WikiHow v OpenAI: When Does AI Training Become Copyright Infringement?

Most businesses using AI tools are not training models themselves, so the growing volume of copyright litigation against AI developers can appear to be someone else's problem. The commercially relevant question for a software business is whether the tools already built into its product rest on something a court will later find unlawful, and what the consequences would be if that happened. The wikiHow claim against OpenAI is a useful way into that question, provided it is read alongside the European position, which is reached by a different legal route and which has already been settled in part.

What Has wikiHow Alleged?

On 21st August 2026, wikiHow filed a complaint in the US District Court for the Southern District of New York against OpenAI and several affiliated entities, alleging copyright infringement and the removal of copyright management information under the Digital Millennium Copyright Act. The complaint asserts infringement of at least 1,200 registered copyrights and states that OpenAI's crawlers reached wikiHow's site more than 185,000 times between March and May 2025 and 148,529 times between May and July 2026.

The claim covers copying to build training datasets, retrieval of wikiHow material at runtime to answer user prompts, outputs that reproduce recognisable passages, and the alleged stripping of titles, bylines and terms of use from articles as they entered the datasets. That last element is worth noting, because claims for removal of copyright management information carry their own statutory damages and do not depend on establishing that the training use was itself infringing. wikiHow is seeking damages and an order restraining further infringement. OpenAI maintains that training on publicly available material is fair use.

Fair use does not exist in EU law. Article 3 of the Digital Single Market Directive allows text and data mining for scientific research by research organisations and cultural heritage institutions, with no opt-out available to rightsholders. Article 4 allows text and data mining for any purpose, including commercial AI training, on lawfully accessible content, unless the rightsholder has expressly reserved their rights in an appropriate machine-readable form.

In the United States the argument is largely about whether training is transformative and whether outputs substitute for the original in the market. In the EU the first questions are narrower and more factual, namely whether the content was lawfully accessible and whether a valid machine-readable reservation was in place at the time of collection.

Where Have European And UK Courts Reached?

The litigation is still developing, but the European and UK cases decided so far are useful because they separate out several questions. Whether material can lawfully be collected for training is not necessarily the same question as whether it can be retained within a model, reproduced in an output or relied upon in a particular jurisdiction. GEMA and Getty illustrate those distinctions particularly clearly. 

In GEMA v OpenAI (case 42 O 14139/24), the Regional Court of Munich I reached a different conclusion on 11th November 2025. The German collecting society demonstrated that ChatGPT could reproduce the lyrics of nine well-known German songs when prompted. The court treated the retention of those lyrics inside the model as a reproduction in its own right, and held that this took the activity beyond mining, on the basis that mining is aimed at drawing out general linguistic structure rather than holding on to the works themselves. The answers the model gave were separately treated as infringing, as reproductions, as adaptations and as making the works available to the public. Liability attached to OpenAI rather than to anyone typing the prompt, because OpenAI chose what went into the model and designed how it worked. The court ordered the company to stop, to account for how extensively the works had been used and what it had earned from them, and to compensate GEMA. OpenAI has announced an appeal, and a reference to the Court of Justice remains possible.

In Getty Images v Stability AI [2025] EWHC 2863 (Ch), the English High Court delivered judgment on 4th November 2025. Getty dropped its principal copyright claim part way through the trial, having been unable to show that the copying it complained of had happened on UK soil, since the training took place elsewhere. The secondary claim then failed as well, on the basis that a set of model weights is not an infringing copy where the works themselves are nowhere held within it.

What Getty was left with was a limited trade mark finding about watermarks surfacing in outputs, confined to earlier versions of the model. For a US business, the relevant lesson concerns territoriality, since where training physically took place can determine whether a claim is available at all, which is a separate question from whether the training was defensible on its merits.

Considered together, these cases show that AI copyright risk cannot be reduced to a single question about whether training itself was lawful. Courts are beginning to distinguish between the initial collection of material, what the model retains, what it can reproduce in response to prompts and where the relevant acts took place. GEMA suggests that even where text and data mining rules may apply at the dataset stage, memorisation and reproducible outputs can create separate infringement risks, while Getty shows that territoriality may determine whether a claim can be brought at all. For businesses, the practical implication is that due diligence should extend beyond training data provenance to model behaviour, output safeguards and the contractual allocation of risk.

What Should A Software Business Consider Doing?

Whatever the courts eventually decide, the AI Act imposes requirements on providers of general-purpose AI models now. Article 53(1)(c) requires a policy to comply with EU copyright law, including identifying and respecting reservations of rights made under Article 4. Article 53(1)(d) requires a publicly available summary of the content used to train the model, prepared using a mandatory template published by the AI Office in July 2025.

These obligations applied from 2nd August 2025 for models placed on the market after that date, with models already on the market having until 2nd August 2027 to comply. What changed on 2nd August 2026 was enforcement, as the AI Office can now verify compliance, require corrective measures and impose fines of up to 3% of worldwide annual turnover or €15 million, whichever is higher.

For businesses buying AI capability rather than building it, this is the most immediately useful development in the area. Training data provenance has moved from something a vendor could decline to discuss into something it is required to describe publicly. The template asks providers to set out where their material came from, broken down by type of source, what they did to it before training, and how they went about respecting copyright along the way. The level of detail in those published summaries varies considerably between providers, which is itself a useful signal during procurement.

There are three separate exposures here, and they are worth keeping apart because the appropriate responses differ.

Where a business trains or fine-tunes models on third-party content, the first question concerns acquisition rather than use. Lawful access is the gateway to Article 4, and material obtained through circumvention or from unauthorised sources sits outside the exception regardless of what happens to it afterwards. The second question, following GEMA, is whether the model can be made to reproduce recognisable portions of its training data, since that is where European courts are currently finding liability. Output filtering and memorisation testing have accordingly become risk controls rather than quality features.

Where a business only deploys another party's model, its exposure runs through its contracts. It is worth reading the intellectual property indemnity carefully rather than assuming that it provides cover, including whether it extends to outputs as well as the model, whether it is capped, whether it survives use of the tool inside a customer-facing product, what conditions must be met to preserve it, and what happens if an injunction is granted against the provider in a European jurisdiction.

The Laws on Copyright and AI

The copyright question around AI is becoming more precise. The debate is no longer simply whether training on copyrighted material is lawful or unlawful. Courts and regulators are beginning to distinguish between how content was obtained, what happened to it during training, what remains recoverable from the resulting model and what the model ultimately produces.

For software businesses, that distinction matters. A business integrating a third-party model may never touch the provider's training dataset, but it can still inherit commercial risk if the model reproduces protected material, if a court restricts its use, or if contractual protection turns out to be narrower than expected.

That does not mean businesses should avoid using generative AI. It means copyright provenance should now form part of ordinary AI procurement and product governance. The practical questions are increasingly identifiable: where did the training material come from, how were rights reservations respected, what testing has been carried out for memorisation, what safeguards apply to outputs, and who bears the risk if the legal position changes?

wikiHow v OpenAI may eventually give US courts another opportunity to define the limits of fair use in AI training. In Europe, however, the direction is already becoming clearer in that lawful access to material is only the starting point. What the model retains and what it can reproduce may prove just as important.

How Can Gerrish Legal Help?

Gerrish Legal is a dynamic digital law firm. We pride ourselves on giving high-quality and expert legal advice to our valued clients. We specialise in many aspects of digital law such as GDPR, data privacy, digital and technology law, commercial law, and intellectual property. 

We give companies the support they need to successfully and confidently run their businesses whilst complying with legal regulations without the burdens of keeping up with ever-changing digital requirements. 

We are here to help you, get in contact with us today for more information.