法學期刊
  • 社群分享
論著名稱:
生成式 AI 的市場競爭與事前規制初探(Preliminary Study on the Generative AI Market Competition and Ex-ante Regulation)
文獻引用
編著譯者: 魏杏芳
出版日期: 2024.05
刊登出處: 台灣/月旦法學雜誌第 348 期/75-95 頁
頁  數: 20 點閱次數: 664
下載點數: 80 點 銷售明細: 權利金查詢 變更售價
授 權 者: 魏杏芳
關 鍵 詞: 生成式 Al技術邏輯基礎模型微調生態系鎖入事前規制
中文摘要: 人工智慧系統其主要目的在自動產出文字、圖影、聲音、影像等內容者,稱為「生成式 AI」,通常是在通過數據預訓練的基礎模型上調整開發而成。「數據的量與質」、「演算法模型與專業人力資源」、「訓練運算成本及財務負擔」是決定生成式 AI 市場競爭的三大要素,可觀察到基礎模型市場已有集中的趨勢,Microsoft、Google、Meta 等事業是重要參與者。競爭主管機關公平會應關注基礎模型、雲端運算等相關市場的結構與限制競爭問題,並以支配事業違法濫用市場力為執法優先項目,「生態系鎖入」類型尤其值得重視。依生成式 AI 的技術邏輯及對人類社會造成的風險,完全依賴被動的事後管制及救濟,並不可行。歐盟是以正式立法對生成式 AI 及基礎模型進行事前規制,我國現行法對於生成式 AI 的規範並非完全真空,但必須配合技術發展的現實進行調整。由於生成式 AI 與基礎模型的同質性,基礎模型本身的不當將由下游應用「繼承」;生成式 AI 演算法不透明;生成式 AI 價值鏈上的參與者眾,對 AI 系統風險的控制力各不相同,責任歸屬難以確定等因素,我國在政策上應仔細籌謀具事前規制效果的法規,至少從根本的數據蒐集處理、演算法設計與應用的透明性、正確性、AI 價值鏈上參與者地位的確定及義務、必要的人為監管等各方面預作安排,應是可行之道。
英文關鍵詞: Generative AlTechnological LogicsFoundation ModelFine-tuningEcosystem Lock-inEx-ante Regulation
英文摘要: An artificial intelligence (AI) system aimed at automatically producing text, image, audio or video content can be classified as a generative AI The emerging paradigm today for developing generative AI is based on the so-called foundation model which is trained on broad data and capable of functioning a wide range of tasks after being fine-tuned or adapted. The quantity and quality of data used for training, algorithmic models and required expertise and talent, as well as computational resources are the three major parameters for the competition in the generative AI market. The concentration of market for foundation models has been observed and the big techs such as Microsoft, Google and Meta continue to be the key players on the market. The competition authority should closely watch the structure and competition concerns of the relevant markets. Any abuse of market power by dominant companies on the market for foundation model and cloud computing will be the enforcement priority and “ecosystem lock-in” should be put on the top of the list. Considering the technological logics of generative AI and the associated risks on human society arising from them, entirely relying on the passive ex-post regulation is arguably infeasible. While the formal ex ante enactment for the purpose of controlling generative AI and foundation models has been introduced in the EU, the applicable laws and regulations for the emerging technology are not in a vacuum at all in our country, but thorough review and making the corresponding adjustments based on the AI technological reasonableness is necessary. Due to the homogenization characteristic of foundation models, the failures and flaws in the foundation model will be inherited by the downstream application models. Generative AI algorithms are lack of transparency and the capability of risks control for each participant on the general AI value chain varies and therefore making the accountability of AI and identifying the one who should be responsible for any harm imposed by generative AI become challenging and difficult. Considering the above- mentioned reasons, it is highly suggested that our government should vigorously and carefully calibrate the ex-ante regulation system for generative AI from now on and incorporate, at least, the clauses dealing with data collection and processing, algorithm design and transparency as well as accuracy, the obligations and responsibilities for each actor on the generative AI value chain, and the necessary human oversight on the generative AI.
目  次: 壹、前言-討論生成式 AI 議題的前提認識
貳、人工智慧的三大要素-數據、演算法與運算能力
參、生成式 AI 與基礎模型概述
肆、生成式 AI 市場及其競爭分析
伍、對生成式 AI 施以事前規制的探討
陸、對我國生成式 AI 政策及措施的建議-代結論
相關法條:
相關判解:
相關函釋:
相關論著:
魏杏芳,生成式 AI 的市場競爭與事前規制初探,月旦法學雜誌,第 348 期,75-95 頁,2024年05月。
返回功能列