UK copyright & AI training: Where the law stands after the opt-out rejection

No safe harbor for commercial AI training

The UK’s sole text and data mining (TDM) exception remains confined to non-commercial research purposes, and despite the recent government consultations, no broader safe harbor for commercial AI training has emerged. The Intellectual Property Office’s 2022 proposal for a broad copyright exception was abandoned in February 2023 following creative sector opposition, and the government’s 2024 consultation – which proposed a TDM exception with opt-out as its preferred option – was overwhelmingly rejected by respondents (81% favoring mandatory licensing). For more detailed analysis, see our client alert on text and data mining in the UK.

In March 2026, the government formally abandoned the opt-out and endorsed no alternative, acknowledging in its impact assessment that existing copyright law acts as a “significant constraint” on general-purpose model training in the UK and that rights holder permission would “usually be needed.” The practical position for businesses is clear: reproduction of copyright works for training purposes requires appropriate licensing under existing UK law. For more detailed analysis, see our client alert on the UK copyright and AI report.

Government confirms: License first, legislate later

The House of Lords Communications and Digital Committee set the tone in March 2026, urging the government to rule out any broad commercial TDM exception, impose statutory transparency obligations on AI developers, and introduce protections against unauthorized digital replicas – arguing it would be “a poor bet to sacrifice the UK’s outstanding creative capacity for speculative AI gains.”

In May 2026, the government’s response puts the legal position beyond doubt, stating that “AI developers must seek permission and acquire licences for the use of copyright works, unless they can show an existing copyright exception applies.” It adds that it “will not introduce reforms to copyright law unless or until it is confident that they will meet our objectives for the economy and for UK citizens,” reaffirming the position that licensing is the framework within which AI developers are expected to operate. The government acknowledges, however, that value from major deals may not be reaching individual creators – a gap it intends to monitor.

The response also sets out four areas of focus:

• Digital replicas consultation. Address harms from unauthorized replication of a person’s likeness.
• AI labeling taskforce. Establish a new taskforce to develop proposals on best practice for labeling AI-generated content (interim report due autumn 2026).
• Creator control review. Publish a review of mechanisms available for creators to control their works online, covering standards, technical solutions, and input transparency for AI model training.
• Working group on smaller creative organizations. Explore whether the government has a role in supporting independent and smaller rights holders to license their content effectively.

On transparency, the government agrees AI developers should disclose training sources but stops short of committing to statutory requirements, preferring to monitor international approaches – a notably cautious position, given that the EU has moved in the opposite direction. 

The government also announced the Sovereign AI unit, backed by £500 million in investment and R&D, alongside the Creative Content Exchange (CCE), which is due to begin its pilot scheme with an ambition to be commercially independent by 2030, signaling continued appetite for domestic AI capability even as copyright constraints tighten.

EU comparison

The contrast with the EU is instructive. The EU AI Act’s transparency obligations have been in force since August 6, 2026, requiring AI providers to disclose AI use and label AI-generated content, and deployers to inform individuals when they are exposed to emotion recognition, biometric categorization, or deepfakes. National authorities and the AI Office can also now act on GPAI obligations, which include training data summaries and establishing copyright compliance policies. Regulatory sandboxes are operational, giving smaller organizations room to experiment responsibly. For more detailed information, see our client alert on EU AI transparency rules.

What courts have (and haven't) decided about AI training

United Kingdom

In Getty Images v. Stability AI, the UK’s leading case on this topic, Getty alleged that Stability AI infringed its copyright by training its generative AI model, Stable Diffusion, on datasets containing millions of Getty’s images. The claims centered on three arguments: primary infringement through copying during training; primary infringement through AI-generated outputs; and secondary infringement via distribution of an “infringing copy.”

The training claim failed on territoriality: Getty could not satisfy the court that copies were made or hosted on UK servers. The output claim fell away after Stability blocked the relevant prompts and challenged Getty’s chain of title and originality to the underlying images. On secondary infringement, the court accepted that an “article” can be intangible but held that Stable Diffusion is not an “infringing copy” because it “does not store, reproduce, or contain any of the copyright works on which it was trained.”

The Getty case is due for appeal later in 2026, but the central question remains untested: whether training an AI model on copyright works in the UK constitutes primary infringement. For AI developers, the practical takeaways are to document and maintain clear records of where training occurs and to implement filtering and prompt-blocking measures that minimize the risk of infringing output. Until the courts provide clarity, the government’s licensing-first position remains the operative guide. For more detailed analysis, see our client alert on the Getty v. Stability AI case.

Germany

The court’s finding in Getty was fact-specific and turned on Stable Diffusion’s particular architecture – it does not preclude claims against models capable of memorizing or reproducing training data. That distinction proved decisive in the German courts.

In GEMA v. OpenAI, the German court reached the opposite conclusion to Getty, holding that where lyrics are “reproducibly contained” in a model, fixation within it constitutes reproduction. More recently, in GEMA v. Suno, the court delivered the first major European ruling that training an AI model on protected works without a license can infringe copyright, even where training occurs outside the EU. The court held that memorization constitutes reproduction, the provider (not the user) is liable, and AI Act compliance is no defense to copyright infringement. The judgment is not yet final, and the Court of Justice of the European Union’s (CJEU) ruling in Like Company (expected September 2026) may reshape the landscape further.

Practical implications

For clients engaged in AI development or content licensing, four points should inform current decision-making:

• Licensing is not optional. The government has confirmed it will not introduce a commercial TDM exception, and the licensing-first position is now unambiguous policy. If your training pipeline touches UK-hosted copyright works without a license, you are exposed under current law.
• Where training happens matters – but only up to a point. Getty v. Stability AI confirmed that UK copyright liability depends on acts occurring in the UK, and offshore training may fall outside the court’s reach under current law. GEMA v. Suno shows that other jurisdictions are willing to take a longer reach, and the CJEU ruling in Like Company (expected September 2026) may further shift the extraterritorial landscape. Maintain granular records of where training activities occur and on what datasets.
• Your model’s architecture affects your exposure. The core distinction emerging from the case law is between models that merely learn from content and those that memorize or can reproduce it. Models in the latter category face materially greater infringement risk under both UK and German authority. Implement prompt-filtering and output controls, and ensure your technical documentation clearly reflects how the model handles training data.
• The compliance landscape will move quickly before year end. The digital replicas consultation, the AI labeling taskforce interim report (due autumn 2026), the CCE pilot, and the Getty appeal could each shift the baseline. Build a monitoring trigger into your compliance calendar rather than treating the current position as settled.

The government is backing licensing and transparency over legislative reform. The key practical message: secure permission before you train, maintain clear records, and implement controls to manage risk.