On October 30, 2026, the digital rights advocacy group Electronic Frontier Foundation (EFF) filed an amicus brief in the In re OpenAI case, defending fair use in the training of generative AI models. The document's authors, Corynne McSherry and Tori Noble, warn: if the court narrows fair use under pressure from publishers and authors, the restrictions will affect not only AI developers but also journalists, researchers, and ordinary users who daily copy and transform data on the internet. The outcome of the case will determine whether training models on copyrighted material remains legal without mandatory licenses.

image
image

What happened

EFF filed a third-party written statement—an amicus brief—in the In re OpenAI case in support of a broad interpretation of fair use in the training of generative models. The document's authors, Corynne McSherry and Tori Noble, point to legal precedents that have already recognized model training as 'transformative' use. In support, EFF cites research, including papers on arXiv, challenging the thesis of AI as a 'violation machine' for copyright. A separate section of the text criticizes the idea of mandatory licenses for training data: EFF quotes Professor Lawrence Lessig, who believes such a structure would leave the ability to train models only to a 'handful of tech giants.'

Context

Fair use is a doctrine of American copyright law that permits the copying and transformation of protected materials without the rights holder's permission when the use is sufficiently transformative. This exception underpins journalists' work with others' materials, researchers' analysis of corpora, and ordinary users' everyday online actions. The In re OpenAI case takes the dispute beyond a single lawsuit: in reality, it resolves the question of who will capture the model training market if entry begins to require licenses. The requirement to prove 'lawful acquisition' of every fragment of training data falls asymmetrically on participants: industrial labs have lawyers and budgets for licensing agreements, while university labs and open-source projects do not. Therefore, narrowing fair use is not a dispute about model quality, but a question of who will be allowed to train them at all.

Why this matters for the industry

There are no direct changes today: an amicus brief is the position of one participant in the discussion, not a court ruling, and it does not affect APIs, prices, or model latencies. However, the outcome of the case will determine access conditions to training data. If courts begin requiring proof of 'lawful acquisition' of materials, small teams and academic projects will feel the barrier first, and web scraping methods used by journalists and NGOs for their investigations will become legally risky. Mandatory data licenses, according to the argument presented by EFF, will lead to the consolidation of pretraining among large companies capable of paying for agreements. A practical step available to engineering teams now and almost for free is auditing data sources in pipelines and logging data provenance: if the scenario of proving lawful acquisition is realized, a recorded history of sources and licenses will serve as insurance. On the horizon of six months and two years (an interpretation, not a guarantee), two forks are likely: either the emergence of a category of compliance-ready products—data rights registries, opt-out tools, and training corpus cleaning services—or, if fair use is preserved, growth in the niche of open datasets and fine-tunes with a low barrier to entry for small teams.

Why this matters for users

The dispute concerns not only AI companies: plaintiffs propose recognizing routine daily internet actions—opening emails, clicking links, streaming—as potentially infringing on copyright. If fair use is narrowed, legal risk will begin to cascade down the stack—from training base models to specific applications used by ordinary people. The availability of applied services like Be My AI, which uses GPT to describe images for blind users, as well as research projects dependent on open data, may be at risk. It is precisely these everyday practices of copying and transforming information that EFF asks the court not to put in question.

What is currently unknown / limitations

The process is in an early legal stage: an amicus brief has been filed by one interested party, and the court has not yet issued any interim or final rulings, so neither the narrowing of fair use nor mandatory licenses is currently a legal reality. The specific studies cited by EFF, including the mentioned arXiv papers, are not named in the available text of the material, so the thesis about refuting the 'violation machine' cannot be verified against primary sources. The thesis about the non-linear growth of model training costs when fair use is narrowed should be calibrated as a scenario hypothesis with explicitly reduced confidence: this is an interpretation, not a measured effect, and cannot be presented as a fact. Development scenarios on the horizon of six months and two years are interpolations of possible outcomes, not forecasts.

Sources

Author

Look at AI, editorial team