Breaking
SEDelays and missorting of early votes in Postnord's deliveriesESABBA's Benny Andersson supports Magdalena Andersson in the final stretch of the Swedish electionsDEHouthi militia intensifies attacks on Saudi ArabiaRUIn the Saratov region, two people were crushed under a slab while dismantling a buildingESBarcelona seeks to maintain its perfect streak against LevanteARThe Houthis control the strategic Bab al-Mandab Strait amid escalating regional tensionsBRCandidates for the AM government participate in a series of interviews on Rede Amazônica; see dates and how to watchDEAI industry under pressure: debate about security risks and decelerationDELow water on the Rhine: Economic effects and challengesRUBoris Johnson and European officials could be hit by a drone on the border between Ukraine and PolandSEDelays and missorting of early votes in Postnord's deliveriesESABBA's Benny Andersson supports Magdalena Andersson in the final stretch of the Swedish electionsDEHouthi militia intensifies attacks on Saudi ArabiaRUIn the Saratov region, two people were crushed under a slab while dismantling a buildingESBarcelona seeks to maintain its perfect streak against LevanteARThe Houthis control the strategic Bab al-Mandab Strait amid escalating regional tensionsBRCandidates for the AM government participate in a series of interviews on Rede Amazônica; see dates and how to watchDEAI industry under pressure: debate about security risks and decelerationDELow water on the Rhine: Economic effects and challengesRUBoris Johnson and European officials could be hit by a drone on the border between Ukraine and Poland
Back生成式AI的下一場法律戰爭:日本推動訓練資料透明度原則
生成式AI的下一場法律戰爭:日本推動訓練資料透明度原則
Tech
自由时报1 hour agoTech4 min readChinaView translation

生成式AI的下一場法律戰爭:日本推動訓練資料透明度原則

日本政府採納「生成式AI智慧財產保護與透明度原則」,要求AI業者提高訓練資料透明度,在創新與權利保護間建立資訊對稱。

Quick Look

日本政府於八月二十五日採納「生成式AI智慧財產保護與透明度原則」,要求AI開發者揭露訓練資料概要與蒐集方式。此舉旨在解決資訊不對稱問題,推動AI著作權治理從「許可或禁止」轉向「透明義務」,為產業發展與權利保護尋求平衡。

AI-generated summary

Why It Matters

日本著作權法第三十條之四長期對AI訓練採取寬鬆態度,允許為資訊分析目的利用著作物。然而,隨著生成式AI發展,權利人面臨資訊不對稱,難以判斷作品是否遭濫用。

Font size

生成式AI的下一場法律戰爭,可能是:用了什麼,能不能說清楚?

日本政府八月二十五日正式採納生成式AI智慧財產權保護相關的「生成式AI智慧財產保護與透明度原則(Principle Code)」,要求AI業者提高訓練資料透明度,包括揭露模型訓練所使用資料的概要、類型與蒐集方式等資訊。值得注意的是,這套規範目前並非具有法律強制力的硬法,也不是要求企業公開完整資料庫,而是透過軟法治理提高AI開發者的說明責任;其適用思維也不限於日本本土企業,而將在日本提供相關AI服務的海外業者納入考量。日本沒有突然關上AI發展的大門,AI可以繼續學習,但訓練資料還能不能永遠是一個誰也看不見的黑箱?

這項政策之所以值得注意,與日本著作權法長期對資料分析及機器學習採取相對寬鬆的態度有關。

日本《著作權法》第三十條之四允許在一定條件下,為資訊分析等「非享受著作物所表達思想或感情」目的利用著作物,因此長期被視為對AI訓練較為友善的法律環境。但這項規定從來不是「AI可以任意使用所有著作物」的免責金牌。法律本身仍設有不得不當損害著作權人利益等界線,而生成式AI快速發展後,更大的問題逐漸浮現:權利人甚至不知道自己的作品是否被使用,又如何判斷自己的權利是否受到影響?當法律只討論「使用是否合法」,卻沒有足夠資訊讓權利人知道「究竟使用了什麼」,實體權利就可能因資訊不對稱而難以行使。

因此,日本這次政策真正重要的地方,是把AI著作權治理從單純的「許可或禁止」往前推進到「透明義務」。

這是一個很重要的法律思維轉變。過去AI業者與創作者的爭議經常被簡化成兩個極端:一方認為訓練AI必須大量取得授權,否則就是侵權;另一方則主張機器學習需要龐大資料,如果每一筆資料都必須事前取得許可,AI產業根本難以發展。但兩者之間其實存在第三種制度選擇:法律可以保留合理的資料分析空間,同時要求大型AI業者對資料來源負擔一定程度的透明責任。換句話說,「可以訓練」與「完全不必說明」並不是同一件事。

透明制度的價值,也不只是保護著作權人。

對AI產業而言,清楚說明資料來源與取得方式,反而可能降低未來法律風險。當AI模型進入企業、政府、醫療、教育及專業服務市場,使用者愈來愈在意模型是否涉及未授權內容、個資、營業秘密或其他高風險資料。如果開發者完全無法說明訓練資料的基本來源,下游企業就很難進行法遵與風險評估。因此,訓練資料透明度未來甚至可能從「法律負擔」轉變成「市場競爭力」。誰能證明自己的資料治理更完整、授權來源更清楚,誰就更容易取得大型企業與政府客戶的信任。日本選擇先以Principle Code推動,而非立即祭出全面強制規範,也反映其政策思維:先讓市場形成透明標準,再觀察是否需要進一步立法。

日本的選擇值得台灣參考。

台灣同樣希望發展自主AI,也擁有新聞、出版、影視、音樂及大量中文內容產業。如果未來政策討論仍停留在「AI訓練是不是合理使用」的二分法,很容易讓科技業與內容產業陷入長期對立。

更務實的方向,是先建立最低程度的資料治理規則,例如要求一定規模以上的基礎模型開發者揭露訓練資料的主要類型、來源方式、是否使用授權資料,以及提供權利人合理的查詢或申訴管道。這並不等於要求企業公布每一筆資料,更不應強迫公開模型架構或營業秘密,而是在創新與權利保護之間建立最基本的資訊對稱。生成式AI時代真正成熟的法律制度,可以先要求開發者回答「你的AI究竟怎麼學」。日本正在傳達的訊息很清楚:可以讓AI繼續學習,但不能讓訓練資料永遠留在黑箱裡。

What to Watch

AI outlook — possibilities, not facts

  • 日本政府將觀察市場對Principle Code的反應,評估是否進一步立法。

    Likely · Within months

Open Questions

  • 未來是否會將原則轉化為強制性法律?
  • 海外業者如何配合日本的透明度要求?

Related Topics

This article was originally published by 自由时报.

Related Stories

Anhui brain-computer interface technology accelerates from laboratory to clinical application
Tech·

Anhui brain-computer interface technology accelerates from laboratory to clinical application

Several brain-computer interface teams in Anhui are promoting the clinical application of the technology. Aibo Intelligent has developed lower-limb exoskeleton robots to achieve "brain-reading" assisted rehabilitation, and Fuqing Medical has developed a non-invasive deep brain regulation system to achieve precise "brain-writing" intervention, which has been applied in more than 200 hospitals across the country.

中国新闻网
2 min read
Gao Wen, academician of the Chinese Academy of Engineering: China's scientific research system should "advance in four directions" to take the lead in scientific research competitions
Tech·

Gao Wen, academician of the Chinese Academy of Engineering: China's scientific research system should "advance in four directions" to take the lead in scientific research competitions

Gao Wen, an academician of the Chinese Academy of Engineering, proposed at the 2026 Pujiang Innovation Forum that China's scientific research system should "advance in four directions" by embracing iterative scientific research paradigms, reforming funding mechanisms, strengthening the closed loop of scientific research and manufacturing, and building AI scientific research infrastructure to gain advantages in the new round of scientific research competition.

中国新闻网
2 min read
Shanghai's "Pearl Constellation Project" was officially launched, aiming to build a space-based computing infrastructure with global coverage
Tech·

Shanghai's "Pearl Constellation Project" was officially launched, aiming to build a space-based computing infrastructure with global coverage

During the 2026 Pujiang Innovation Forum, Shanghai’s “Pearl Constellation Project” was officially launched. The plan has the long-term goal of building a "gigawatt-level space-based computing constellation" and aims to build a space-based intelligent computing infrastructure with global coverage and space-ground coordination by 2030, and promote the extension of intelligent computing capabilities to low-Earth orbit.

中国新闻网
2 min read
2026 AI empowers the green development of manufacturing industry theme event held in Ningbo
Tech·

2026 AI empowers the green development of manufacturing industry theme event held in Ningbo

The 2026 AI-empowered manufacturing green development themed event was held in Ningbo. The event released the "Ningbo Green and Low-Carbon Manufacturing One-stop Service Platform" and the first batch of AI industrial application products, aiming to optimize the production process through artificial intelligence technology, reduce energy consumption and pollution emissions in the manufacturing industry, and promote the green transformation of the industry.

中国新闻网
4 min read
More on this topic生成式AI