Wikipedia licenses for AI companies
The discussion about data licenses for AI companies increasingly involves non-profit organizations like Wikipedia. Jimmy Wales, co-founder of Wikipedia, is seeking more licensing agreements with AI companies, similar to the existing Google agreement. This is a reaction to rising server and storage costs incurred by AI bot scraping. The central question is how to find a fair balance between open data and AI without compromising the principle of free knowledge sources.
Source: Own Representation
Wikipedia and AI Licenses
Wikipedia, operated by the non-profit Wikimedia Foundation, is mainly funded by donations. The content is available under Creative-Commons-Lizenz CC BY-SA 3.0 and the GNU Free Documentation License, which also permit commercial use under the terms of the license. This openness makes Wikipedia attractive to AI companies, as the articles are editorially reviewed, versioned, sourced, and structured – an ideal foundation for Knowledge Graphs and Training Data.
AI bots crawl according to Wales "the entirety of the site", , leading to a disproportionate increase in caches, RAM, and bandwidth. Wikipedia donors want to support free education, not subsidize the infrastructure costs of multi-billion dollar AI corporations.
The first major licensing agreement is the one between the Wikimedia Foundation and Google for the service „Wikimedia Enterprise“. This commercial API service, launched in 2021, is aimed at large-scale users such as search engines and AI firms. While regular users can continue to use Wikipedia for free, companies requiring millions of requests, real-time mirroring, and customized data feeds receive stable, contractually regulated access for a fee. Google pays for this service , while the Internet Archive, as a non-profit, receives free full access.
This approach shifts value creation: instead of free scraping, there is now a product tailored for data-intensive applications. In parallel, AI companies are negotiating license packages with media houses. OpenAI has concluded agreements with publishers like Axel Springer and the Financial Times. Reddit stated in its IPO filing that it has generated approximately US$203 million through data licenses, , including a contract with Google for about US$60 million per year. Wikipedia therefore faces the question of why it should continue to provide free infrastructure for AI models when other platforms sell their data.
Source: wikimediafoundation.org
The Wikimedia Foundation is developing an AI strategy that puts people at the center.
Legal Aspects
The legal situation for training large language models is a global gray area. . In the EU, the Copyright Directive for the Digital Single Market (DSM Directive) permits text and data mining, provided rights holders do not explicitly opt out. The upcoming EU AI Act tightens requirements for "General Purpose AI": providers must publish a detailed summary of their training data and respect copyrights and opt-out signals. This serves transparency, enabling rights holders to track the use of their content and, if necessary, negotiate or sue.
In the US, the debate is dominated by the concept of „Fair Use“. . AI companies argue that training on copyrighted works constitutes transformative use, as no 1:1 copies are distributed, but rather statistical patterns are learned. However, courts do not consistently accept this argument. A Munich court ruled that training ChatGPT on protected song lyrics violates German copyright law and ordered OpenAI to pay damages to GEMA. The licensing of LLM training data depends on exceptions, opt-out mechanisms, transparency obligations, and court proceedings, prompting platforms like Wikipedia to seek predictable licensing models.
Source: digitalzentrum-berlin.de
The EU AI Act establishes a legal framework for the use of Artificial Intelligence and influences licensing models.
A concise introduction to the fundamental issues of "Copyright and AI Training" is provided by this English-language expert discussion.
Costs of AI Scraping
For non-profit platforms like Wikipedia, AI scraping primarily incurs three cost blocks: additional infrastructure, technical countermeasures, and governance overhead. The demand for servers, bandwidth, and caching resources increases as AI bots automatically scan large portions or entire dumps of the project. Non-profit organizations must decide how to defend themselves against excessive scraping. In addition to traditional methods like robots.txt, the Wikimedia Foundation is discussing specialized solutions such as Cloudflare „AI Crawl Control“. . Furthermore, governance structures are needed to decide which companies to negotiate with and when free access undermines community goals.
Other platforms react differently: Reddit relies on paid data licenses but is under surveillance by the US-Federal Trade Commission for selling user data. At the same time, Reddit is suing Anthropic for breach of contract in scraping to protect its licensing model. For projects like Wikipedia, the room for maneuver is smaller. They can offer paid enterprise models but must explain to their community why certain forms of AI scraping, beyond a certain extent, are no longer considered legitimate use but a cost trap.
Insights into the tension between free culture and sustainability are offered by Jimmy Wales in lectures.
Models for Fair Compensation
The central question is what a fair balance between open data and AI looks like, protecting innovation and the common good. A tiered access model is an obvious approach: individuals and small projects can use content freely, respecting the CC license. Large commercial entities, especially operators of AI systems, enter into license agreements with clear terms regarding scope, attribution obligations, deletion rights, and liability, as demonstrated by Wikimedia Enterprise or the deals struck by major publishers with OpenAI.
Source: user-added
Two glowing jellyfish, one blue, one green, against a black background with grid lines and brackets.
A second building block is technical signaling: rights holders must be able to indicate machine-readably whether their content may be used for AI training, and crawlers from AI companies must respect these signals. A third element is standardized compensation models: instead of individual deals, collective rights management organizations or industry-wide framework agreements could collect and redistribute license fees for defined usage categories, comparable to music licensing.
YouTube allows creators to explicitly agree that their videos may be used for AI system training by external companies; this setting is disabled by default. This suggests a future where open content is no longer automatically considered a free raw material source for AI corporations, but where usage rights, compensation, and opt-out options are transparently negotiated – including for non-profit projects like Wikipedia.
The basics of Creative Commons licenses are explained in this German-language explainer video. A compact introduction to the European perspective on AI regulation and copyright is offered by this lecture.
Conclusion and Outlook
Wikipedia stands for the idea of free knowledge, but "freely accessible" does not mean "freely exploitable for any business model." The decision by Jimmy Wales and the Wikimedia Foundation to pursue more licensing deals with AI companies is an attempt to preserve Wikipedia's principles in an AI-driven world. Open content should remain open, but those who use it commercially on an industrial scale should contribute a fair share to the infrastructure's financing.
Whether this approach is successful depends on how courts, regulators, and the public answer the question of a fair balance between open data and AI. The goal is to stop passing on the true costs of AI scraping to volunteers and donors.