Wikimedia's Two-Part Copyright Test for AI: What Changes Between Model Training and Model Output
Most debates about AI and copyright collapse into a single question — is model training lawful? The position put forward by the Wikimedia Foundation suggests the more productive question is which stage of the pipeline is being judged. Read as a whole, it sets out a two-part test: collecting material from a website is not automatically an infringement, and placing copyrighted works into large language model pre-training without permission can be covered by fair use as long as what the model produces ends up "different" from the original. The standard tightens at the delivery stage. If a chatbot hands back an article in a form that removes the reader's need to visit the source, the AI company behind it should, in this reading, satisfy licensing requirements.
The argument matters beyond the encyclopedia that inspired it. It is an attempt to convert an emotionally charged dispute into a sequence of separable acts — access, ingestion, transformation, reproduction — so that courts and industry can reach different conclusions about different steps instead of issuing one verdict that governs the entire technology stack.
The claim, stated in four moves
- Access alone is not the offence. Scraping site content does not, on its own, necessarily constitute copyright infringement.
- Training can be tolerated conditionally. Where copyrighted works are used in LLM pre-training without permission, fair use may apply if the generated result is "different" from the original.
- Substitution is the red line. When a chatbot outputs an article directly and readers no longer need the underlying work, licensing requirements come into play.
- Open licences remain conditional. Material published under an open licence still carries obligations that downstream reuse must respect.
Where the encyclopedia's own licence shapes the argument
The Foundation's stance cannot be separated from the licensing model Wikipedia runs on: a Creative Commons attribution and share-alike arrangement. Wikimedia's concern, as described, is that models reproduce text published under that open licence without necessarily meeting the conditions attached to it — crediting the source, and distributing derivative material on the same terms. Openness, in other words, is not the same as a waiver.
From there the position becomes deliberately two-sided. If a court were to treat AI training as categorically unrestricted, the Foundation's worry is that such a ruling would make it easier to ignore licence conditions and weaken the flow of content between authors and readers. If, conversely, rights holders win control that is too broad, the room for innovative and transformative uses of copyrighted work shrinks. The recommendation is therefore not a verdict for one party but a demand for a finer-grained analysis.
Training versus output: a side-by-side view
| Stage of the AI pipeline | How the position frames it | Practical consequence for AI companies |
|---|---|---|
| Crawling or scraping website content | Does not automatically amount to copyright infringement | Access by itself is not treated as the decisive legal event |
| Including copyrighted works in LLM pre-training without permission | Fair use may apply where the generated output is "different" from the original | A conditional path to train without prior authorisation, tied to how far outputs depart from sources |
| A chatbot reproducing an article for the user | Licensing requirements should be met once the answer replaces a visit to the work | Permission becomes necessary at the point of delivery, not at training |
| Reuse of Wikipedia text released under an open licence | Attribution and share-alike conditions still have to be honoured | Compliance obligations follow the licence even when access was lawful |
The balance sheet: risk and usefulness together
The Foundation's framing does not treat AI tools as purely a threat to encyclopedia projects. On the maintenance side, such tools are described as capable of assisting Wikipedia editors in several concrete ways:
- Spotting destructive edits.
- Flagging copyright violations.
- Detecting personal information published in doxing fashion.
- Locating entries that need improvement.
The offsetting concern is behavioural rather than technical. If a chatbot simply supplies the information, users may stop going to Wikipedia and to other original websites. For open knowledge projects that depend on community participation, that loss of return visits is not a traffic statistic alone; it erodes the volunteer upkeep on which the content itself rests.
The litigation backdrop
Underneath the policy argument sits an active dispute. In 2023 The New York Times sued OpenAI and Microsoft, alleging that the two companies had scraped large volumes of news articles without permission and used them to train AI systems. Both companies have countered that the training at issue is fair use.
- Plaintiff: The New York Times, claiming unauthorised scraping of news articles for AI training.
- Defendants: OpenAI and Microsoft, arguing that the relevant training is fair use.
- Year filed: 2023.
- Third-party concern: the Wikimedia Foundation warns that a ruling failing to weigh the differences between training and generated content could shape future AI development, the open web, and the way users obtain information.
Facts and figures cited in the position
- 2023 — The New York Times brings suit against two companies over article scraping for AI training.
- Two companies named as defendants: OpenAI and Microsoft, both asserting fair use for the training in question.
- One licence model at stake: Wikipedia's Creative Commons attribution and share-alike terms, which the Foundation says models often reproduce without fulfilling.
- Two legal outcomes flagged as risky: blanket permission for training, and overly broad control by rights holders.
What to watch next
- Whether any ruling separates the training stage from the generation and delivery stage, rather than treating AI as a single act.
- How courts or settlement terms define "different" enough output to qualify as fair use.
- Whether attribution and share-alike obligations attached to openly licensed text are enforced in model outputs.
- How licensing duties are applied when an assistant's answer functionally replaces a visit to the source.
- The knock-on effect for open knowledge projects whose maintenance model depends on readers returning to the original site.
Frequently asked questions
Is the Foundation saying website scraping is legal?
Not in absolute terms. The position is that scraping does not automatically constitute infringement — a rejection of blanket characterisation rather than a clean clearance.
When could unlicensed pre-training be acceptable?
Where copyrighted works are used in large language model pre-training and the model's generated result comes out "different" from the original, fair use may apply.
When is a licence required?
Once a chatbot outputs the article itself and readers no longer need to access the underlying work, the expectation is that AI companies meet licensing requirements.
Why does Wikipedia's own licence feature so prominently?
Because the encyclopedia's content is published under an open Creative Commons attribution and share-alike licence, and models that reproduce that text may not honour the conditions that travel with it.
Is the position anti-AI?
No. It also points out that AI can help editors detect vandalism, copyright violations and doxing, and identify articles that need work.
Where can readers send questions or corrections about this brief?
Contact the AI & Data News editorial desk at editorial@aidatanews.com. Author: David Nakamura, AIdataNewsGlobal.