I.Introduction

As artificial intelligence technologies advance rapidly and spread across industries, the focus of AI governance worldwide has shifted from declaring ethical principles to building concrete rules. On January 14, 2026, Taiwan promulgated the Artificial Intelligence Basic Act (the “AI Basic Act”), which took effect on the same day and serves as the legal foundation for AI governance in Taiwan. Pursuant to Article 16 of the Act, the Ministry of Digital Affairs (MODA) developed the AI Risk Taxonomy and Assessment Framework (the “Framework”). After approval by the Executive Yuan, MODA released Version 1.0 of the Framework on July 7, 2026.

Notably, AI in Taiwan concerns not only economic and industrial development but also national security, and this consideration is reflected in how the Framework’s risk types are designed. This article focuses on the Framework. It explains the structure of the AI Basic Act, Taiwan’s AI risk governance system, and the Framework’s positioning and operational workflow, and it briefly compares Taiwan’s approach with those of other major jurisdictions. The aim is to help the international community understand the direction of AI risk governance in Taiwan.

II.The Two Pillars of the AI Basic Act

The AI Basic Act aims to promote innovation while safeguarding human rights and human autonomy. Broadly, the Act rests on two pillars. The first is promoting development, which covers industrial innovation, data governance, and talent cultivation. The second is risk management, which covers the Framework, warnings for high-risk applications, and the attribution of liability.

In terms of its legal nature, the AI Basic Act can be understood as a national AI strategy established in the form of law. It applies to government agencies and does not directly regulate businesses. In other words, the specific obligations of businesses continue to come from regulations that the competent authorities for the relevant industries (the “competent authorities”) adopt within their respective areas of responsibility. Competent authorities are the central government agencies that, by law, oversee particular industries or affairs. For example, the Ministry of Health and Welfare oversees healthcare, and the Financial Supervisory Commission oversees financial services.

On the risk management side, the relevant provisions are as follows. Article 5, Paragraph 1 requires the government to prevent AI applications from infringing on people’s life, body, freedom, or property, or from disrupting social order, national security, or the ecological environment. Under Article 16, MODA promotes an AI Risk Taxonomy and Assessment Framework, and the competent authorities establish management regulations in accordance with that framework. Article 17 requires the government to clarify the attribution of liability and the conditions for such liability for high-risk AI applications, and to establish relief, compensation, or insurance mechanisms. Article 18 requires government agencies to complete the review and adaptation of laws and regulations by January 13, 2028.

III.A Three-Tier Risk Governance Structure

Taiwan’s AI risk governance system can be understood in three tiers.

The first tier consists of development principles. Article 4 of the AI Basic Act sets out seven 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. These principles are similar to the AI Principles of the Organisation for Economic Co-operation and Development (OECD) and form the foundation of the overall governance system.

The second tier is the Framework. Adopted under Article 16, Paragraph 1, the Framework groups AI risks into three types: technical design flaws in AI systems; issues in post-deployment operation and human-AI interaction; and societal-structural and environmental impacts.

The third tier consists of the governance work of individual ministries and agencies. Each of the three risk types is matched with a different governance tool. Technical design flaws are addressed through assessment and verification tools. Issues in post-deployment operation and human-AI interaction are addressed through regulatory measures. Societal-structural and environmental impacts are addressed through government policies. In short, technology is checked through assessment and verification, use is governed by regulation, and broad social impacts are guided by policy. This design matches governance tools to the nature of each risk and avoids handling every risk through regulation.

IV.Positioning and Design of the Framework

First, the Framework serves the competent authorities. It helps them establish risk-based regulations for AI applications in the industries under their purview. “Risk-based” means that the higher the risk, the stricter the rules, while lower-risk applications retain room for innovation.

Second, the Framework is positioned as a methodological tool. It guides the competent authorities and does not directly regulate businesses. In this respect, it is similar to the AI Risk Management Framework (AI RMF) of the U.S. National Institute of Standards and Technology (NIST). Both are process-oriented, and neither method is in itself a form of regulation. Their legal status, however, differs. The NIST AI RMF is adopted by businesses on a voluntary basis, whereas the Framework has a statutory basis and binds the procedures that government agencies follow in developing regulations.

Third, the Framework neither relies on a list of high-risk AI predefined in law nor sorts AI into fixed risk levels. Instead, it uses specific use cases as the unit of analysis, and the competent authorities determine whether a use case is high-risk and state their reasons. This approach reflects three considerations: the same technology can pose very different risks in different contexts; technology changes rapidly; and each authority knows the industries under its purview best. To prevent inconsistent standards across agencies, the Framework also provides three common tools: a unified Table of AI Risk Types, a unified set of protected legal interests for determining serious harm, and a Table of Illustrative Examples of Serious Harm.

In addition, the Framework is centered on promoting innovation. When planning measures, agencies should follow the principle of proportionality and assess the opportunity costs of not adopting AI, so that an overly cautious approach does not leave policy lagging behind social needs. The Framework applies to the civilian domain and excludes military uses.

V.The Four-Step Operational Workflow

The Framework’s operational workflow consists of four steps, described below in order.

Step 1: AI Use Case Inventory

Each agency takes inventory of existing and planned AI use cases in the industries under its purview and records their purposes, technologies, and stakeholders.

Step 2: Risk Identification

Each agency checks each use case against the Table of AI Risk Types, item by item, to identify the risks involved. The table sets out 20 subtypes under three categories:

Among these, C6 (Cognitive Warfare and Information Sovereignty) addresses the fact that AI greatly lowers the cost of large-scale information manipulation. Malicious actors can use generative AI, automated accounts, and deepfake content to systematically interfere with public discourse, electoral processes, and policymaking. Few jurisdictions include this item in an official risk taxonomy, and its inclusion reflects Taiwan’s geopolitical environment.

Step 3: Risk Assessment

The assessment is based on the objective characteristics of the use case itself and does not take into account the mitigating effects of existing measures. A use case that may cause serious harm to any of six protected legal interests, namely national security, the fundamental rights of the people, life safety, property protection, social order, or the ecological environment, is a high-risk AI application.

Step 4: Risk Response

Each agency reviews how well the laws and regulations under its purview cover the identified risks and confirms whether they are sufficient to address them. If they are not, the agency plans measures in support of development or management measures in accordance with the principle of proportionality. For high-risk applications, the AI Basic Act requires that warnings be labeled and that liability be clearly attributed. Where a risk cannot be effectively controlled, the competent authority should restrict or prohibit the application in accordance with the law.

Consider a hypothetical example from the healthcare sector. After the Ministry of Health and Welfare identifies AI-assisted diagnosis as a use case in its inventory, it identifies risks such as erroneous outputs and overreliance. Because a misdiagnosis could endanger lives, the use case is assessed as a high-risk AI application. The Ministry then reviews its healthcare regulations and requires that warnings be labeled and that liability be clearly attributed. If the existing regulations fall short, it adopts new rules.

VI.Implementation Status

To support implementation, MODA has built an online system for internal government use. Agencies use the system to report their inventory and assessment results, which allows MODA to track progress and provide timely assistance. MODA is also giving priority to working with ministries whose responsibilities directly affect the significant rights and interests of the public. It provides explanations and advice on applying the Framework, on risk concepts, and on international cases. Risk determinations and rulemaking, however, remain the responsibility of each agency within its own authority. The results of this work will serve as the basis for completing regulatory adaptation by January 13, 2028, and the Framework will be revised on a rolling basis in light of practical experience.

VII.Comparative Observations

Internationally, the European Union’s Artificial Intelligence Act (AI Act) defines high-risk AI systems in advance through Annexes I and III and imposes obligations on providers and deployers. However, the European Commission’s Digital Omnibus on AI, which entered into force in July 2026, postponed the application of high-risk obligations to December 2027 and August 2028. South Korea’s Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trustworthiness took effect on January 22, 2026. It uses the concept of “high-impact AI” and lists the regulated areas in the statute itself. Japan enacted the Act on the Promotion of Research, Development and Utilization of Artificial Intelligence-Related Technologies in 2025. That Act focuses on promoting AI research, development, and utilization and contains no penalties.

Compared with these systems, Taiwan’s approach has three distinctive features. First, high risk is determined at the level of specific use cases, and the competent authorities make case-by-case determinations and state their reasons. Second, risk governance follows the existing legal framework and is carried out through each agency’s review of its laws and regulations. Third, the Framework’s risk taxonomy draws on mainstream international norms, so multinational companies can continue to use their existing global compliance frameworks.

VIII.Conclusion

Overall, Taiwan is re-examining its laws and regulations across sectors on a risk basis while building on its existing regulatory structures, so that its legal system can respond flexibly to rapidly changing AI technologies. As agencies gain practical experience and international developments evolve, the Framework will continue to be reviewed and revised, with the aim of striking a balance between promoting innovation and protecting people’s rights and interests.

I.Web Crawling Governance Issues: An Overview

The development of generative artificial intelligence technology in Taiwan relies heavily on training with massive amounts of data, driving a surge in large-scale automated data collection by web crawlers. The traditional Internet ecosystem has long operated on a tacit quid pro quo between search engine indexing and website content operators, whereby websites permit crawling in exchange for traffic and advertising exposure.

However, in the era of generative AI, model training "internalizes" external web content into model parameters and generates derivative outputs with commercial substitutability. This severs the original traffic feedback mechanism, substantially crowding out the advertising and subscription revenues of content producers.

This structural shift from "traffic-referral reciprocity" to "substantive market competition" has exposed a long-standing disconnect between technical signals and legal declarations of intent. Consequently, Taiwan's web crawling governance is confronting a dual bottleneck that constrains both legal regulation and industrial innovation.

II.International Governance Approaches: A Comparison

Internationally, governance models for crawling AI training data have mainly developed along two strategic paths:

1.The U.S. ex post judicial balancing model: This model relies primarily on the fair use doctrine under Section 107 of the Copyright Act, wherein courts conduct a case-by-case evaluation of the transformative use and its potential market substitution effect. Recent judicial trends also show that private-law contracts such as website terms of service, alongside diverse legal norms, are increasingly leveraged to build a protective net around rights.

2.The EU ex ante statutory exception model: Articles 3 and 4 of the Directive on Copyright in the Digital Single Market (CDSM Directive) establish statutory exceptions for text and data mining (TDM). These provisions are combined with a rights reservation (opt-out) mechanism that allows rightsholders to expressly opt out through machine-readable means, shifting the governance baseline toward ex ante regulation and technical labeling.

III.Legal Framework and Practical Developments in Taiwan: The Current Landscape

Unlike the EU, Taiwan's Copyright Act has not yet introduced a statutory TDM exception as an institutional safe harbor. When AI developers collect and train on large-scale corpora, their only statutory defense under the Copyright Act is an ex post defense under the general fair use provision of Article 65, Paragraph 2. Pursuant to this provision, in determining whether the exploitation of a work constitutes fair use, all surrounding circumstances shall be taken into account, with particular attention paid to the following factors:

● the purpose and nature of the exploitation, including whether it is of a commercial nature or for nonprofit educational purposes;
● the nature of the work;
● the amount and substantiality of the portion exploited in relation to the work as a whole; and
● the effect of the exploitation on the potential market for and current value of the work.

Taiwan's courts have not yet accumulated sufficient judicial precedents concerning generative AI training to establish a settled legal doctrine. Case-by-case standards therefore remain highly uncertain, and compliance costs for developers rise accordingly.

Furthermore, Taiwan's judicial practice has extended the protection of rightsholders' property interests into the realm of criminal liability. In the "Lawbank v. Lawsnote case" (Taiwan New Taipei District Court, Criminal Judgment 111 Zhi-Su No. 8)（臺灣新北地方法院111年度智訴字第8號刑事判決）, the court differentiated its legal reasoning based on the nature of the crawled content:

● Legislative history of laws and regulations: The court held that it constitutes a compilation work under Article 7 of the Copyright Act and is protected by copyright.
● Texts of laws and regulations: Under Article 9, Paragraph 1, Subparagraph 1 of the Copyright Act, these materials are not subject to copyright. However, the defendant violated the website's terms of use and, without authorization, deployed a crawler to harvest this material for its database. The court held that this constituted the offense under Article 359 of the Criminal Code of unlawfully obtaining another’s electromagnetic records.

This judgment demonstrates that even where crawled materials are not protected by copyright, crawling may still incur criminal liability for violating a website's terms of use. Web crawling in Taiwan therefore faces significant criminal liability risk. (This case is a first-instance judgment.)

IV.Industry Challenges and Future Trends in Taiwan

The current challenges of web crawling governance in Taiwan essentially stem from the gap between the technical limits of expression and "legal declarations of intent":

1.Semantic limitations of technical protocols: Existing machine-readable protocols (such as robots.txt) are structurally simplistic, issuing only binary instructions to allow or deny access. They cannot convey nuanced layers of legal authorization, nor can they accurately distinguish at the level of machine language between two entirely different kinds of authorization intent: referral-based indexing by search engines and training-based internalization by AI models.

2.The gap between natural-language terms and technical signals: Terms of service drafted by website operators in natural language are difficult for automated crawlers to parse in real time. As a result, even developers who subjectively intend to comply can easily become vulnerable to breach-of-contract or infringement disputes in practice. Meanwhile, lacking mutual trust and effective communication mechanisms, content producers have increasingly turned to aggressive technical blocking.

3.Lack of commercial licensing mechanisms: Under the dual constraints of an absent clear licensing and pricing framework and the lack of a statutory TDM safe harbor, enormous legal risks are generating a chilling effect across the domestic AI research and development environment.

V.Conclusion

Taiwan's future governance framework should not rely solely on ex post case-by-case judicial discretion. It should instead establish buffer mechanisms that narrow the gap between technical signals and legal declarations of intent, thereby striking a balance between safeguarding the substantive revenues of content producers and promoting the innovation of Taiwan's domestic AI industry.

To make the application of the Cyber Security Management Act clearer, the Administration for Cyber Security under the Ministry of Digital Affairs (MODA) recently issued two important interpretive rules:

I.Interpretation on Requirements for Restricting the Use of Products Harmful to National Cyber Security at Directly Operated or Outsourced Facilities

Paragraph 1 of Article 11 and Paragraph 1 of Article 27 of the Cyber Security Management Act both include provisions stipulating that agencies shall not download, install, or use products that are harmful to national cyber security. The same shall apply to any broadcasting equipment or Internet access services provided to the public at facilities operated directly or outsourced by the government agency, when necessary to maintain cyber security.

This provision takes into account that information currently circulates extremely fast and its content directly affects public life and public safety; given that broadcasting equipment and Internet access services at facilities operated directly or outsourced by government agencies are closely linked to information flow, regulatory oversight is deemed necessary.

Pursuant to Order No. 11550002132 (數授資法字第11550002132號) issued by the Ministry of Digital Affairs on June 29, 2026:

1."Broadcasting equipment" refers to all types of electronic equipment capable of playing text, images, or audio.
2."Internet access services" refers to wireless or wired Internet access services provided by agencies, including the software and hardware equipment used, as well as the information and communication services accessed through such Internet access services.
3."When necessary to maintain cyber security" refers to instances where the aforementioned equipment and services possess network connectivity capabilities and pose a risk of endangering cyber security as defined in Subparagraph 3 of Article 3 of the Act.

II.Interpretation on Regulations Governing Regulated Specific Non-Government Agencies That Are of Importance to Cyber Security

Under the Cyber Security Management Act, regulated entities primarily consist of Government Agencies and Specific Non-Government Agencies. By definition, a "Specific Non-Government Agency" refers to a critical infrastructure provider, a government-owned enterprise, a designated foundation, or any enterprise, organization, or institution under control of the government.

Among these, an "enterprise, organization, or institution under control of the government" refers to enterprises, organizations, institutions publicly announced by the Ministry of Civil Service pursuant to Article 77, Paragraph 1, Subparagraph 2, Item 3 and 4 of the Public Service Pension, Severance and Compensation Act, that are of importance to cyber security, as designated by the central competent authority in charge of the relevant sector and approved by the competent authority.

Pursuant to Order No. 11500012432 (數授資法字第11500012432號) issued by the Ministry of Digital Affairs on March 26, 2026: The requirement of being "of importance to cyber security" means that if an entity's core business system cannot resume normal operation within an acceptable timeframe, it may immediately and significantly impact daily social operations, key industrial operations, or financial operations. Therefore, the missions of enterprises, organizations, or institutions selected by the central competent authority in charge of the relevant sector under this Subparagraph must possess one of the following characteristics:

1.Impact on daily social operations: If business operations come to a halt, it could immediately cause a major impact on the daily lives of the general public.
2.Impact on key industrial operations: Playing a key role in industrial supply chains, holding a leading position in the industry market, belonging to the defense industry, or possessing special critical technologies, etc.
3.Impact on financial operations: Domestic systemically important banks (D-SIBs) that possess influence in the financial market.