UK Watchdog Secures Data Protection Promises From the Ten Biggest AI Developers

Britain's data protection watchdog has secured promises of change from the ten companies behind the world's largest AI models. The Information Commissioner's Office, known as the ICO, published a report on 13 November 2026 confirming that the developers have made, or committed to make, improvements to how they handle personal data when training their systems.

The ten developers named in the report are Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. Each of them went through supervision by the ICO, and each has now agreed to fix weaknesses the regulator found. The ICO says it is monitoring their progress against those commitments.

Why the ICO looked at AI training

Foundation models sit behind chatbots, image generators, search features and a growing number of workplace tools across the UK. They are trained on vast datasets, and those datasets would be expected to include significant volumes of personal data. That makes the training stage a data protection issue, not just a technical one.

The ICO focused its supervision on five areas of UK data protection law: legitimate interests, special category data, transparency, the right of access and the right to object. In plain terms, it asked whether the companies had a proper legal reason to use people's data, whether they were open about it, and whether people could ask what was held about them or object to its use.

What the regulator found wrong

The report found that many of the developers' legitimate interests assessments, the documents that are supposed to justify using personal data, were too vague. Some did not identify a specific interest for each type of personal data at each stage of model development. Others lacked evidence that collecting the data was actually necessary, or failed to explain whether a less intrusive approach could have been taken.

Broad justifications such as improving AI products were not enough on their own, according to the ICO. Developers are now expected to make a robust assessment of whether processing is necessary, including an analysis of whether less intrusive options, such as synthetic data, could be used instead. They should also provide evidence that the safeguards they rely on actually work.

Transparency was another weak spot. The ICO found privacy information written in vague or imprecise language, including statements that training data may incidentally include personal data. Information about the specific categories of personal data used in training was often limited, and privacy details were spread across several different documents or locations, which made them hard to follow.

In many cases it was not clear whether the privacy information applied to users of a developer's services or to people whose personal data had been obtained from other sources. The ICO also raised concerns where developers took a blanket approach to refusing requests from people exercising their data rights, and where it was unclear how individuals could exercise those rights.

The changes the companies have agreed to make

In response, the developers have committed to changes. These include updating their legitimate interests assessments and privacy policies, and providing evidence of the efficacy of their training approaches. Several have published standalone privacy notices for model training, or updated existing notices to explain how and why personal data is processed during training.

The ICO also addressed special category data, the most sensitive kind of personal data, which covers things like health information, religious beliefs and political opinions. The regulator found that foundation model developers are likely to be processing special category data, whether they intended to or not, unless they can show their training datasets do not contain it. Where no legal condition permits that processing, the ICO recommends measures to prevent the collection of special category data or to filter it out.

Can an AI model itself hold personal data?

One of the most significant parts of the report is the ICO's position on whether AI models themselves can contain personal data. The regulator's view is that when training datasets include personal data, the resulting embeddings can still be personal data. In December 2024, the European Data Protection Board confirmed that AI models can embed their training data in ways that mean they are not anonymous, and that they therefore can be subject to data protection law.

Within the UK, the ICO says answering the question for a specific model requires a case-by-case assessment. To reduce the risk, organisations should test whether a model can single out an individual across records, and whether the same person can be identified by combining multiple records. The ICO is also monitoring the field of machine unlearning, the techniques for removing an individual's data from a trained model, to understand how mature those approaches are.

What happens next

The report sets standards that will matter well beyond the ten named developers. Any organisation that uses personal data to develop, fine-tune or adapt AI systems should expect the same expectations to be applied to it, particularly around lawful basis, transparency, data subject rights and special category data.

The ICO is preparing further guidance, including a statutory code of practice on AI and automated decision-making. At the same time, it has launched a call for evidence on agentic AI, systems that can take actions on a user's behalf, which raises fresh questions about who is responsible when an AI agent acts.

For people in the UK, the report is a reminder that the data used to build AI systems is regulated, and that the companies behind the biggest models must now prove they are handling it properly.

Frequently Asked Questions

Which AI developers did the ICO investigate?

The ICO supervised ten foundation model developers: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. Each of them has made, or committed to make, data protection improvements, and the ICO says it is monitoring their progress.

What was wrong with how AI companies handled personal data?

The ICO found vague legitimate interests assessments, limited evidence that collecting personal data was necessary, unclear privacy notices, information spread across several documents, and blanket refusals of data subject requests. Developers are now expected to justify their data use with specific, evidence-based assessments and clearer privacy information.

Can an AI model itself contain personal data?

According to the ICO, yes, it can. When training datasets include personal data, the resulting embeddings can still be personal data, and the European Data Protection Board confirmed in December 2024 that models can embed training data in ways that mean they are not anonymous. Whether a specific model contains personal data needs a case-by-case assessment.