proposed TDM exception. By May 2026, it appears that the Administration has moved into drafting a code of practice, and “there is no concrete timetable for submitting the legislative proposal.”

Korea has also opted for guidance over legislation. The Ministry of Culture, Sports and Tourism’s fair use guidebook on generative AI surveys approaches elsewhere, including those of Japan and in Singapore, while noting that under Korea’s fair use provision, AI training will be assessed case by case.

China’s Supreme People’s Court in September 2026 released the Opinions on adjudicating disputes involving AI. The Opinions give no general answer on whether unauthorised use of copyrighted works is lawful, but they require a developer claiming non-infringement to produce evidence of its training data sources, training records and how the model operates.

Taken together, Asia is not converging on a single model so much as experimenting with several at once. Given how far IP law has already been harmonised internationally, I remain optimistic that these approaches will eventually converge. Will we one day see a cohesive, Berne-style international framework for AI training?‍ ‍

Next
Next

Thailand's Emerging Text and Data Mining Exception: Between Copyright Orthodoxy and AI Competitiveness