Copyright, Creativity, and Generative Models

Generative AI has transformed creative production at a speed that legal frameworks were not built to accommodate. Models capable of generating text, images, music, and video have expanded the boundaries of what machines can create — while raising fundamental questions about intellectual property that existing copyright law cannot cleanly answer. The legal system's core concepts of authorship and originality were developed for human creativity, and they fit the outputs of generative AI awkwardly at best.

Two distinct but related questions are at the centre of ongoing legal and regulatory debate. First, does training AI models on copyrighted works constitute infringement — and if so, under what conditions? Second, who, if anyone, holds copyright over AI-generated output? Courts and regulators across multiple jurisdictions are working through these questions in parallel, producing an evolving patchwork of rulings and guidance with significant practical implications for organisations building and deploying AI systems.

Training Data and the Infringement Question

Training AI models requires large datasets that typically include substantial quantities of copyrighted material — books, articles, images, source code, musical recordings. The training process does not copy these works in the traditional sense; it uses them to adjust the parameters of a model that subsequently generates novel outputs. Whether this process constitutes infringement under copyright law is one of the most consequential unsettled questions in intellectual property.

In the United States, AI developers have largely relied on the fair use doctrine, arguing that training constitutes transformative use analogous to the way a human author reads and learns from existing works without permission. Several lawsuits filed by authors, artists, and news organisations challenge this interpretation, contending that the commercial exploitation of copyrighted material to build commercial AI products cannot fall within fair use regardless of how different the outputs are. UK and European law present different frameworks with different outcomes, and cases are working through various national court systems simultaneously.

Creative Output and Authorship

The question of who owns AI-generated output is legally distinct from the training data question, though the two are often conflated in public debate. Copyright law in most jurisdictions requires human creativity as a prerequisite for protection — purely machine-generated works receive no copyright protection in the United States or United Kingdom as the law currently stands. This creates a practical problem for organisations that use AI extensively in creative production: content generated without sufficient human creative input may be unprotectable.

In practice, the relevant question is often one of degree. If a human author provides detailed creative direction, selects from multiple AI outputs, and makes substantive edits, the resulting work likely meets the creativity threshold. If a user types a simple prompt and publishes the output verbatim, it likely does not. Organisations using AI in creative workflows should document their creative processes carefully, as evidence of meaningful human creative contribution may become legally significant as court decisions develop.

Ethical Dimensions

The legal questions sit within a broader ethical debate about the obligations that AI companies have towards the creators whose work has trained their systems. Artists, writers, and musicians whose work has been ingested by training datasets have received no compensation and, in most cases, no notice that their work was used. Many argue that this constitutes a form of extraction that undermines creative ecosystems — the very infrastructure that makes generative AI valuable was built on creative labour that the technology is now competing with.

Proposed remedies range from mandatory licensing frameworks — in which AI companies pay into funds that compensate rights holders collectively — to opt-out mechanisms that allow creators to exclude their work from future training datasets. The EU AI Act includes transparency obligations requiring disclosure of training data, which may provide a basis for licensing frameworks to develop. Industry practice is also evolving: some AI companies have negotiated licensing agreements with publishers and media organisations, suggesting that voluntary solutions are achievable where parties have incentives to reach them.

Industry Responses and Regulatory Development

Technology companies and creative industries are adapting to change at different paces and in different directions. Some AI developers have pursued proactive licensing arrangements with major content holders, seeking to establish legal clarity and build relationships with industries that might otherwise oppose them in regulatory forums. Others continue to rely on fair use defences, treating the litigation risk as manageable compared to the cost of licensing at the scale their training pipelines require.

Regulators are moving cautiously. Copyright reform is politically contested, with powerful interests on multiple sides, and legislators are reluctant to resolve through statute questions that courts are still working through case by case. The most likely medium-term outcome is continued jurisdictional divergence, with different courts and legislatures arriving at different answers to broadly similar questions.

Implications for Organisations

For organisations building on or deploying generative AI, the practical implications of the current legal environment are significant. The provenance and licensing status of training data is now a material risk factor; AI models trained on uncleared commercial content may face litigation exposure that affects their value as assets. Due diligence on training data sourcing is becoming a standard element of AI procurement and partnership assessment.

On the output side, organisations should establish clear policies for AI-assisted creative work that document human creative contribution, identify works where IP protection may be uncertain, and establish contractual clarity with commercial partners about the copyright status of AI-generated deliverables. The legal landscape will continue to shift as cases proceed through the courts and regulatory frameworks develop; organisations that engage with these questions proactively will be better positioned when clarity arrives.