A Line, Not a Verdict: Why Wikipedia's Operator Wants AI Copyright Split in Two
Copyright litigation against AI companies has a habit of arriving as a morality play. Either training a model is transformative fair use and the last decade of machine learning was lawful, or it is mass infringement and the industry was built on stolen paper. The Wikimedia Foundation — the nonprofit behind Wikipedia — has now walked into that room and asked the judges to stop framing it as a play at all. Its position, filed as an amicus in a cluster of landmark cases including The New York Times' suit against OpenAI, is that the law should treat two separate acts separately: ingesting copyrighted work to train a model, and emitting that work back to a user.
That sounds like legal hair-splitting. It is actually the most consequential design decision in the whole debate, because it determines not just who pays whom, but what kind of internet survives the next five years of AI. Below: the mechanism, the trade-offs, who wins under each rule, and what it means for teams that ship AI features for a living.
The argument in one sentence: training and output are different products

Figure: Editorial context image from the source page (preview only; rights remain with the publisher) · Credit / further reading: www.theverge.com
The Wikimedia Foundation's framing is roughly this: courts should give broad protection to training, because a model that learns statistical structure from a corpus is not the same thing as a machine that redistributes the corpus. But that protection should not automatically extend to what a chatbot says to a user — especially when the model produces a near-verbatim passage that substitutes for the original.
Layer one: ingestion
At the ingestion layer, the question is whether copying millions of documents into a training pipeline is the kind of use that copyright's fair-use doctrine tolerates. The arguments on the permissive side are familiar: the copy is intermediate rather than expressive, the model learns patterns rather than storing a library, and the downstream use is different from reading the source. The arguments on the restrictive side are equally familiar: the copying is total, it is commercial, and it directly enables a product that competes with the people whose work was copied.
What Wikimedia adds is an institutional stake. Wikipedia's text is freely licensed but heavily attributed, and the volunteer community depends on being able to use AI-assisted tooling — translation, summarization, patrolling edits at scale, spotting vandalism. If training on publicly available text becomes presumptively unlawful, the nonprofit sector's access to those tools narrows sharply, while deep-pocketed licensees carry on. That is a distributional argument, not just a doctrinal one.
Layer two: emission
At the emission layer, the analysis changes shape. When a model outputs a passage that closely tracks a copyrighted work, the harm is direct and measurable in a way that training harm is not: the user may have substituted the output for a purchase, a subscription, or a page view. Here the permissive arguments weaken, because the copy is no longer intermediate — it is the product.
Crucially, emission is the layer engineers can actually control. Retrieval, output filters, memorization testing, refusal policies, and citation behavior are all product decisions. Training corpora are chosen once; output behavior is tuned continuously.
Why the split matters procedurally
If courts treat training and output as one question, then evidence about regurgitation becomes evidence about training. A model that leaks a paragraph of a paywalled article in one prompt suddenly imperils the legality of the entire corpus. Conversely, if courts treat training as lawful without ever reaching output, plaintiffs lose their most concrete, demonstrable harm. Splitting the questions gives each side the evidence that actually matches the claim — and gives builders a rule they can engineer against.
What a black-and-white ruling actually breaks
The permissive extreme
A blanket holding that training on anything public is fair use, with no meaningful output-side constraint, would do two things. It would legitimize a market in which the cheapest strategy is to ingest everything and let the model paraphrase — good for deployment speed, corrosive for the licensing markets that fund professional content. It would also hand a permanent cost advantage to firms large enough to absorb litigation risk, which is precisely the asymmetry the nonprofit sector worries about.
The restrictive extreme
A blanket holding that unlicensed training is infringement would be worse in a subtler way. It would not retroactively delete models, but it would convert every future corpus into a licensing negotiation with unmanageable transaction costs. Small teams would not license; they would simply train on less and worse material, or buy from whoever already had a deal. The result is consolidation dressed up as protection. Wikimedia's warning — that Wikipedians lose access to useful AI tooling — is the same mechanism applied to a nonprofit.
The middle path's own costs
Honesty requires naming the downside of the split: it is less predictable. A two-part test means more litigation phases, more expert evidence about memorization and substitutability, and more years before anyone knows the rules. Parties who want certainty will not get it. The counterargument is that certainty purchased by collapsing two genuinely different acts is false certainty — it will be litigated anyway, just later and more expensively.

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: www.theverge.com
Follow the money: who wins under each rule
| Rule | Model developers | Rights holders | Open/knowledge commons | Enterprise buyers |
|---|---|---|---|---|
| Training broadly permitted, output policed | Broad corpus access; must invest in output controls | Licensing leverage shifts to output and brand claims | Best case: tools stay cheap and available | Clear compliance target |
| Training broadly restricted | Barriers to entry; incumbents with licenses favored | Strong licensing market; long negotiation cycles | Worst case: tooling priced out of reach | Vendor risk and indemnity costs rise |
| No split (single question) | Binary outcome, high volatility | Winner-take-all litigation | Policy hostage to one verdict | Contracts renegotiated after each ruling |
The table is deliberately crude, but it shows the structural point: the split converts a binary bet into a compliance problem. Compliance problems are boring, expensive, and — for anyone who has to ship software — vastly preferable to a coin flip.
The open web as collateral damage
The under-discussed stake here is not Wikipedia's servers; it is the feedback loop that made the web usable. Publicly accessible reference material — encyclopedias, forums, documentation, code repositories, community Q&A — is the substrate of both search and modern AI. Those producers mostly do not have legal departments. If the rules make ingestion risky without a license, the rational move for a model developer is to license from large publishers and skip the long tail. The long tail then loses the traffic and tooling that justified its openness in the first place.
That is the real asymmetry in the Wikimedia brief: not that AI owes Wikipedia something, but that a rule written to protect the most commercially valuable text can quietly strip the commons of the tools built on it.
What this means if you ship AI features
The practical read-through does not wait for a verdict. Four moves are worth making now:
- Separate your training posture from your output posture. Document what you train or fine-tune on, and independently document how you constrain outputs — filters, refusal rules, memorization tests, retrieval boundaries. If case law splits the question, you want evidence for both halves.
- Treat verbatim output as a product defect class. Build regression suites that probe for regurgitation of high-risk sources (paywalled news, licensed reference works, code). Track it the way you track latency.
- Prefer retrieval plus citation over memorization. Architectures that quote and link are both legally safer and more auditable. They also degrade gracefully when a source changes.
- Map your jurisdictions. Fair-use doctrine is a US construct; other markets use different exceptions, and cross-border commerce pulls multiple regimes into one product. A storefront serving EU and US customers cannot assume a single answer.
Key takeaways
- Wikimedia's intervention is a request to split AI copyright into an ingestion question and an emission question, not to bless or ban AI training wholesale.
- Ingestion and emission have different economics: the first is about market entry and corpus access, the second about substitution and direct harm.
- A single binary rule produces either licensed-only consolidation or litigation-driven chilling — both concentrate capability in the largest players.
- The split is less predictable but more engineerable: output behavior is a product surface teams can actually control.
- For builders, the defensible posture is documented training provenance plus hard output-side controls, not a bet on one verdict.
Next steps
- Audit your AI stack for where copyrighted material enters (training, fine-tuning, retrieval, user uploads) and where it can leave (chat, exports, embeddings, logs).
- Add memorization and verbatim-overlap tests to your release checklist, prioritized by source risk.
- Review vendor contracts for indemnity scope and whether they cover training, output, or both.
- Draft an internal position paper on your training posture so procurement and legal answer the same way.
- For cross-border operations, maintain a jurisdiction matrix mapping each market's text-and-data-mining exceptions to your product surfaces.



