人工智慧（Artificial intelligence，AI）技術並非新興產物，早在1940年代即已開始發展，近年因為算力提升、數位資料大量增加及如ChatGPT等生成式AI工具之問世，使AI更貼近日常生活，企業亦積極導入AI以提升效率與服務品質，但其快速應用也帶來資訊安全、系統偏誤、隱私侵犯等風險與疑慮。AI之「黑盒子」特性使決策過程不透明，導致使用者難以理解決策邏輯，增加了不公平性與倫理風險，醫療、自駕車、信用評分等AI應用若演算法或資料出錯，將威脅人身安全、財產或基本權利，而深偽（Deepfake）技術更被用於詐騙，加劇治理挑戰。國際間為因應此趨勢，紛紛提出治理原則及法制架構，OECD於2019年提出「價值為基礎之AI原則」，涵蓋包容性成長、人權尊重、透明性、穩健性與問責，2024年更新版本更強調法治與民主價值；G20、UNESCO亦分別發佈AI倫理與發展原則，強調公平、隱私、非歧視及人類監督。OECD在2022年推出AI系統分類框架，協助各國依風險進行監管，WHO也針對健康產業發佈AI治理指引。各國陸續發展類似政策並漸趨具體化，歐盟於2024年通過《人工智慧法》，成為全球首部全面水平式AI法規，強調以人為本、可信賴AI發展，針對高風險系統設立透明性、人為監督及資料治理等嚴格規範；英國提出AI治理白皮書，採風險為基礎、脈絡特定之監管策略，強調兼顧創新發展與風險管理；日本2025年立法推動AI研究及應用透明化、防止濫用，強化國際競爭力；韓國2025年施行《AI基本法》，要求開發者與應用者進行衝擊評估、確保透明性與安全性；澳洲也考慮針對高風險AI進行強制規範，並採取風險為基礎之治理邏輯。歐盟之《人工智慧法》採「以風險為基礎」之監管模式，將風險分為不可接受、高風險、特定透明性與最低風險等四類，對於不可接受風險如潛意識操縱、侵害知情決定之應用予以禁止，高風險系統則需符合嚴格要求，包括風險管理、資料治理、技術文件維護與監管責任等，透明性風險如聊天機器人必須標示為AI，最低風險則無強制規範，僅採自願性準則。為促進負責任之創新，各國陸續推動法規沙盒，如歐盟在《人工智慧法》中要求會員國建立AI法規沙盒，以受控環境進行測試，協助符合規範並加快產品上市；英國也設有針對個人化AI、AI醫療器材、金融科技等不同領域之沙盒計畫，新加坡則推出生成式AI測評沙盒與金融科技沙盒，透過這些機制平衡創新發展與風險管控。綜言之，全球AI治理已由原本之原則倡議逐漸轉向實質立法與具體實踐，並普遍採風險為基礎之監管思維可以給予台灣人工智慧風險治理路徑借鑑。

Artificial intelligence (AI) is not a newly emerged technology; it began developing as early as the 1940s. In recent years, driven by advances in computing power, the explosive growth of digital data, and the emergence of generative AI tools like ChatGPT, AI has become increasingly integrated into daily life, with enterprises actively adopting AI to improve efficiency and service quality. However, its rapid deployment also brings concerns over information security, system bias, privacy infringement, and other risks. The“black box”nature of AI makes decision-making processes opaque, leading to unfairness and ethical risks. AI applications in healthcare, autonomous vehicles, and credit scoring could endanger personal safety, property, or fundamental rights if algorithms or data are flawed. Additionally, deepfake technology has been exploited for fraud and to manipulate public opinion, posing serious threats to social trust and democratic systems. These challenges highlight the urgent need for AI regulation that balances technological advancement with human rights protection. In 2019, the OECD released its“Principles on Artificial Intelligence,”covering inclusive growth, respect for human rights, transparency, robustness, and accountability; the 2024 update further emphasizes rule of law and democratic values. Similarly, G20 and UNESCO released AI ethics and development principles, highlighting fairness, privacy, non-discrimination, and human oversight. In 2022, OECD launched an AI system classification framework to help countries regulate based on risk, while WHO published AI governance guidelines for the healthcare sector. Countries have followed suit with similar policies and are moving toward concrete regulations. The EU passed the“Artificial Intelligence Act”in 2024, becoming the world’s first comprehensive horizontal AI legislation, emphasizing human-centric and trustworthy AI, and imposing strict requirements on high-risk systems regarding transparency, human oversight, and data governance. The UK issued an AI regulation white paper, adopting a risk-based, context-specific supervisory approach to balance innovation and risk management. Japan enacted legislation in 2025 to promote transparent AI research and applications, prevent misuse, and strengthen international competitiveness. South Korea’s“AI Basic Act,”effective in 2025, requires developers and users to conduct impact assessments and ensure transparency and safety. Australia is considering mandatory regulations for high-risk AI and also adopts a risk-based governance logic. The EU’s AI Act follows a“risk-based”regulatory model, categorizing AI risks into unacceptable, high, specific transparency, and minimal risk. Applications with unacceptable risks, such as those manipulating subconsciousness or impairing informed decision-making, are banned. High-risk systems must comply with strict requirements, including risk management, data governance, technical documentation maintenance, and clear regulatory responsibilities. Applications with transparency risks, such as chatbots, must be labeled as AI, while minimal risk applications have no mandatory regulation and rely on voluntary codes of conduct. To promote responsible innovation, various countries have introduced regulatory sandbox mechanisms. The EU’s AI Act mandates member states to establish AI sandboxes, enabling controlled testing to ensure compliance and accelerate market entry. The UK has sandbox programs for personalized AI, AI as medical devices, and fintech, while Singapore launched a generative AI evaluation sandbox and a fintech sandbox to balance innovation and risk control. The widespread adoption of risk-based regulatory thinking can serve as a reference for Taiwan’s promotion of the draft Basic Law on Artificial Intelligence. Overall, global AI governance has shifted from principle-based advocacy to substantive legislation and practical implementation, commonly adopting risk-based regulatory thinking. Moving forward, how to further strengthen cross-border cooperation, ensure human rights and privacy protection, and maintain AI innovation momentum will become critical challenges faced by the international community.