Privacy & Cybersecurity Law Blog

The UK Information Commissioner’s Office (“ICO”) has published a report following its scrutiny of ten of the world’s largest artificial intelligence (“AI”) foundation model developers operating in the UK. The report sets out the ICO’s regulatory positions on how UK data protection law applies to AI foundation model development and signals certain areas where the ICO expects stronger compliance measures, as detailed below.

Transparency

AI foundation models are often trained on large and diverse datasets that include personal data. Such personal data may have been collected directly from users, obtained from third parties or scraped from public sources. In the ICO’s view, developers must do more to explain what data they use, where it comes from, how it is processed and how data subjects can object or otherwise exercise their rights. For example, with respect to data obtained from third parties, the ICO states that a standalone privacy notice for third parties is a sufficient method of providing information on how data subjects’ personal data is being processed.

Lawful Basis and Data Subject Rights

The report also makes clear that developers cannot rely on the technical complexity of the relevant system or model to justify non-compliance with data protection obligations. Particularly, the ICO states that organizations training AI foundation models must identify and document a valid lawful basis for processing personal data and put in place effective mechanisms to enable data subjects to exercise their rights, including rights of access, erasure, restriction and objection. Where the lawful basis relied on is legitimate interests, the report makes clear that developers should carry out a legitimate interests assessment to demonstrate that this basis has been properly established.

Models Containing Personal Data

The ICO indicates that a model which processes personal data may fall within the scope of UK data protection law where personal data from training datasets is embedded in the model in a way that could permit singling out, linkability or extraction. In this context, “singling out” refers to whether the model can be used to distinguish one person across records or isolate records relating to that person from a dataset, and “linkability” refers to whether the same person or a group of people can be identified by combining multiple records or pieces of information. The ICO indicates that this assessment must be made case by case, taking into account the technical characteristics of the model and the safeguards used to reduce re-identification and disclosure risks.

Special Category Data

The ICO recognizes that special category data may be embedded in training datasets in ways that make it difficult to identify and manage. These difficulties do not, however, reduce developers’ obligations. Processing special category data requires both a valid lawful basis under Article 6 of the UK General Data Protection Regulation and a separate condition under Article 9. Depending on the circumstances, potentially relevant Article 9 conditions may include data manifestly made public by the individual under Article 9(2)(e) or the scientific research condition under Article 9(2)(j). The ICO cautions, however, that these conditions are unlikely to cover the processing of all special category data within a training dataset.

Alongside the report, the ICO launched a six-week call for evidence on agentic AI, citing growing concerns about systems that can complete tasks, use tools and interact with external systems with limited human oversight. The ICO indicated that evidence gathered through that process will inform future guidance and its forthcoming statutory code of practice on AI and automated decision-making. The deadline to respond to the call for evidence is November 20, 2026.