OpenAI told a UK parliamentary committee that leading AI systems could not be developed without using large amounts of copyrighted material. In written evidence, the company argued that modern copyright protection covers so much online expression that limiting training to clearly public-domain works would leave models unable to meet contemporary needs.
The submission places a central industry assumption into direct public debate. Developers want broad, recent and varied data, while creators and publishers want control over how protected work is used. OpenAI maintained that its practices comply with the law and pointed to potential partnerships and compensation, but it did not propose abandoning large-scale collection of online material.
The Argument Begins With the Breadth of Copyright
OpenAI said that news, forum discussions, images and many other kinds of online expression are protected. Because copyright is widespread and long-lasting, a training policy that excluded all protected material would remove much of the modern internet from consideration. The company argued that public-domain books and drawings created more than a century ago would not be enough to build systems relevant to today’s citizens.
This is partly a claim about capability and partly a claim about practicality. A model trained only on old material would miss current language, technology and cultural context. Yet the usefulness of copyrighted data does not by itself settle whether a particular use is lawful or fair. The submission explains why developers value broad access; courts and policymakers still have to weigh that interest against the rights of creators.
Training and Reproduction Raise Different Questions
AI developers often describe training as learning statistical patterns rather than storing a conventional library of copies. Rights holders focus on the fact that protected works are collected and processed at scale, sometimes without permission, and that generated output can compete with the people whose material contributed to the system. Both perspectives concern different stages of the same pipeline.
The source notes lawsuits alleging copyright breaches, including action from The New York Times. OpenAI appeared prepared to rely on broad interpretations of fair use rather than fundamentally change its training approach. The final legal boundaries were unresolved in the article. That uncertainty matters for customers as well as model providers because organisations need to understand the provenance, contractual protections and output risks of the systems they deploy.
Partnerships Could Create a More Explicit Exchange
OpenAI acknowledged that publisher partnerships and compensation mechanisms could help support creators. A licence can clarify which material is included, how it may be used and what the rights holder receives in return. It may also provide higher-quality data and metadata than indiscriminate collection, benefiting the model as well as the publisher.
Licensing cannot solve every issue on its own. The web contains millions of individual creators who cannot each negotiate a bespoke agreement, and one platform may not control every right in user-contributed material. A market dominated by large publishers could also leave independent creators with little leverage. Scalable permissions, meaningful opt-outs and transparent reporting would be needed if compensation is to extend beyond a small number of prominent deals.
“Impossible” Is a Policy Position, Not the End of Debate
OpenAI’s wording frames copyrighted training data as an unavoidable foundation for frontier AI. That claim puts pressure on lawmakers to permit broad use or risk slowing model development. Creators may reasonably answer that an industry’s preferred business model should not determine the scope of their rights. The political challenge is to decide what access supports innovation without making creative labour effectively free for any company able to train a large model.
More transparency would improve that debate. Model developers could describe the categories and sources of training material, the safeguards used to prevent memorised reproduction and the process for rights holders to raise concerns. Creators could then assess actual practices rather than infer them from model behaviour. Transparency does not resolve the legal question, but it makes accountability possible.
Technical choices also matter. Filtering, deduplication and output controls may reduce some forms of harm, while licensed or purpose-built datasets can replace parts of a broad crawl. The source does not suggest that these measures eliminate the need for protected works, but they can affect how responsibly those works are handled.
Clear records would also help distinguish training disputes from cases where a system reproduces a recognisable passage. Those situations may involve different evidence and remedies. Without documentation of data sources, filtering and model behaviour, every disagreement risks becoming an argument over assumptions rather than a review of what actually occurred.
The committee submission highlights a structural conflict. Frontier models gain value from absorbing an enormous range of human expression, yet much of that expression exists because creators and publishers can control and finance their work. Weakening either side can damage the ecosystem: overly narrow data access may limit useful models, while uncompensated use can undermine the production of future material.
OpenAI’s position is clear that it will not voluntarily restrict training to the public domain. That makes legal judgments, legislation and commercial agreements the likely arenas for deciding the balance. Companies using AI should follow those developments and avoid treating a provider’s assurance of compliance as the final word.
The sustainable outcome will need more precision than “all protected data is off-limits” or “all training is permitted.” It must distinguish collection, learning, reproduction and market impact while creating rules that can operate at internet scale. OpenAI’s testimony did not close the issue; it made the stakes harder to ignore.
