Taiwan - AI Evaluation Centre

Overview

The AI Product and System Evaluation Centre (AIEC) was established on December 6, 2023, by Taiwan's Ministry of Digital Affairs (MODA) in cooperation with the National Institute of Cyber Security and the Industrial Technology Research Institute. The centre provides certification mechanisms, evaluation guidelines, and testing services for AI products and systems deployed in Taiwan. Operating under the 'AI Product and System Evaluation System' framework, the AIEC scrutinizes ten critical aspects of AI systems including safety, fairness, privacy, transparency, and accountability. The initiative aligns with European and American AI standards and has secured support from global technology companies including Meta and Microsoft, positioning Taiwan as a leader in trustworthy AI governance in the Asia-Pacific region.

Definitions

The framework establishes key terminology for AI governance in Taiwan. Artificial Intelligence System refers to any machine-based system that makes predictions, recommendations, or decisions influencing real or virtual environments. AI Product encompasses commercial applications incorporating AI capabilities offered to consumers or enterprises. Trustworthy AI describes systems meeting standards for safety, fairness, privacy, transparency, and accountability. Evaluation means the systematic assessment of AI products against established criteria and guidelines. Certification refers to formal recognition that an AI product meets AIEC standards. High-Risk AI includes systems deployed in sensitive contexts requiring enhanced scrutiny, such as healthcare, finance, and public safety applications.

Governance and Institutional Framework

The Ministry of Digital Affairs (MODA) serves as the primary authority overseeing the AIEC, with the Administration for Digital Industries (ADI) responsible for establishing evaluation systems and guidelines. The centre operates through collaboration with the National Institute of Cyber Security for security assessments and the Industrial Technology Research Institute (ITRI) for technical evaluation capabilities. Cross-departmental cooperation extends to the National Science and Technology Council (NSTC), which coordinates broader AI policy including the draft AI Basic Act. The Taiwan AI Center of Excellence (AICoE), officially opened on March 28, 2024, serves as a platform to streamline efforts by government agencies formulating AI governance rules. This multi-institutional approach ensures comprehensive oversight while leveraging specialized expertise across safety, security, and technical domains.

Key Focus Areas

• Safety Assessment: Evaluating AI systems for potential harms to users, third parties, and society, ensuring systems operate within acceptable risk parameters.
• Fairness and Non-Discrimination: Testing for algorithmic bias across protected characteristics including gender, ethnicity, age, and disability status.
• Privacy Protection: Assessing compliance with data protection requirements including the Personal Data Protection Act and emerging privacy standards.
• Transparency Standards: Evaluating the explainability and interpretability of AI decision-making processes for users and regulators.
• Accountability Mechanisms: Ensuring clear responsibility chains and documentation for AI system decisions and outcomes.
• Robustness Testing: Assessing system performance under adversarial conditions, edge cases, and unexpected inputs.
• Security Evaluation: Testing cybersecurity posture including resistance to data poisoning, model extraction, and prompt injection attacks.
• Reliability Verification: Assessing consistent performance across operating conditions and over time.
• Linguistic Inclusivity: Ensuring AI systems work effectively with Taiwan's national languages and diverse linguistic communities.
• Human Oversight Capability: Verifying that systems maintain appropriate human control and intervention mechanisms.
• International Standard Alignment: Ensuring compatibility with EU AI Act requirements and US AI governance frameworks.
• Generative AI Assessment: Specialized evaluation protocols for large language models and content generation systems.

Implementation Framework

Implementation proceeds through a structured evaluation process beginning with voluntary submission by AI developers and deployers. The AIEC conducts initial assessment to determine risk classification and applicable evaluation criteria. Technical testing follows established protocols for each of the ten evaluation dimensions, with specialized methodologies for different AI application categories including generative AI, computer vision, and autonomous systems. Successful products receive certification enabling market deployment with recognized trust credentials. The centre provides free GPU computing resources through MODA to help AI startups train models, reducing development costs and accelerating responsible AI adoption. A phased approach prioritizes government agency AI systems for initial evaluation, with expansion to private sector applications following. The NT$10 billion AI ecosystem investment secured from the National Development Council supports infrastructure development and talent programs enhancing evaluation capabilities.

Monitoring and Evaluation

The AIEC maintains ongoing oversight through periodic reassessment requirements for certified AI products. Continuous monitoring tracks deployed systems for emerging risks, performance degradation, or compliance drift. The centre analyzes industry needs and trends to update evaluation criteria, ensuring relevance as AI technology evolves. Post-market surveillance mechanisms enable identification of issues not apparent during initial evaluation. Stakeholder feedback channels allow users, civil society, and industry to report concerns about certified products. The AIEC coordinates with international evaluation bodies to share best practices and harmonize methodologies. Annual reporting provides transparency on evaluation activities, certification decisions, and emerging risk patterns. Technical audits verify that certified products maintain compliance with conditions of certification over their operational lifecycle.

Penalties, Liability, and Appeals

The current framework operates primarily through voluntary certification and market incentives rather than mandatory compliance penalties. Products failing evaluation may not receive certification, limiting access to government procurement and trusted AI designations. The draft AI Basic Act, currently under Executive Yuan review, would establish more comprehensive penalty provisions for high-risk AI violations. Product liability for AI-related harms follows existing consumer protection and product safety frameworks pending dedicated AI liability legislation. Developers whose products cause harm may face civil liability under tort principles and existing regulatory frameworks. Appeals against evaluation decisions proceed through administrative channels within MODA. The AIEC provides detailed feedback on evaluation failures to enable remediation and resubmission. Certification may be suspended or revoked if post-market monitoring reveals non-compliance or emerging safety concerns.

Relationship to Other Instruments

The AIEC framework operates in coordination with Taiwan's broader AI governance ecosystem. The draft AI Basic Act, published by NSTC in July 2024, would establish overarching principles and requirements that the AIEC would help implement through technical evaluation. The Personal Data Protection Act (PDPA) provides the foundational privacy framework that AI products must respect, with the AIEC assessing PDPA compliance as part of privacy dimension evaluation. Cybersecurity requirements under the Cyber Security Management Act inform security assessment protocols. The framework aligns with industry-specific regulations in sectors like healthcare, finance, and telecommunications that have their own AI deployment requirements. The Taiwan AI Action Plan 2.0 provides strategic context for AIEC activities within national AI development objectives. International instruments including the OECD AI Principles and EU AI Act inform evaluation criteria and support interoperability with global AI governance standards.

International Alignment

The AIEC explicitly aligns with European and American AI standards to ensure Taiwan's AI products can compete in global markets. Evaluation criteria incorporate requirements from the EU AI Act's risk-based classification system, enabling Taiwanese products to demonstrate compliance with the world's most comprehensive AI regulation. Alignment with US frameworks including the NIST AI Risk Management Framework and White House AI Executive Order principles ensures compatibility with major technology markets. Partnerships with global technology companies including Meta and Microsoft provide input on international best practices and emerging standards. The centre participates in international forums on AI governance, contributing Taiwanese perspectives while learning from global experiences. Mutual recognition arrangements are being explored with evaluation bodies in other jurisdictions to reduce duplicative certification burdens. This international alignment strategy positions Taiwan as a trusted origin for AI products while supporting domestic companies' export capabilities.

Implementation Timeline

Sources and References