维基媒体基金会在 OpenAI 版权案中寻找文件 Amicus 简介

条例

维基媒体基金会在 OpenAI 版权案中寻找文件 Amicus 简介

在 Google 上的首选来源中添加 Unite.AI

2026年10月30日Wikimedia Foundation 宣布表示已要求美国纽约南区地方法院准许他提交法庭之友(amicus)，或称法庭之友(friend-the-court)，在In re: OpenAI, Inc. 版权侵权诉讼，合并了对OpenAI的12项诉讼，其中最著名的成员是纽约时报对OpenAI和Microsoft的诉讼。

基金会于2026年10月29日向西德尼·H·施泰因法官递交的呈文是一份请求准许呈文的请求，其中附有拟议的书状作为展品，并被称之为支持任何一方。 它根据法院的当地规则和2026年8月7日的关于法庭之友参与的命令提交，该命令于2026年9月24日修改；由于该命令指示未来的法庭之友迁移请假，维基媒体没有征求双方同意。 提议的书状有5 831个字，在法庭为法庭之友书状规定的6 000字限制之内。

Lex Lumina LLP的Rhett O. Millsaps II与基金会总法律顾问Stephen LaPorte、副总法律顾问Jacob Rogers和法律顾问Aly Marino一起担任记录律师。

在合并诉讼中，法院将裁定人工智能(AI)公司使用版权内容开发大型语言模型(LLMs)是否构成公平使用。 称基金会的裁决将有很大的影响，因为其管辖权、所涉各方以及迄今处理该问题的少数裁决。

"纽约时报"于2023年提起诉讼，指控OpenAI和Microsoft将Times拥有的内容用于任何目的，包括AI模式培训，直接违反了版权法。 OpenAI和Microsoft认为，基于公开数据的培训和部署模型是属于公平使用的技术发展。

基金会认为，过于限制性的公平使用版本可能会抑制技术创新，限制诸如维基百科等开源项目的工具可用性，而全面公平使用则可以产生一个更加封闭的网络，其中在线知识因缺乏可用的二级源材料而削弱。

在维基媒体运动博客Diff的一篇文章中，Foundation 称指出，如果培训步骤不被视为公平使用，称则鼓励出版商将更多信息锁定在屏障后面

基金会CEOBernadette Meehan 称的立场是“捍卫维基百科所依赖的自由和开放的人类知识”，将通过AI扩大获取知识的机会描述为必须尊重人民、贡献者和社区的强大目标，使可靠的知识成为可能。

逐项使用公平使用框架

在最高法院对Andy Warhol Foundation for the Visual Arts v. Goldsmith案的2023年裁决中，拟议简短地解释了其做法，认为在开发和运作聊天机的各个阶段，对复制作品的每一种不同用途都需要自己的四要素公平使用分析。 它援引了2025年Bartz诉Anthropic案的裁决，认为该裁决对在不同阶段评估相同副本的单独用途，而不是将它们纳入单一分析具有启发性。

该简报还争辩说，刮片——自动向网站提出技术请求索取作品副本——是获取，而不是版权使用，只有在复印后才应进行公平使用分析。 指出认为，某些形式的刮刮仍可能违反网站使用条款，国家合同或侵权法，数字千年版权法，或计算机欺诈和滥用法。

培训可能公平，检索-增强的一代可能不

简言之，在预训方面，用法可能是公平的：由于拷贝提供语言模式，而不是服务于作品的信息目的；鉴于百科全书作品的事实性质，在第二个因素上是中立的；在第三个因素上是正当的，因为开发者需要数十亿字来训练一个模型，使得整个作品的使用是合理必要的；在第四个因素上是公平的，因为预训本身不能取代市场上的原著。 基金会的公告总结了这一立场，即当它训练一些与原始来源不同且不需要满足版权许可条件的模型时，使用版权内容进行示范培训应被视为公平。

简介将检索到的生成(RAG)描述为一个聊天员搜索网页或其他数据库，然后识别、复制、排列和粘贴到模型上下文窗口中的相关内容块与用户的查询。 因为这样的系统将作品用于其最初的告知读者的目的，取了这些作品中最重要的部分，并且与人类观赏的源头竞争，简言之，这种使用可能失败了四因子测试。

基金会认为，不应将RAG和类似地基技术视为公平使用，提供这些技术的公司应尊重每个平台或出版商的许可证条款。 维基百科中的相关条目：CC BY-SA 4.0 许可

书状的结论要求法院裁定ChatGPT通用的LLM的预训是公平使用的，而OpenAI在ChatGPT的检索-强化生成功能中使用副本是不公平的，必须遵守维基媒体项目内容的创意通用许可证等许可证。

维基百科中的相关条目

维基百科拥有超过6700万篇文章，使用300种语言，每月有近25万志愿者献计献策，基金会将其描述为用于训练LLMs的优质数据集之一。 简言之，截至2025年1月，每月有超过26万志愿贡献者在基金会的13个自由知识平台上创建和温和的内容，维基百科每月浏览数十亿次。

简言之，来自刮刮和爬行的非人流量会给维基媒体的服务器带来压力，迫使它采用技术保障措施，而不良行为者往往会绕过这些保障措施，而且这种压力和保障措施都威胁到其网站对读者的无障碍访问。 称基金会刮掉其项目必须遵守其公开的API费率限制，或通过它的Wikimedia Enterprise API报价。 它还告诉法庭，机器学习工具已经帮助志愿者打击破坏行为，发现侵犯版权行为和不适当的个人信息，并将最需要改进的文章列为优先事项，认为将LLM开发贴上绝对不公平的标签违反版权法的宗旨。

Stephen LaPorte 称总法律顾问说，公平使用对话“不能限于营利组织，”并补充说，不认真适用公平使用原则的裁决可能威胁到自由知识的可持续性，并扼杀开源技术的发展。

维基媒体的动向加入了一波法庭之友活动：备案条目显示，2026年10月28日和29日，引擎宣传，作者联盟，公众知识，马修·萨格教授，布莱恩·L·弗莱(Brian L. Frye)，数字内容Next，以及包括美国出版商协会在内的一批出版组织搬来提交自己的法庭之友简报，而新闻出版商OpenAI和Microsoft之间的简易判决简报继续。 法院有广泛的酌处权来决定是否接受法庭之友书状，基金会的指出案状，在正式接受基金会提出的书状之前，Stein法官必须批准基金会的许可。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Regulation

Wikimedia Foundation Seeks to File Amicus Brief in OpenAI Copyright Case

Add Unite.AI to your preferred sources on Google

The Wikimedia Foundation announced on October 30, 2026, that it has asked the U.S. District Court for the Southern District of New York for leave to file an amicus, or friend-of-the-court, brief in In re: OpenAI, Inc. Copyright Infringement Litigation, the consolidation of 12 lawsuits against OpenAI whose best-known member is The New York Times’ case against OpenAI and Microsoft.

The Foundation’s filing, dated October 29, 2026, and addressed to Judge Sidney H. Stein, is a letter motion for leave to file, with the proposed brief attached as an exhibit and styled as supporting neither party. It was submitted under the court’s local rules and an August 7, 2026 order governing amicus participation, as modified on September 24, 2026; because that order directs prospective amici to move for leave, Wikimedia did not seek the parties’ consent. The proposed brief contains 5,831 words, within the 6,000-word limit the court set for amicus briefs.

Rhett O. Millsaps II of Lex Lumina LLP is counsel of record, alongside the Foundation’s General Counsel Stephen LaPorte, Associate General Counsel Jacob Rogers, and Legal Counsel Aly Marino.

In the consolidated litigation, the court will rule on whether artificial intelligence (AI) companies’ use of copyrighted content to develop large language models (LLMs) constitutes fair use. The Foundation said the decision will be highly influential because of the jurisdiction, the parties involved, and the small number of rulings that have addressed the question so far.

The New York Times filed its lawsuit in 2023, alleging that OpenAI and Microsoft’s use of Times-owned content for any purpose, including AI model training, directly violates copyright law. OpenAI and Microsoft contend that training and deploying models based on publicly available data is technological development that falls within fair use.

The Foundation argues that a too-restrictive version of fair use could inhibit technological innovation and limit the availability of tools that benefit open-source projects like Wikipedia, while blanket fair use could produce a more closed web in which online knowledge is weakened by a lack of available secondary source material.

In a companion post on Diff, the Wikimedia movement’s blog, the Foundation said that if the training step is not considered fair use, AI development would likely remain possible only for the largest corporations or the best-funded developers, leaving open-source developers unable to afford or even access the technology. A blanket ruling, it said, would instead encourage publishers to lock more information behind barriers, reducing the visibility that Wikipedia and many publishers depend on for readership and for recruiting new volunteers.

CEO Bernadette Meehan said the Foundation is taking a position “to defend the free and open human knowledge that Wikipedia depends on,” describing expanded access to knowledge through AI as a powerful goal that must respect the people, contributors, and communities that make reliable knowledge possible.

A Use-by-Use Fair-Use Framework

The proposed brief grounds its approach in the Supreme Court’s 2023 decision in Andy Warhol Foundation for the Visual Arts v. Goldsmith, arguing that each distinct use of copied works across the stages of developing and operating a chatbot requires its own four-factor fair-use analysis. It cites the 2025 Bartz v. Anthropic ruling as instructive for assessing separate uses of the same copies at different stages rather than blending them into a single analysis.

The brief also contends that scraping — making an automated technical request to a website for a copy of a work — is acquisition rather than a copyright use, and that only what is done with the copies afterward should undergo fair-use analysis. It notes that some forms of scraping can still violate site terms of use, state contract or tort law, the Digital Millennium Copyright Act, or the Computer Fraud and Abuse Act.

Training Likely Fair, Retrieval-Augmented Generation Likely Not

On pre-training, the brief argues the use is likely fair: transformative because the copies supply linguistic patterns rather than serving the works’ informational purpose; neutral on the second factor given the factual nature of encyclopedic works; justified on the third because developers need billions of words to train a model, making the use of entire works reasonably necessary; and fair on the fourth because pre-training itself does not substitute for the originals in the market. The Foundation’s announcement summarizes the position as holding that using copyrighted content for model training should be considered fair when it trains models that are something new and different from the original sources, with no requirement to meet copyright licensing terms.

The brief describes retrieval-augmented generation, or RAG, as a process in which a chatbot searches the web or another database, then identifies, copies, ranks, and pastes relevant chunks of content into the model’s context window alongside the user’s query. Because such a system uses the works for their original purpose of informing readers, takes the most important parts of those works, and competes with the sources for human viewership, the brief argues, the use likely fails the four-factor test.

The Foundation maintains that RAG and similar grounding techniques should not be treated as fair use and that companies offering them should respect each platform or publisher’s licensing terms. In Wikipedia’s case, that is the Creative Commons CC BY-SA 4.0 license, which requires credit to original creators and requires that adaptations remain under the same license.

The brief’s conclusion asks the court to rule that the pre-training of ChatGPT’s general-purpose LLM was fair use, while OpenAI’s use of copies in ChatGPT’s retrieval-augmented generation feature is not fair and must abide by licenses such as the Creative Commons license for Wikimedia Project content.

Wikipedia’s Stake and Other Amicus Filings

Wikipedia holds more than 67 million articles across 300 languages, with nearly 250,000 volunteers contributing each month, and the Foundation describes it as one of the highest-quality datasets used to train LLMs. The brief states that more than 260,000 volunteer contributors per month, as of January 2025, create and moderate content across the Foundation’s thirteen free-knowledge platforms, and that Wikipedia is viewed billions of times per month.

The brief states that non-human traffic from scraping and crawling strains Wikimedia’s servers, pushing it to adopt technical safeguards that bad actors often bypass, and that both the strain and the safeguards threaten the accessibility of its sites to human readers. The Foundation said scraping of its projects must respect its public API rate limits or go through its Wikimedia Enterprise API offerings. It also told the court that machine-learning tools already help volunteers fight vandalism, detect copyright violations and inappropriate personal information, and prioritize articles most in need of improvement, arguing that it would be contrary to the purposes of copyright law to label LLM development categorically unfair.

General Counsel Stephen LaPorte said the fair-use conversation “cannot be limited to for-profit organizations,” adding that a ruling that does not apply fair use principles carefully could threaten the sustainability of free knowledge and stifle the development of open-source technologies.

Wikimedia’s motion joined a wave of amicus activity: docket entries show that on October 28 and 29, 2026, Engine Advocacy, the Authors Alliance, Public Knowledge, Professor Matthew Sag, Brian L. Frye, Digital Content Next, and a group of publishing organizations including the Association of American Publishers moved to file their own amicus briefs, while summary-judgment briefing between the news publishers, OpenAI, and Microsoft continued. The court has broad discretion to decide whether to accept an amicus brief, the Foundation’s filing notes, and Judge Stein must grant the Foundation leave before its proposed brief is formally accepted.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.