美國科技巨頭 Meta（前 Facebook 公司） 近日再度陷入爭議，被指控非法盜用 2396 部成人影片作為人工智慧訓練數據，引發產業界與法律界的高度關注。根據加州聯邦法院的正式訴訟文件，兩家成人影視公司已向法院提起訴訟，指控 Meta涉嫌透過BT網路 下載並分發其版權影片，用於訓練生成式AI模型。

原告方強調，他們自主研發的追蹤系統鎖定涉案影片的流向，發現大量涉及下載與分發的IP地址均隸屬於Meta，而且整個數據傳輸與存取的紀錄完整可查，堪稱「鐵證如山」。在美國版權法中，若屬於故意侵權，每一部影片的最高賠償額可達15萬美元。以涉案的2396部影片計算，最高總額可達3.59億美元（約新台幣113億）。

訴狀中，原告律師特別闡述為何這些影片對AI訓練極具價值。他們指出，成人影片往往具備 高畫質、分辨率清晰、鏡頭持續時間長、表情自然、對話有節奏、動作具連貫性，且場景變化相對簡單。對於訓練生成式影片模型而言，這些特徵可說是「理想素材」，甚至具備其他一般影視作品無法比擬的優勢。此外，成人影片還具有某些「獨特場景」，使得這類內容在AI訓練資料中別具價值。

從商業角度來看，透過BT網絡的「以種換種」機制，Meta不僅能以極低成本快速下載這些影片，還能藉由熱門片源交換到更多數量、更大規模的數據集。也因此，雖然表面上看似「免費」取得的資料，最後可能轉化成最昂貴的法律賬單。

這起案件不僅挑戰Meta在數據蒐集與使用上的合法性，更再度點燃業界對AI訓練資料版權問題的討論。外界普遍認為，若訴訟成立，這將成為生成式AI發展過程中最具指標性的版權糾紛之一，並可能對全球AI企業未來如何合法蒐集與使用內容資源帶來深遠影響。

U.S. tech giant Meta (formerly Facebook) has once again found itself embroiled in controversy, this time accused of illegally using 2,396 adult films as training data for artificial intelligence, sparking heated debate within both the industry and legal circles. According to official filings with the California federal court, two adult film companies have filed suit, alleging that Meta downloaded and distributed their copyrighted works via BitTorrent networks for the purpose of training generative AI models.

The plaintiffs emphasized that their self-developed tracking system was able to trace the distribution of the films, uncovering that a large number of the IP addresses involved in the downloads and redistribution were registered to Meta. The records of data transmission and storage were complete and verifiable, described by the plaintiffs as “ironclad evidence.” Under U.S. copyright law, willful infringement can incur statutory damages of up to $150,000 per work. With 2,396 works in question, the total damages could reach as high as $359 million.

In the complaint, the plaintiffs’ lawyers explained why these films are especially valuable for AI training. They pointed out that adult films typically feature high resolution, clear image quality, long continuous shots, natural expressions, rhythmic dialogue, coherent motion, and relatively simple scene changes. For training generative video models, these qualities make adult content “ideal material,” offering advantages unmatched by many mainstream films. Furthermore, the genre provides certain “unique scenarios” not found in other types of training data.

From a business perspective, through BitTorrent’s “seed-for-seed” exchange mechanism, Meta was allegedly able to download such content at virtually no cost, while simultaneously trading popular titles for access to larger and more diverse datasets. As such, what seemed like “free” data could ultimately become the most expensive bill Meta has ever faced.

This lawsuit not only challenges the legality of Meta’s data collection and usage practices but also reignites broader debate over copyright and AI training datasets. Many observers believe that, if successful, the case could become one of the most significant copyright disputes in the development of generative AI, with far-reaching implications for how AI companies worldwide legally acquire and use content in the future.