Generative AI Under Taiwan Law: Copyright, Fair Use and the AI Fundamental Act
Observing current Copyright Act in Taiwan, Article 3, Paragraph 2 stipulates: “Author: refers to the person who creates the work.” Based solely on textual interpretation, Taiwanese regulations appear to recognize only “persons” as eligible authors. Taiwanese courts have supplemented this through judicial rulings: “The latter part of Article 1 of the Copyright Act stipulates that matters not provided for in this Act shall be governed by other laws. According to the relevant provisions of the General Principles (Chapter I) and Persons (Chapter II) of the Civil Code, ‘persons’ include natural persons (Section I) and legal persons (Section II). Furthermore, Article 3(1)(2) of the Copyright Act contains no provision excluding legal persons from being authors. Therefore, both natural persons and legal persons may be authors.” Furthermore, the majority of rulings consistently reiterate that only “creations resulting from human intellectual activity” qualify for copyright protection. Based on the above, according to the plain meaning of Taiwan's current legislation and past judicial interpretations, it appears reasonable to infer that Taiwanese law protects only the crystallization of human intellectual creations. Independent creations by artificial intelligence do not meet this requirement and thus cannot be protected under Taiwan's Copyright Act. However, the aforementioned interpretations remain centered on the question of “who acquires copyright.”
In recent years, discussions regarding artificial intelligence copyright have emerged in Taiwanese practice. The Taiwanese copyright authority, the Intellectual Property Office under the Ministry of Economic Affairs, has addressed the copyright controversy surrounding “independent creations by artificial intelligence.” Regarding public inquiries about “the training results of speech recognition models obtained by analyzing and processing voice data from the 1999 public hotline using robots,” regarding whether AI-generated outcomes qualify for copyright protection, the bureau stated: “Since AI is neither a natural person nor a legal entity, the intellectual outcomes it produces do not constitute works protected under the Copyright Act and generally cannot enjoy copyright protection. However, if the experimental results involve creative participation by a natural person or legal entity, and the robot's analysis is merely a mechanically executed process, then the copyright for the expression of such results belongs to that natural person or legal entity.” Regarding copyright uncertainties surrounding AI-driven “automatic music systems,” it further states: “Music created by an ‘automatic music system’ is merely the result generated by the machine or system through automated computation, lacking human ‘originality’ and ‘creative input.’ Such works may not qualify as protected works under copyright law.” Regarding Taiwanese court rulings, concerning computer-generated analytical charts created by AI-enabled software, it has been held: "The generation of analytical charts relies on the user inputting relevant parameters, after which the computer software performs calculations based on these parameters and produces the analytical charts. Thus, the generation or variation of the aforementioned analytical charts constitutes the result of the computer software's calculations based on the input parameters. Since such results are derived from mathematical calculations, they are not creations by ‘humans’ and thus cannot be regarded as subjects protected under the Copyright Act.” In summary, it is evident that Taiwan's judicial practice maintains a reserved stance on the copyrightability of ”independent creations by artificial intelligence."
Taiwanese scholars contend that if a work is entirely produced by a computer, it cannot be protected under Taiwan's Copyright Act—for example, translations automatically generated by translation software. Some scholars argue that this interpretation would make it difficult to distinguish between human translations and AI-generated translations, thereby hindering efforts to clarify the scope of copyright protection.
In reviewing Taiwan's copyright protection for “AI-generated original works,” current legislation largely adopts a reserved stance. Most discussions revolve around whether a “creator” can be non-human, relying solely on exhaustive legal methods to interpret the definition of traditional copyright. In response, an alternative perspective has emerged: why not draw upon the framework of neighboring rights to regulate the copyright of AI-generated works? This approach could safeguard the economic interests of those investing in programming while sidestepping the controversy over AI's non-natural person status. It would seek legislative solutions outside the traditional copyright system to address the current legal challenges posed by AI-generated creations.
When a generative artificial intelligence model scrapes data from the Internet for training purposes and stores such data, the conduct falls within the statutory definition of “reproduction” under Article 3, Paragraph 1, Subparagraph 5 of the Taiwan Copyright Act, which defines reproduction as the direct or indirect, permanent or temporary duplication of a work by means of printing, reprography, sound recording, video recording, photography, transcription, or other methods.
If the scraped online data constitutes works protected under the Taiwan Copyright Act, Article 22, Paragraph 1 grants authors the exclusive right of reproduction, except as otherwise provided in the Act. According to an official interpretive letter issued by the Intellectual Property Office of the Ministry of Economic Affairs (Letter No. 11252800520, June 16, 2023), reproduction of a protected work requires the authorization or consent of the copyright owner unless the use falls within the scope of fair use under Articles 44 to 65.
The scraping of Internet data by generative AI models for training purposes does not fall within the scope of temporary reproduction that is transient, incidental, an essential part of a technological process, and without independent economic significance, as provided under Article 22, Paragraphs 3 and 4 of the Taiwan Copyright Act. In other words, such activity cannot be considered a lawful temporary reproduction made solely for the purpose of lawful Internet relay transmission or other permitted uses of a work. Separately, it is generally difficult to invoke the specific fair use provisions enumerated in Articles 44 to 63 of the Copyright Act in this context.
In assessing whether such conduct falls within the scope of the general fair use provision under Article 65, Paragraph 2 of the Taiwan Copyright Act, and in the absence of judicial precedents specifically addressing the reproduction of works for AI training purposes, the analysis must rely on the balancing factors set forth in the same paragraph.
Specifically, in determining whether the exploitation of a work is within a reasonable scope as envisaged in Articles 44 to 63 or otherwise constitutes fair use, all relevant circumstances should be considered. Particular attention should be paid to the following factors:
1. The purpose and nature of the exploitation, including whether it is commercial in nature or undertaken for non-profit educational purposes;
2. The nature of the work itself;
3. The amount and substantiality of the portion used in relation to the work as a whole;
4. The effect of the exploitation on the work’s current market value and potential future market.
Article 65 of the Taiwan Copyright Act was modeled after Section 107 of the U.S. Copyright Act. In earlier U.S. jurisprudence, unauthorized commercial use was presumed not to constitute fair use. However, the U.S. Supreme Court gradually shifted its position, holding that the commercial nature of the use is not determinative. If a use is “transformative,” adding new value, expression, or meaning to the original work, and the degree of transformation is higher, the importance of other fair use factors diminishes, making fair use more likely.
The concept of transformative use has also influenced Taiwanese judicial practice. Courts in Taiwan, when applying the general fair use clause under Article 65, Paragraph 2, have introduced the concept of transformative use and evaluated the degree of transformation. Examples include Taiwan Taichung District Court Civil Judgment No. 5 of 2023 (臺灣臺中地方法院112年度智字第5號民事判決, Intellectual Property), Intellectual Property Court Civil Judgment No. 5 of 2019 (智慧財產法院108年度民商上字第5號民事判決), Intellectual Property Court Civil Judgment No. 1 of 2018 (智慧財產法院107年度民商訴字第1號民事判決), Intellectual Property Court Criminal Appeal Judgment No. 41 of 2008 (智慧財產法院97年度刑智上訴字第41號刑事判決), and Taiwan Taichung District Court Criminal Judgment No. 1865 of 2005 (臺灣臺中地方法院94年度訴字第1865號刑事判決).
In Taiwan Taichung District Court Criminal Judgment No. 1865 of 2005 (臺灣臺中地方法院94年度訴字第1865號刑事判決), the court held that the defendant’s use of character descriptions from a Japanese game manual in a strategy guidebook constituted fair use. Although the use was commercial in nature, both the quantitative and qualitative portions of the work used were minimal, and it did not affect the current or potential market for the original work. The court characterized this as a “transformative use,” consistent with the U.S. Supreme Court’s formulation in Campbell v. Acuff-Rose Music, Inc., describing it as a “productive use,” and therefore concluded that no infringement occurred. By contrast, in other copyright infringement cases where Taiwanese courts applied transformative use analysis, the defendants’ use was generally found insufficient to meet the threshold for transformative use.
When a generative AI model, for commercial purposes, scrapes copyrighted works for training and reproduces them in a database solely as materials for statistical analysis and parameter training—without commenting on, criticizing, or creating new expression based on the works themselves, and without engaging in any creative transformation of their expressive content—such conduct differs from the transformative use recognized in judicial practice. In the absence of clear legislative provisions or judicial precedents, where the activity is commercial, involves large-scale reproduction of works for model training, and may impact the licensing market for those works, the likelihood of qualifying as fair use under Article 65, Paragraph 2 appears comparatively low.
With respect to whether AI-generated outputs infringe copyrighted works, it remains necessary to evaluate the traditional criteria of “access” and “substantial similarity.” In the context of AI, “access” typically refers to whether the work was included in the model’s training data, while “substantial similarity” refers to a high degree of resemblance between the generated output and the original work, or the replication of a qualitatively significant portion of the work. If the model had access to the work during training and the generated output is substantially similar to the original, copyright infringement may be found. When assessing alleged infringement of artistic works—including graphic, photographic, fine art, or audiovisual works—courts should pay particular attention to the “total concept and feel” of the works, as perceived by an ordinary reasonable observer.
When a generative AI model is employed merely as an auxiliary tool in the creative process based on an existing work, and there is meaningful human creative contribution, the resulting work may still qualify for copyright protection as an adaptation of the original work. Conversely, if the creative process—though grounded in an existing work—is carried out entirely through the AI’s computational functions without any human intellectual input, the resulting output does not enjoy copyright protection, as Taiwan’s Copyright Act safeguards only works that embody human intellectual creation.
Whether the use of AI-generated output that involves the reproduction or adaptation of another person’s work qualifies under Article 65, Paragraph 1 of the Taiwan Copyright Act—“Fair use of a work shall not constitute infringement of economic rights in the work”—depends on whether such use meets the specific fair use provisions enumerated in Articles 44 to 63, or falls within the scope of the general fair use clause under Article 65, Paragraph 2.
Where the use is undertaken for legislative, administrative, judicial, or educational purposes, and remains within the reasonable scope prescribed by the relevant statutory provisions, reproduction of the original work may qualify as fair use. Similarly, for purposes such as the compilation of textbooks, classroom teaching, or personal or family non-profit use, adaptation of the original work may also be justified under the fair use framework.
By contrast, where the generated output is used for commercial purposes, the competent authority has recommended that users ascertain whether the developer or operator of the generative AI model has obtained authorization from the copyright owners of the original works and whether sublicensing for commercial exploitation is permitted, so as to mitigate the risk of potential copyright disputes.
In recent years, the rapid development of generative artificial intelligence (“generative AI”) has drawn increasing attention worldwide, particularly with respect to its impact on content generation, automated decision-making, and data utilization, as well as the legal and governance risks associated therewith. In response to these developments, Taiwan formally promulgated and implemented the Artificial Intelligence Fundamental Act on January 14, 2026. As Taiwan’s first comprehensive statute dedicated specifically to artificial intelligence, the Act establishes a foundational governance and institutional framework applicable to all AI technologies, including generative AI.
The legislative objectives of the Artificial Intelligence Fundamental Act are to build a smart nation, promote human-centered AI research and industrial development, safeguard human dignity and fundamental rights, and advance digital inclusion and sustainable social development (Article 1). From a legislative design perspective, rather than imposing immediate and highly detailed regulatory obligations specifically targeting generative AI, the Act adopts a principle-based and framework-oriented approach. Through high-level institutional design, it seeks to address the potential risks and uncertainties arising from generative AI while preserving regulatory flexibility.
With respect to government roles and governance mechanisms, the Act designates the National Science and Technology Council as the central competent authority, responsible for overall policy planning, inter-agency coordination, and the promotion of AI-related initiatives. Where AI applications involve specific industries or professional sectors, the relevant competent authorities are tasked with implementation in accordance with their respective mandates (Article 2). In addition, the Executive Yuan is to establish a National Artificial Intelligence Strategic Advisory Committee, comprising experts from academia, industry, and civil society, to coordinate, promote, and supervise national AI affairs and to formulate a National AI Development Strategy as a basis for medium- and long-term policy planning and resource allocation (Article 6). This governance structure leaves substantial room for policy adjustment in response to the cross-sectoral applications and regulatory demands that generative AI may give rise to in the future.
Regarding the risk governance for generative AI, the Act does not adopt a blanket prohibition or rigid regulatory approach. Instead, it articulates a governance direction centered on risk classification and impact assessment. The government is required to establish appropriate risk management and supervisory mechanisms for AI products and systems that are classified as high-risk or that may have significant social impacts, and to mandate appropriate information disclosure and warning measures. In particular, where AI applications involve minors or other vulnerable groups, their best interests must be given priority (Article 5). The Act further contemplates the development of more concrete regulatory arrangements concerning the allocation of legal liability, standards of attribution, and mechanisms for remedies, compensation, or insurance in connection with high-risk AI applications, with a view to enhancing the accountability and societal trustworthiness of generative AI systems (Article 17).
Moreover, given that generative AI relies heavily on large volumes of data for training and operation, the Act places particular emphasis on data governance and privacy protection (Article 4). On the one hand, the government is required to promote data openness, sharing, and reuse in order to enhance the quality and availability of data used in AI systems (Article 13). On the other hand, the Act requires that unnecessary collection, processing, or use of personal data be avoided in the development and deployment of AI technologies, and that data protection principles be incorporated into the design of AI systems from the outset (Article 14). These requirements serve as important guiding principles for the development of generative AI.
Overall, the Artificial Intelligence Fundamental Act is not a statute that immediately imposes stringent regulatory restrictions on generative AI. Rather, it constitutes a foundational, directional, and flexible governance framework. Its practical impact will gradually emerge through subsequent subsidiary legislation, administrative guidelines, and implementation practices. It therefore remains worth continued observation how Taiwan will shape a trustworthy AI governance model that balances the promotion of generative AI innovation with the protection of fundamental rights.