Artificial Intelligence and Intellectual Property in Taiwan: Innovation Ambition Meets the Copyright Bottleneck

Introduction

The year 2026 marks a turning point for the relationship between artificial intelligence (AI) and intellectual property (IP) in Taiwan. On 14 January 2026, the Artificial Intelligence Basic Act (人工智慧基本法, the “AI Basic Act” or AIA) was promulgated and entered into force, establishing Taiwan’s first framework statute dedicated to AI governance. The AIA is an innovation-first, principle-based law, and it deliberately declines to resolve the hardest IP questions, instead committing the government to revisit the relevant statutes within two years.

This restraint is significant because Taiwan is simultaneously pursuing an ambitious sovereign AI agenda: the development of a Traditional Chinese large language model (LLM) and a “Taiwan Sovereign AI Training Corpus” that reflects Taiwanese language, culture and values. Yet the same body of Traditional Chinese material that Taiwan most wants to train on is, for the most part, protected by copyright. As a result, copyright, not patent, trade mark or trade secret, has become the single most consequential IP constraint on Taiwan’s AI development. Against the backdrop of the new AIA, this article traces the copyright bottleneck while also examining developments reshaping patents, trade marks, and trade secrets.

The AI Basic Act and its intellectual property touchpoints

The AIA is a short, twenty-article framework law. It designates the National Science and Technology Council (NSTC) as the central competent authority and tasks the Ministry of Digital Affairs (MODA) with building a risk-classification framework. The AIA codifies seven core principles: sustainable development and well-being; human autonomy; privacy protection and data governance; cybersecurity and safety; transparency and explainability; fairness and non-discrimination; and accountability. On 7 July 2026, MODA published Version 1.0 of the AI Risk Classification Framework (the “Framework”), providing sector regulators with a common methodology for identifying, assessing and responding to AI risks while emphasising a proportionate, innovation-supportive regulatory approach. Although the Framework is not specific to intellectual property, it is highly relevant to IP practitioners. Notably, it expressly identifies intellectual property infringement, including the unauthorised use of copyrighted training materials and the generation of infringing AI outputs, as a recognised category of AI risk. This reflects the government’s view that IP protection is an integral component of trustworthy AI governance rather than merely a matter of private rights enforcement.

For IP practitioners, the most important provision is Article 13 on data governance. It directs the government to establish mechanisms for data openness, sharing and reuse to improve the availability and quality of AI training data, and, critically, to ensure that training and output results sufficiently reflect the nation’s diverse cultural values while protecting intellectual property rights. Two further features shape the IP landscape:

• A Pro-Innovation Priority Rule: Article 11, Paragraph 1 expressly provides that, where measures promoting AI research and development, application or governance conflict with existing laws or regulations, priority should be given to facilitating the development of AI technologies and services, provided that the seven core principles of the AIA are upheld. This indicates that future copyright reform may favour enabling AI training while maintaining appropriate safeguards for rights-holders.
• A Two-Year Review Clock: Within two years of the effective date, the government must review, enact or amend the relevant laws and administrative measures to conform to the AIA. The Copyright Act sits squarely within the scope of that mandate.

The AIA does not itself create any copyright exception for text and data mining, nor does it answer who owns AI-generated output. It is an enabling framework that defers the substantive IP questions to sector regulators and to a future amendment cycle. The practical consequence is that, for at least the next two years, the existing IP statutes and the courts will continue to govern AI disputes.

Copyright: the central battleground

Copyright is where AI and IP collide most directly in Taiwan, and the area of greatest legal uncertainty. The Taiwan Intellectual Property Office (TIPO) has issued only one interpretive letter of real significance, dated 16 June 2023. In it, TIPO took the position that feeding copyright-protected works into an AI model for training constitutes reproduction that infringes the right-holder’s reproduction right, unless it qualifies as fair use under Articles 44 to 65 of the Copyright Act. The letter also maintained that AI-generated output, lacking a human author, is not itself a protectable “work”.

The difficulty is that Taiwan’s Copyright Act contains no statutory exception for text and data mining or for AI training. Whether training qualifies as fair use must therefore be decided case-by-case under the open-ended factors of Article 65, with no controlling precedent yet on point. This leaves both developers and right-holders in a grey zone.

As of mid-2026, there is still no decided AI-training copyright case in Taiwan. But two recent disputes have cast a long shadow over the sector and illustrate the criminal dimension that distinguishes Taiwan from many other jurisdictions:

• The Lawsnote case, decided only at first instance and so still subject to appeal, did not itself concern AI training but shows how severely Taiwanese courts treat large-scale unauthorised copying. In June 2025, the New Taipei District Court convicted the two founders of legal-tech start-up Lawsnote of both copyright infringement and computer-use offences under the Criminal Code (treating the unauthorised crawling as illicit access to another party’s database) after the company used a crawler to scrape nearly 100,000 legislative amendment-history records from the commercial Lawsbank database and load them directly into its own competing search product. A threshold question, which the defendants dispute on appeal, is whether such records are protected by copyright at all: the text of statutes is itself public domain, and the conviction rested on treating Lawsbank’s compilation of those amendment histories as an original compilation work. The founders received prison sentences of four and two years respectively, with civil compensation of roughly TWD105 million, an unprecedented level of combined criminal and civil liability for data scraping in Taiwan.
• The Central News Agency (CNA) dispute bears far more directly on AI training. In July 2025, an open-source Traditional Chinese dataset circulating in a generative-AI community was found to contain more than 140,000 news articles belonging to CNA, Taiwan’s national news agency, none of which CNA had authorised for use in the dataset; CNA filed a criminal complaint and prosecutors questioned the person who had compiled and released the dataset. Notably, the dataset had been distributed for free with no commercial gain, yet the unauthorised reproduction of news content to build a training corpus still drew criminal investigation. The case is the clearest signal to date that assembling, or even sharing, a Traditional Chinese training set drawn from copyright-protected news material can carry criminal exposure in Taiwan, whether or not the use is commercial.

Because copyright infringement in Taiwan can carry criminal penalties, not merely civil damages, using Traditional Chinese materials to train an LLM may expose developers to criminal liability. This is a materially harsher risk profile than in the United States, where the question is litigated as a matter of civil fair use. It explains why copyright has become the principal brake on indigenous LLM development in Taiwan. Until either the courts recognise training as fair use or the legislature enacts an exception, the criminal-liability overhang will persist.

The policy stakes are now openly acknowledged at the highest levels of oversight. In 2025, the Control Yuan concluded an investigation into a domestic LLM controversy involving a Traditional Chinese corpus and found clear deficiencies in the government’s data governance, singling out the failure to adapt the copyright framework. The Control Yuan observed that a lone TIPO interpretive letter is an inadequate legal basis for resolving fast-moving AI copyright questions, and urged the Executive Yuan and TIPO to pursue active legislative adjustment. The AIA’s two-year review clock is the most likely vehicle for that reform.

The sovereign AI training corpus problem

Taiwan’s sovereign AI ambition is driven by both cultural and security concerns. Models trained predominantly on Simplified Chinese data risk importing the political framing and value system of the People’s Republic of China; Taiwan’s National Security Bureau has publicly flagged content bias in several China-developed AI tools. A Traditional Chinese model trained on Taiwanese material is therefore framed as a matter of digital sovereignty, not merely commercial competitiveness.

To address the copyright bottleneck, MODA launched the “Taiwan Sovereign AI Training Corpus” in late December 2025, built initially from open government data, policy reports and government publications contributed by more than 200 agencies, and published the “Taiwan Sovereign AI Training Corpus License - Version 1”, a standardised set of terms under which contributing agencies release their data with a clear grant of rights and AI developers apply to use the corpus for training. By replacing case-by-case copyright clearance with a single uniform licence, the mechanism is intended to contain the infringement exposure that would otherwise attach to training on this material. This approach mirrors the earlier Taiwanese Native Large Language Model (Trustworthy AI Dialog Engine; TAIDE) project, whose training data was deliberately sourced from government agencies to sidestep copyright disputes. The strategy is sound but inherently limited: a corpus confined to government text cannot capture the linguistic and cultural richness held in privately owned news archives, literature and media, precisely the material that gives a model its Taiwanese voice, and precisely the material that remains locked behind copyright and criminal-liability risk.

Patents: human inventorship and AI-assisted inventions

Taiwan’s patent regime has taken a clear position on AI inventorship. In its 2021 decision concerning the DABUS application (an AI system named as the sole inventor), the Intellectual Property and Commercial Court (IPC Court) upheld TIPO’s rejection, holding that an inventor must be a natural person who has made a substantial mental contribution to the claimed invention. AI, being neither a natural nor a legal person, is treated as an “object” under Taiwan law and cannot be an inventor. This aligns Taiwan with the United States, the United Kingdom, the European Patent Office, Japan and China.

The more practical questions concern AI-assisted inventions and the patentability of AI technology itself. Taiwan does not yet have a dedicated set of AI patent examination guidelines. However, in October 2025, TIPO published AI-related invention case examples illustrating how existing patentability standards apply to AI technologies. The amended Examination Guidelines for Computer Software-Related Inventions, effective 1 July 2021, therefore remain the principal framework governing AI-related paten examination. Under those guidelines, the inventive step of software and AI inventions is assessed using the general standard, weighing factors that support and factors that deny inventive step, with “simple variations”, such as merely systematising tasks previously performed by humans, weighing against patentability. Applicants seeking to protect AI inventions in Taiwan should focus claims on concrete technical effects rather than abstract algorithms, and should ensure the human inventive contribution is clearly identifiable, as AI tools become routine in the R&D process.

Trade marks: AI in branding and enforcement

Trade mark law has so far been the least disrupted of the four IP fields, but AI is reshaping practice at the margins. Generative tools are increasingly used to create brand names, logos and marketing assets, which raises familiar questions about whether AI-generated logos attract protectable rights and whether they inadvertently reproduce existing marks. Under prevailing principles, trade mark rights flow from use and registration rather than authorship, so AI involvement in creating a mark does not in itself defeat registrability, but applicants should conduct careful clearance searches, because a generative model may output a sign confusingly similar to a registered mark.

On the enforcement side, AI is becoming a tool for both infringers and brand owners. AI-driven counterfeiting and the automated generation of look-alike branding increase the volume of potential infringement, while brand owners deploy AI-based monitoring to detect counterfeits and unauthorised use across online marketplaces. TIPO has itself begun to deploy AI: in March 2024 it launched an AI-powered “search by image” trade mark retrieval tool, which lets a user upload a mark and obtain visually similar prior applications, registrations and refused marks across all 45 classes, without requiring the user to input separate image-path parameters, making pre-filing clearance searches faster and more accessible. Released initially as a trial tool to supplement, rather than replace, the existing graphic-similarity search, it placed TIPO alongside offices such as the EUIPO, IP Australia and WIPO in offering AI image-based mark searching, even though no AI-specific trade mark rules have yet been adopted.

Trade secrets: from commercial disputes to national security

Trade secret protection has undergone the most dramatic escalation of any IP field in Taiwan, driven by the convergence of AI, semiconductors and national security. As AI accelerates the value of proprietary process technology and training data, Taiwan has reframed its most advanced know-how as a strategic national asset rather than ordinary commercial IP.

The key instrument is the 2022 amendment to the National Security Act, which created a category of “national core key technologies” that covers advanced semiconductor processes and similar strategic know-how. Misappropriation of such technologies for use outside Taiwan now carries criminal exposure far exceeding ordinary trade secret claims, and corporate entities, not only individuals, can be punished. Economic espionage benefiting hostile foreign forces attracts the most severe penalties.

On 27 April 2026, the IPC Court delivered the first conviction under this regime. In a case involving the alleged theft of TSMC’s 2-nanometre and sub-14-nanometre process secrets, the lead defendant, a former TSMC employee who had joined the Taiwan subsidiary of a Japanese equipment supplier, received a ten-year sentence, the maximum for the lead actor, while co-defendants received terms ranging from two to six years and the corporate employer was fined for failing to supervise its staffs. The ruling establishes a published sentencing benchmark and confirms that Taiwan will treat leading-edge process technology as a matter of national survival.

For businesses, the lessons are concrete. Courts scrutinise whether the holder took “reasonable confidentiality measures”, so robust internal controls, access logging, and enforceable non-disclosure and non-compete obligations are now essential. Companies handling advanced semiconductor process and chip-design know-how should map which assets may qualify as national core key technologies (the designated list squarely targets sub-14-nanometre IC manufacturing and, in its later expansion, AI-computing chip design) and calibrate compliance, separation and “bring-your-own-knowledge” policies accordingly.

Outlook

The defining tension in Taiwan’s AI-IP landscape is between an innovation-first policy posture and a copyright regime that, uniquely, backs infringement with criminal sanctions. The AIA sets the direction of travel and starts a two-year clock, but it leaves the decisive copyright questions unanswered. How the legislature and TIPO resolve the training-data problem (whether through a statutory text-and-data mining exception, a collective or statutory licensing mechanism, or judicial recognition of training as fair use) will largely determine whether Taiwan can build a genuinely sovereign Traditional Chinese model or remain constrained to a government-data corpus.

In the meantime, businesses operating in Taiwan should treat AI and IP as an integrated risk. Developers must assume that training on copyright-protected Traditional Chinese material carries criminal as well as civil exposure until the law changes. Patent applicants should secure clearly human inventive contributions and frame claims around technical effect. Brand owners should fold AI clearance and AI-enabled monitoring into trade mark strategy. And any company touching strategic technology must build its trade secret programme to withstand scrutiny under both the Trade Secrets Act and the National Security Act. The next two years will be decisive, both legally and technologically, for the future of AI in Taiwan.