Where AI meets the law: a map

An AI product passes through stages: data is collected, a model is trained on it, a system is built around the model, the system produces output, and people use the output. The model is a large file of numbers that training adjusts, example after example, until its answers improve, and those numbers hold everything the model learned. The system is the product built around the model, such as a chat assistant, which adds the screen a user types into, safeguards and connections to other software. Different law attaches at each stage, so this module takes the legal questions in the order of the stages, from copyright in the training data to the professional duty of a lawyer who relies on the output. Each question is stated as it stood when this module was last reviewed, with the primary source it rests on and the Advanced module that treats it at length. Several of the questions are open, and where courts or regulators have taken a view, the view is attributed to them.

Copyright in training and output

Training a model involves copying works, often in very large numbers, into a training collection, so copyright law asks whether a licence or an exception permits those copies. In the European Union, Article 4 of Directive (EU) 2019/790, the Digital Single Market Directive, allows copies made for text and data mining, the use of software to analyse text and data in digital form for patterns, trends and correlations. The exception applies on condition that the rightholder has not expressly reserved the use “in an appropriate manner, such as machine-readable means in the case of content made publicly available online”, so a rightholder can opt out of the exception. Machine-readable means are instructions written for software to read, such as a notice in a website’s files telling automated collection tools not to copy its pages. European courts have begun to consider what counts as a valid reservation, and how far Article 4 of the Digital Single Market Directive covers the training of generative models, meaning models that produce new text, images or music.

Models built to perform a wide range of tasks, such as the models behind chat assistants, are known in the EU’s AI Act as general-purpose AI models, and the companies that develop and supply them are their providers. The AI Act adds, in Article 53(1)(c), a duty on providers of general-purpose AI models to put in place a policy to comply with Union copyright law, and to identify and comply with reservations made under Article 4.

In the United Kingdom, section 29A of the Copyright, Designs and Patents Act 1988 permits copies made for computational analysis, the UK statute’s phrase for text and data mining, only “for the sole purpose of research for a non-commercial purpose”. The government consulted on a wider exception with an opt-out, and its report of 18 March 2026 said that “a broad copyright exception with opt-out is no longer the government’s preferred way forward”, while proposing further evidence-gathering.

Whether the trained model is itself a copy of the works it learned from is a separate question, which bears on whether holding or supplying the model, after training has finished, can infringe copyright. The numbers a model holds are called its weights or parameters, and the works the model learned from are not kept inside them as text or pictures. Even so, a model can sometimes reproduce, word for word, a passage it saw often in training, which is called memorisation and is the reason to ask whether the weights hold a copy. Courts in London and Munich came to different results at first instance, on different evidence. In Getty Images v Stability AI the High Court found that the model weights were not an infringing copy. In GEMA v OpenAI the Munich Regional Court held that memorisation embodied song lyrics in the models’ parameters, and that embodying them there was a reproduction under German and EU copyright law. The High Court has since given permission to appeal in Getty, and the Munich court’s press release of 11 November 2025 says that its judgment in GEMA is not final.

Output raises its own question, which is whether text or images produced with little human input are protected by copyright at all, and who owns them. Section 9(3) of the Copyright, Designs and Patents Act 1988 treats the author of a computer-generated work as “the person by whom the arrangements necessary for the creation of the work are undertaken”, so UK law can protect a work that a computer generated without a human author. The government’s report of March 2026 proposes removing that protection for wholly computer-generated works while keeping it for works made with AI assistance. Copyright is treated in detail in “Training data and copyright”, an Advanced module.

Personal data in models

Training data gathered from the web usually includes personal data, such as names, photographs and posts, so the GDPR in the European Union and the UK GDPR in the United Kingdom apply to collecting it and training on it. A developer therefore needs a lawful basis for that processing, meaning one of the grounds on which data protection law permits it. One of those grounds is legitimate interests, which calls for a developer’s interests to be weighed against the rights of the people the data is about. The European Data Protection Board, which brings together the EU’s national data protection regulators, known in the GDPR as supervisory authorities, examined the use of legitimate interests for training in Opinion 28/2024.

Whether a trained model itself contains personal data is less clear, since a model holds numbers and the text it learned from is not stored in it as such, although memorisation can lead a model to reproduce passages it saw in training, details about people among them. The Board took the view in the same opinion that not every model trained on personal data can be considered anonymous, and that supervisory authorities should assess a claim that a model is anonymous case by case. Personal data might come out of a model by extraction, meaning technical attacks that draw training data out of the model, or in answer to queries, meaning questions or instructions that a user puts to the model. In the Board’s view, a model is anonymous only where the likelihood of extracting personal data from it, or of obtaining such data through queries, is insignificant.

The Board also took the view that the GDPR does not apply to a model that is anonymous. The rights the GDPR and the UK GDPR give individuals, such as the rights to have personal data corrected or erased, therefore depend in part on whether a model is anonymous. Correcting or erasing personal data is straightforward in a database, where each person’s record sits in one place that can be found and changed. In a model’s weights, what the model learned about one person is spread across many of the same numbers that hold everything else it learned, so removing it may mean retraining the model. The GDPR and the UK GDPR are treated at length in “Personal data in models”, an Advanced module.

The AI Act

The EU AI Act, Regulation (EU) 2024/1689, regulates AI by the risk a use poses, so most of its duties attach at the system and use stages, with separate duties for general-purpose AI models at the model stage. The AI Act prohibits some practices outright, imposes detailed requirements on high-risk systems, such as those used in recruitment or credit scoring, sets transparency duties for some others, and adds obligations for providers of general-purpose AI models. In the AI Act’s terms, a provider is, broadly, a business that develops an AI system or model and puts it on the market under its own name, and a deployer is a person or organisation that uses an AI system in the course of its work. A bank that buys a credit-scoring system from a provider and uses it to assess loan applications is therefore a deployer. Under Article 2(1) the AI Act applies to providers that place AI systems or general-purpose AI models on the EU market, wherever the providers are established, and to deployers of AI systems where the deployer is established or located in the EU, among others.

The AI Act’s obligations apply in stages under Article 113, as amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744, which changed some of the dates. The prohibitions have applied since 2 February 2025, apart from some that the Digital Omnibus added, which apply from 2 December 2026. The obligations for general-purpose AI models have applied since 2 August 2025, and most of the rest since 2 August 2026. The requirements for high-risk systems apply from 2 December 2027 for the uses listed in Annex III, such as recruitment and credit scoring, and from 2 August 2028 for products covered by Annex I, the AI Act’s list of EU product-safety laws. An AI system can come within Annex I by forming part of a product such as a medical device, since the EU’s law on medical devices is among the laws that Annex I lists.

Open questions under the AI Act include where the line falls between a high-risk use and an ordinary one, and how the duties of a provider and a deployer divide when a business builds its own product on another company’s model. That division may be hard to draw for a software company that builds a contract-review tool on a general-purpose model from another developer and sells the tool to law firms. The developer has duties as provider of the model, the software company may have duties as provider of the tool, and the tool’s behaviour depends partly on a model the software company did not build.

The United Kingdom has followed a different course, and it had not passed a statute comparable to the AI Act when this module was last reviewed. The UK government’s response of February 2024 to the consultation on its white paper on AI regulation set out an approach under which existing regulators apply five cross-sectoral principles, among them fairness and transparency, within their remits. Both the AI Act and the UK’s approach are treated at length in “Regulating the technology”, an Advanced module.

Liability and professional duty

Who answers for harm caused by an AI system’s output, whether the developer of the model, the business that built the product or the person who relied on it, is largely an open question. In the European Union, the new Product Liability Directive, Directive (EU) 2024/2853, sets common rules on liability for damage caused by defective products. The Directive defines a product in Article 4(1) to include software, so AI systems supplied as software appear to fall within it. EU member states have until 9 December 2026 to bring their laws into line with the Directive under Article 22. Under Article 2(1) the Directive covers products placed on the market or put into service after that date.

For lawyers in England and Wales, the Divisional Court of the High Court has stated one part of the answer to the question of responsibility for AI output. In R (Ayinde) v London Borough of Haringey, the court said that freely available generative AI tools are not able to conduct reliable legal research. The court went on to say that lawyers who use AI for legal research have “a professional duty therefore to check the accuracy of such research by reference to authoritative sources” before using it in their work, and the case is worked through in Does AI reason?, a Beginner module. AI agents, the systems that take actions such as placing orders on a person’s behalf, raise a further question, which is whether a contract an agent makes binds that person. Liability for AI output, contracts made by AI agents and the professional duties of lawyers are covered in “Liability, agents and professional duty”, an Advanced module.