Asian Publishers Week
Asia Publishing Insight
Q&ATaiwan
Taiwan Has No AI Copyright Rules Yet. Publishers Can’t Wait for Them.
Drawing on recent cases in Taiwan where copyright law, court rulings and policy are making legal risks concrete, the author outlines the legal safeguards publishers need to put in place for the age of AI.
Chun-Wei Tseng Attorney, Taiwan
Question 1. How is AI training data currently treated under Taiwan’s copyright law?
Taiwan has not yet enacted copyright rules specifically governing “AI training data,” and no court has issued a ruling directly determining whether the use of AI training data is lawful. At this stage, therefore, the legal status of AI training data must be assessed under Taiwan’s existing Copyright Act.
1. AI training may reproduce source materials and potentially infringe the copyright owner’s reproduction right
Taiwan’s Intellectual Property Office stated in an August 29, 2025, interpretation that activities undertaken during AI training—including data collection, preprocessing, model training and optimization—may involve the “reproduction” of source materials and therefore may potentially infringe a copyright owner’s reproduction right.
2. AI-generated outputs may infringe the copyright owner’s right of adaptation
In the same August 29, 2025, interpretation, the Intellectual Property Office explained that where AI makes no separate creative contribution to an original work, the process does not constitute an “adaptation” and therefore does not, on that basis, raise an infringement issue involving the copyright in the original training data. Conversely, if works contained in a training database are “collaged” or “recombined” and the resulting output is substantially similar to the original training work, the output may infringe the copyright owner’s right of adaptation.
3. Copyright protection for AI-generated output depends on sufficient human creative contribution
With respect to AI-generated content, the Intellectual Property Office stated in interpretations issued on June 11, 2018 and April 1, 2025 that purely AI-generated works with no human creative input are not protected by copyright. Where a human makes an actual creative contribution and AI is used merely as an assisting tool, however, the resulting work may still qualify for copyright protection.
Question 2. Have there been significant court decisions or legislative changes in Taiwan?
1. Significant court decision
Taiwan has not yet produced a court ruling directly addressing the legality of AI training data. There has, however, been a case involving the use of web crawlers to copy a legal database in which the court indicated that the fact that data were acquired “to train AI and enable AI to learn” cannot, by itself, serve as a blanket defense. Taiwanese courts have also held that where reproduction of another party’s work negatively affects that party’s market, the reproduction may fail the fair-use test and constitute copyright infringement.
New Taipei District Court ruling
In New Taipei District Court Judgment No. 111-Zhi-Su-8, issued June 24, 2025, and currently under appeal, Qifa Co., Ltd. (operator of the “Lawsnote” legal database service) and two founders were prosecuted for using web-crawling software to copy and download 98,068 records of legislative histories from another legal database, place them in their own database, and monetize access through monthly membership fees. The court found that the conduct constituted a copyright offense involving unauthorized reproduction for the purpose of sale. The two founders were sentenced to four years and two years’ imprisonment, respectively; approximately NT$4.625 million and NT$3.7 million in criminal proceeds were confiscated, and the legal database company was fined NT$1.5 million.
Reference to a U.S. case
The district court further cited Thomson Reuters v. Ross Intelligence in the United States, a case involving extensive copying of legal database materials to train an in-house AI-powered legal research system. The court noted that the defendant’s product could substitute for the complainant’s product in the market and could affect the potential market for licensing the complainant’s data as AI training material. On that basis, it concluded that the defendant did not satisfy the requirements for fair use. The decision also emphasized that using data “to train AI so that AI can learn” is not, as a general rule, a blanket exemption from liability.
2. Legislative developments
As of now, Taiwan has not enacted dedicated copyright rules for AI training data. A potentially important policy development is the Artificial Intelligence Basic Act, promulgated and brought into force on January 14, 2026. Article 13 requires the government to improve the quality and quantity of AI data while protecting intellectual property rights. Article 18 requires the competent authorities to review and adjust relevant laws and administrative measures within two years. More specific copyright rules concerning AI training data could therefore emerge and take effect within that period.
Question 3. What legal measures is Taiwan’s publishing industry currently taking?
Taiwan’s publishing industry does not yet have a unified AI-use policy or a collective licensing system for copyrighted works. Individual publishers and media organizations are nevertheless taking several measures.
1. Publicly prohibiting AI use
Some media organizations have begun expressly prohibiting users, through membership terms or copyright notices, from providing articles, images, audio, video, or database content to generative AI systems for training, fine-tuning, generation, or derivative uses. For example, the membership terms of Commonwealth Magazine Group include restrictions of this kind. They prohibit members, without the group’s express written authorization, from inputting, uploading, integrating, or otherwise providing the group’s content to generative AI systems, platforms, or models for training, fine-tuning, generation, reuse, or other derivative purposes. Where cooperation or licensing is required, members are instructed to contact the group in advance to obtain appropriate authorization and guidance. The terms add that respecting intellectual property is a civic duty, and that any member who violates the clause shall be liable to Commonwealth Magazine Group for damages, including but not limited to litigation costs and attorneys’ fees.
2. Establishing internal AI-use policies
Some media organizations have also adopted internal AI-use policies. On July 17, 2023, the head of content at Commonwealth Magazine Group set out AI-use guidelines requiring all published content to undergo human review, editing, and fact-checking, while prohibiting the direct use of AI-generated articles, images, and videos. The organization has also emphasized transparency by disclosing AI-related applications used by its editorial operation and continuing to work with credible organizations so that AI serves as a tool for media professionals rather than a source of fake news or misinformation.
3. Self-regulation at the creation stage
Major Taiwanese literary competitions, including the Lin Rong-san Literary Award, Taipei Literary Award, and the Taiwan Mystery Writers’ Association writing competition, strictly prohibit AI ghostwriting. Sanctions may include disqualification, recovery of prize money and trophies, public disclosure of the violation, and lifetime bans from participating in the competitions.
Question 4. What should the publishing industry be preparing for now as it enters the AI era?
1. Establish internal AI-use guidelines
A publisher faces two risks if it opposes unauthorized AI use of its own content while simultaneously allowing editors to paste manuscripts into public generative AI tools. First, the manuscript may be exposed or leaked. Second, AI-generated material may be incorporated into a publication, creating uncertainty over whether the publication qualifies for copyright protection and potentially resulting in infringement of third-party copyrights. The publisher may bear responsibility for these risks.
Strict internal guidelines—such as requiring human review of all content and prohibiting the direct use of AI-generated material—can materially reduce these risks. Publishers should also address whether authors used AI in their submission agreements, potentially requiring authors to sign a “declaration of non-use of AI.” Strong internal controls will provide a firmer basis for asserting rights externally in the future.
2. Even where the law is silent, AI-training provisions must be included in notices and contracts
Taiwan’s Copyright Act does not currently expressly determine whether AI training constitutes infringement. As a result, copyright notices and contracts are the only tools publishers can control themselves.
Externally, publishers should expressly prohibit AI training in membership terms, copyright notices, and copyright pages. A clause like the one in Commonwealth Magazine Group’s membership terms, prohibiting users from providing content to generative AI for training, fine-tuning, generation, or derivative purposes, can serve as evidence that the copyright owner did not implicitly consent to AI training. It may also preserve a basis for arguing that a future infringement was willful and for negotiating a license.
Internally, new publishing agreements should expressly address whether AI training, fine-tuning, retrieval, and related uses fall within the licensed rights, how revenue will be allocated, and whether authors retain an opt-out right. Under Article 37 of Taiwan’s Copyright Act, where the scope of a license is unclear, the principle is that the use is presumed not to have been authorized. In other words, if the contract does not address an AI-related use, the publisher may not possess that right. Failure to address the issue proactively could later call into question the publisher’s authority to negotiate AI-training licenses on behalf of authors.
3. The Readmoo incident: prevent data theft and prepare for future infringement litigation
On August 8, 2026, Taiwan’s leading local e-book platform Readmoo (讀墨) suffered a large-scale, high-intensity malicious-traffic and hacker attack. Although the platform’s core asset—the digital files of e-books—was not breached or stolen, the incident nevertheless disrupted the service for several days.
In the AI era, publishers must strengthen access controls, encryption, and backup systems for copyrighted files. No matter how comprehensive their contracts and notices may be, once large quantities of content are stolen and incorporated into someone else’s training dataset, it becomes extremely difficult to recover the material afterward.
Prevention, however, is only half the task. If a publisher later seeks to establish infringement, it must be able to produce evidence showing who stole what content and when. Publishers should therefore establish procedures at an early stage for identifying, documenting, and preserving evidence of data theft. Without preserved evidence, even stringent legal rules and carefully drafted notices and contracts may provide little practical leverage in asserting rights.
4. Establish a collective copyright licensing system
A “collective licensing system” would pool works from multiple rights holders and allow a representative organization to negotiate with AI training organizations on a unified basis, collect licensing fees, and distribute the proceeds to the individual rights holders. Such a system would enable smaller rights holders to receive compensation as well.
Taiwan currently has no collective licensing gateway specifically for AI training, and licensing depends largely on individual negotiations. The National Science and Technology Council’s TAIDE project is a representative publicly disclosed example of paid licensing in Taiwan. Through the National Applied Research Laboratories’ Center for Science and Technology Policy, the project negotiated individually with content organizations and ultimately reached agreements with the Central News Agency (30 years of news, 1993–2023), Business Today (articles from 2008–2023), ETtoday, and Taiwan Panorama (光華雜誌), among others. Each organization received several million New Taiwan dollars in licensing fees.
The example illustrates the limitation created by the absence of a collective licensing system: only large organizations with substantial volumes of content and concentrated rights are well positioned to negotiate AI-training transactions, while small publishers and individual authors may be excluded from the market. The lack of a collective licensing mechanism can therefore also make it harder for AI developers to secure sufficient high-quality training data.
About the Contributors
Chun-Wei Tseng Copyright attorney in Taiwan.