概述
高風險系統、禁止的行為、特定透明度義務以及系統性風險模型義務仍可能適用。
深入探討
Article 2(12) excludes certain free and open-source AI systems from the Regulation only where they are released under qualifying licences and are not placed on the market or put into service as high-risk systems, prohibited practices, or systems subject to specified transparency obligations. This provision is limited: it does not mean that every open-source project falls outside the Act, and it does not remove duties that may attach when another actor integrates a component into a regulated product or use. The Act has a separate, narrower rule for providers of general-purpose AI models. Some documentation and downstream-information obligations in Article 53(1)(a) and (b) do not apply when model parameters, including weights, architecture information, and usage information are publicly available under a free and open-source licence. The exemption does not cover the copyright-policy and training-content-summary duties in Article 53(1)(c) and (d). It also does not apply to a GPAI model presenting systemic risk; the systemic-risk obligations remain relevant. “Free” licensing is not the only condition. Public availability and what is actually released matter. A model card alone is not the weights; a source-code repository may not include model parameters. The legal treatment can also differ between a model and an AI system built using that model. Commercial distribution does not always erase every open-source provision, but monetized components and service arrangements can affect whether a condition is met. Read the relevant article and recitals rather than relying on a project’s label. Regulation 2026/1744 also amended Article 25(4): its written-information and assistance agreement duty does not apply to third parties publicly providing tools, services, processes, or components (but not GPAI models) under a free and open-source licence. This narrow exception does not exempt the integrated high-risk system provider; track each component and role separately.
現實世界的實施
某專案以開放授權釋出權重與架構,但另行檢查是否適用系統性風險義務。
某公司將開放模型整合至履歷排序服務中,並評估所產生系統的高風險用途。
某維護者將付費代管與支援服務與免費程式碼釋出分開記錄。
儘管享有有限的透明度豁免,模型提供者仍將著作權政策與訓練內容摘要工作保留在適用範圍內。
歐盟人工智慧法案中開源豁免的未來
開放且協作的人工智慧開發將持續演進,人工智慧法案本身也透過Regulation 2026/1744進行了修改。在依賴例外規定之前,應先查閱委員會的實施材料與彙編的EUR-Lex文本。維護具版本控制的釋出記錄,特別是當新增營利化、代管推論、支援服務或下游功能時。套件的變更可能會改變法律分類以及誰承擔義務。當專案新增代管推論、專有權重或下游整合時,請重新檢視授權與套件內容。
付諸實踐
為什麼這很重要
- 風險與安全. 災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
- 更明確的決策. 民眾和專業素養決定強而有力的安全政策在政治上是否可行。
- 突破炒作. 清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
留意
- 將存在風險視為科幻小說,同時能力複合。
- 混淆了表面產品安全與高度自治下的對準。
- 只給非英語和非專業觀眾留下低品質的資源。
後續步驟
- 單獨的產品危害、誤用和失控/失調風險。
- 詢問哪些證據會改變您對時間表和嚴重性的看法。
- 比起行銷主張,更喜歡主要來源和具體評估。
- 確定一條行動路徑:職業、政策、資金或技能——而不僅僅是意識。
不斷探索
常見問題
什麼是歐盟人工智慧法案中的開源豁免?
歐盟人工智慧法案對免費與開源軟體和模型包含有限的排除與豁免,但「開源」並非全面的豁免。高風險系統、禁止的行為、特定透明度義務以及系統性風險模型義務仍可能適用。
開源標籤是否能豁免人工智慧法案的所有義務?
這些條文是有條件的,且因系統與模型義務而異。
以開放授權釋出基礎模型,是否能決定招募產品是否為高風險?
根據本法,模型與人工智慧系統是不同的物件。
維護者應記錄哪些釋出事實?
元件層級的記錄有助於進行精確的範圍評估。
為什麼在數位綜合規則(Digital Omnibus)之後要檢查彙編法規?
這些修正案會影響現行的法定條文與解釋。