AI, Copyright and the Creative Industries: Where Are We Now?

The UK has not created a new copyright exception for commercial AI training. After sustained opposition from the creative industries, the Government has stepped back from its proposed opt-out model, but questions around licensing, transparency and performer protection remain unresolved.

AI, Copyright and the Creative Industries: Where Are We Now?

For the creative industries, the most important fact about the UK’s debate on AI and copyright is also the easiest to lose sight of: the law has not been rewritten.

There is currently no general UK copyright exception allowing generative AI developers to use protected works for commercial training. Nor is there a statutory opt-out regime under which works are assumed to be available for training unless their owners reserve their rights. The Government proposed such a model in late 2024, but it has not become law and, in March 2026, the Government expressly said it was no longer its preferred approach.

That leaves businesses operating in a commercially important but still unsettled landscape.

What does the law currently say about AI training?

The starting point remains the Copyright, Designs and Patents Act 1988. Copyright owners control specified acts including copying, and the Government’s own March 2026 report confirms that, although exceptions apply in particular circumstances, permission is generally required before a restricted act is carried out.

The UK does have a specific text and data mining exception under section 29A, but it is limited to research for a non-commercial purpose, subject to conditions including lawful access. There is also a narrow exception for temporary copies made as an integral part of a technological process. There is no broad statutory exception for commercial AI training.

There is, however, an important qualification. No UK court has yet finally decided the central question of whether a particular act of training a generative AI model on protected works without a licence infringes the reproduction right. That remains a matter for the courts on the facts of an appropriate case.

So it would be too simplistic to say that every disputed act of AI training has already been judicially declared unlawful. But equally, AI developers cannot point to a general commercial AI-training safe harbour in UK copyright law.

How did the proposed opt-out regime unravel?

The Government’s December 2024 consultation proposed four broad approaches. Its preferred model at that stage was a new data-mining exception covering commercial use, provided the developer had lawful access and the rights holder had not expressly reserved its rights. The proposal was intended to be accompanied by technical rights-reservation systems and greater transparency about training data.

The creative industries objected strongly. One fundamental concern was that the proposal would reverse the usual practical operation of copyright: instead of the user seeking permission, the burden would move to creators and rights owners to prevent use.

The response was striking. The Government subsequently reported that only 3% of consultation respondents supported the opt-out option, with opposition particularly strong among individual creators, performers and creative-industry organisations.

The Data (Use and Access) Act 2025 did not introduce the proposed exception or a mandatory training-data transparency regime. Instead, sections 135–137 required the Government to produce an economic impact assessment, a substantive report on copyright and AI, and an interim progress statement.

When the report arrived on 18 March 2026, the Government confirmed that the broad exception with opt-out was “no longer” its preferred way forward. It now says it has no preferred option and will gather further evidence and consider alternative approaches before changing copyright law. Its impact assessment states that copyright reform will not be introduced until the Government is confident it meets its objectives.

That is a significant change of direction. It is not, however, the same as permanently ruling out all future reform.

Licensing is becoming the real commercial question

If copyright remains permission-based, the next issue is how licensing can operate at the extraordinary scale required by AI.

A market is beginning to develop. The Government has identified direct deals between AI businesses and owners of valuable datasets and content libraries, while collective licensing models are also being explored. Its current position is that licensing should remain principally market-led rather than being replaced by a statutory licensing scheme or levy.

For rights holders, the commercial principle is straightforward: valuable content used to improve commercial AI systems may itself have licensing value. For developers, the difficulty is equally real. Large models may involve enormous numbers of works, fragmented ownership, incomplete metadata and potentially prohibitive transaction costs.

This is why transparency matters so much. A licensing right is considerably less useful if a rights holder cannot discover whether its content has been used.

More than 90% of respondents to the Government consultation supported disclosure by developers of the sources of training material. Yet, as of September 2026, UK law does not generally require AI developers to publish information identifying the copyright works used to develop their models or how they were obtained. The Government is currently favouring work on best practice and technical standards while keeping possible future legislation under review.

The House of Lords Communications and Digital Committee has argued for a considerably stronger position: a licensing-first approach backed by mandatory statutory transparency. Those are influential recommendations, but they are not themselves law, and the Government has so far stopped short of accepting the Committee’s proposed statutory framework.

Getty v Stability AI did not settle the training question

The High Court’s November 2025 decision in Getty Images v Stability AI was widely anticipated as the first major UK test of generative AI and copyright.

But it did not produce the definitive ruling on training that many expected.

Getty abandoned its primary training claim after accepting that it could not establish that the relevant Stable Diffusion training and development had taken place in the UK. The Court therefore did not decide whether that training infringed UK copyright. It rejected a separate argument that the Stable Diffusion model made available in the UK was itself an “infringing copy”, although Getty was subsequently given permission to appeal that aspect of the decision.

The case exposes a practical problem which remains highly relevant: copyright is territorial, while AI development is global. Establishing what was copied, where it was copied and by whom may be as important as the underlying legal principle.

Performers and digital replicas

For performers, AI raises additional questions beyond copyright in scripts, recordings or footage.

Existing performers’ rights, copyright, contract, data protection, passing off and other causes of action can offer protection in particular circumstances. But the Government itself accepts that existing UK law does not cover every situation in which a person’s voice or likeness is replicated without permission.

The March 2026 report therefore proposed further examination of a new digital replica or personality right. Government subsequently announced a consultation on digital replicas, alongside work on AI-generated-content labelling and creator-control mechanisms. These remain policy developments: no general UK statutory personality or digital-replica right has yet been enacted.

For producers and talent representatives, that makes contractual protection particularly important now rather than something to revisit once legislation catches up.

What should media businesses and rights holders be thinking about now?

The absence of a final Government settlement is not a reason to wait.

Review rights grants carefully. Commissioning, contributor, talent, archive and acquisition agreements should address AI uses expressly where appropriate. Training, fine-tuning, synthetic performance, voice or likeness replication and downstream model use should not simply disappear inside vague language about “all media”.

Separate different AI rights. Permission to use a work in a programme is not necessarily the same commercial proposition as permission to place it in a training dataset, fine-tune a specialist model on it or create a reusable digital replica of a performer.

Protect confidential material. Scripts, treatments, formats, rushes, unreleased footage and sensitive talent information should not be uploaded to external generative AI systems without understanding the provider’s terms, data-retention arrangements and whether inputs may be used to improve or train its services.

Strengthen provenance and record-keeping. Producers should know where important assets originated, what permissions were obtained and where AI has materially contributed to development or production. That will increasingly matter for chain of title, distributors, commissioners, insurers and international exploitation.

Audit AI supplier terms. Questions about intellectual-property indemnities, training on customer inputs, confidentiality, deletion, output rights and responsibility for infringement are becoming part of ordinary media procurement.

Consider licensing strategically. Well-organised catalogues, archives, formats and specialist datasets may have value in the emerging AI licensing market. Good metadata and clear ownership can therefore be commercial assets, not merely compliance exercises.

The policy debate will continue. But the practical position today is clearer than some of the headlines suggest. Copyright has not been suspended while Government decides what to do about AI. Existing rights remain relevant, licensing is becoming increasingly important, and contracts are already doing much of the work that future legislation may or may not eventually address.

For media businesses, the sensible approach is therefore not to predict where AI law will end up, but to know what rights they own, what rights they are granting and acquiring, and exactly what happens when valuable creative material enters an AI system.