邁入AI智能的新紀元，美國作為生成式AI發展重鎮，其當前美國法院審理之數宗生成式AI侵害著作權之案件，已成為全球關注之焦點。本研究動機即源於生成式AI技術變革對當前法制不確定性帶來的法律與技術衝突，並聚焦三項核心爭點進行研討：首先審視生成式AI侵權之訴訟，觀察美國法院於合理使用四大因素之運用上，是否有過度向「利用之目的與性質(第一因素)」傾斜之失衡現象；其次探討生成式AI複雜之生命週期中，其侵權責任邊界之界定與歸責主體之判定，最後探討在數位人工智慧時代中，法定損害賠償制度於數位侵權下質變為過度懲罰的失衡現象，容易於個案中累積出較高金額，從而引發民事救濟過度偏向懲罰性質的疑慮。 生成式AI於資料擷取與預訓練階段對資料進行重製，首先應該定性為「非表達性使用」之轉化性特質，抑或以實質構成對原著市場的替代？正衝擊現行美國《著作權法》第107條合理使用之四大法律因素，法院應如何在司法實務中動態調適各項審查要件，以在技術創新與原創作者之權益保障中建構動態平衡，實為現階段法制研究之首要課題。再者同法第504條之損害賠償機制，在面對生成式AI巨量數據資料之使用，相較當前以每件著作為計算基準之法定賠償，極易因乘數效果致生天價賠償，於此，因應如何避免懲罰性賠償效果，如何引導回歸損害填補原則之合理本質，亦是人工智慧時代下《著作權法》的重要使命與課題。 美國著作權法制長期將第107條之合理使用規範奉為圭臬，其開放式之例示特質為法規範賦予高度之動態彈性，卻也為法院援引時帶來諸多不確定性，致使實務判決與見解流於各自表述。觀察近來美國司法實務之演變趨向，法院適用合理使用四大審查要素時，顯有漸次向「使用的目的和性質(第一因素)」傾斜發展，由此衍生出「轉化性」與「可認知性傷害(市場的替代性)」，並將此概念躍升為核心的論證標的；後者具有經濟性之內涵，作為評估「對市場的影響(第四因素)」時之輔助指標。據此演進下，轉化性與資料取得之合法性成為主要論證，致使「對市場的影響(第四因素)」退居為次要地位。此外，自1909年美國全面修訂《著作權法》以來，即確立以區間數額(明訂法定上下限)作為法定賠償規範之主軸，立法機關賦予寬泛之量刑裁量權，此一制度導致司法實務上為達預防與威嚇之政策目的，頻生過苛裁定，進而累積為天價之法定賠償額。於救濟程序中，只要賠償數額落於法定處置區間內，上訴法院多傾向維持地方法院之原審裁量結果，加之原告選擇依法定賠償作為請求救濟時，則免除其對實際損害與所失利益之舉證責任，致使個案賠償之公平性，成為當前數位時代下，一個不容忽視的潛在問題。 綜觀諸多美國司法案例，關鍵爭點多聚焦於生成式AI直接引用原創作品，導致產出內容與原創作品間產生替代性，進而實質影響原創市場之占有率。有鑒於當前美國司法實務對生成式AI產出作品之「轉化性」利用之認定範圍似有過於寬泛之傾向，我國未來於法制運用上，應透過歸納司法實務之類型，建構明確之操作性定義。於立法政策層面，立法機關宜針對生成式AI系統開發與平台業者增修規範，要求落實對訓練資料與數據之風險評估義務，並強制導入AI產出物之數位浮水印標示機制，藉以降低構成輔助侵權或間接侵權之風險。此外對人工智慧系統之開發目的，司法裁量應依不同目的採取不同裁量標準，倘若生成式AI系統專為教育或其他公益之目的開發，應給予調降或免除法定賠償額度。綜上所述，面對新興之生成式AI侵權浪潮，我國除需明確規範「轉化性」使用的操作定義，更應積極增修人工智慧系統服務提供者之民事免責事由，並最終落實公平補償的精神，方能降低法律規範與技術發展間之衝突。;As we enter a new era of AI， the United States， a major hub for generative AI development， has seen several generative AI copyright infringement cases currently being heard in its courts become a global focus. This research is motivated by the legal and technological conflicts arising from the uncertainties in the current legal system caused by the technological transformation of generative AI， and focuses on three core issues: First， examining generative AI infringement lawsuits to observe whether US courts have an imbalance in their application of the four factors of fair use， leaning excessively towards the "purpose and nature of use (the first factor)" second， exploring the definition of the boundaries of liability for infringement and the determination of the responsible party within the complex lifecycle of generative AI; and finally， discussing the imbalance in the statutory damages system in the digital age， where it has degenerated into excessive punishment for digital infringement， easily accumulating high sums in individual cases and raising concerns about the excessive punitive nature of civil remedies. Generative AI， in its data acquisition and pre-training stages， reproduces data. Should this be characterized as "non-expressive use，" a transformative characteristic， or does it constitute a substantial substitution of the original work in the market? This challenges the four legal elements of fair use under Section 107 of the current U.S. Copyright Act. How courts should dynamically adjust various review requirements in judicial practice to establish a dynamic balance between technological innovation and the protection of original authors′ rights is a primary issue in current legal research. Furthermore， the damage compensation mechanism under Section 504 of the same Act， when faced with the use of massive amounts of data by generative AI， is prone to exorbitant damages due to the multiplier effect caused by the traditional statutory damages calculated per work. Therefore， how to avoid punitive damages and how to guide a return to the reasonable essence of the principle of damage compensation are also important missions and issues of copyright law in the era of artificial intelligence. The U.S. copyright law system has long regarded Section 107′s fair use principle as its cornerstone. Its open-ended nature grants the legal norms a high degree of dynamic flexibility， but also introduces considerable uncertainty into court applications， leading to divergent interpretations and opinions in practice. Observing recent trends in U.S. judicial practice， courts applying the four elements of fair use assessment have shown a gradual shift towards the "purpose and nature of the use (the first factor)，" giving rise to "transformationalism" and "perceptible harm (market substitutability)，" elevating these concepts to core arguments. The latter， with its economic implications， serves as a supplementary indicator when assessing "market impact (the fourth factor)." Consequently， transformationalism and the legality of data acquisition have become primary arguments， relegating "market impact (the fourth factor)" to a secondary position. Furthermore， since the comprehensive revision of the U.S. Copyright Act in 1909， a range of legally stipulated amounts (with clearly defined upper and lower limits) has been established as the main framework for statutory damages. The legislature has granted broad sentencing discretion. This system has led to frequent overly harsh rulings in judicial practice， driven by the policy objectives of prevention and deterrence， resulting in exorbitant statutory damages. In remedies， as long as the damages fall within the statutory range， appellate courts tend to uphold the original judgment of the district court. Moreover， when plaintiffs choose statutory damages as their remedy， they are exempt from the burden of proof regarding actual damages and lost profits. This makes the fairness of individual cases a potentially significant issue that cannot be ignored in the current digital age. A review of numerous US judicial cases reveals that key disputes often focus on generative AI directly referencing original works， leading to substitutability between the generated content and the original works， thereby substantially impacting the market share of original content. Given that current US judicial practice seems to have an overly broad scope for defining the "transformative" use of generative AI-generated works， Taiwan should， in its future legal application， construct clear operational definitions by summarizing types of judicial practice. At the legislative policy level， the legislature should amend regulations for generative AI system developers and platform operators， requiring them to implement risk assessment obligations for training materials and data， and mandating the implementation of a digital watermarking mechanism for AI outputs to reduce the risk of constituting accessory or indirect infringement. Furthermore， regarding the purpose of developing artificial intelligence systems， judicial discretion should adopt different standards depending on the purpose. If a generative AI system is developed specifically for educational or other public interest purposes， the statutory damages should be reduced or waived. In conclusion， in the face of the emerging wave of generative AI infringement， Taiwan needs to not only clarify the operational definition of "transformative" use， but also actively amend the civil exemption clauses for AI system service providers， and ultimately implement the spirit of fair compensation， in order to reduce the conflict between legal regulations and technological development.
