ICO secures data protection improvements from leading AI developers: What this means

The Information Commission's Office (ICO) has published a significant new report confirming that ten of the world's largest foundation model developers have made, or committed to make, data protection improvements following regulatory supervision by the ICO. The report also sets out the ICO's broader expectations for model development, making the report of interest to developers beyond the ten named.

This comes at the same time as the ICO has launched a call for evidence on agentic AI. We explore this update further in our recently published "The ICO's agentic AI call for evidence: what you need to know" article.

Why this matters

The ICO identified the importance of foundation models, which can be adapted for a vast range of different purposes, given their widespread usage in AI systems operated by organisations across the UK. These are trained on vast datasets that would be expected to include significant volumes of personal data.

The ten developers scrutinised by the ICO were Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. Following engagement with the ICO, each of these organisations has made or committed to make improvements. The ICO has confirmed that it is monitoring developers' progress against their commitments.

However, the standards set out in the report will be of note to any organisation that uses personal data to develop, fine-tune or adapt AI systems, particularly when it comes to the specific areas identified in the report: lawful basis, transparency, data subject rights, and special category data.

What did the ICO identify?

Based on information provided by model developers and the ICO's wider assessment of regulatory uncertainty, the ICO focused its supervision on:

• Legitimate interests (Article 6(1)(f) UK GDPR);
• Special category data (Article 9 UK GDPR);
• Transparency (Articles 13 and 14 UK GDPR);
• Right of access (Article 15 UK GDPR);
• Right to object (Article 21 UK GDPR).

From this, the ICO secured two broad categories of improvement, as follows:

The ICO also found that foundation model developers are likely to be processing special category data, whether intentionally or not, unless they can demonstrate that the datasets used for training do not contain special category data.

Where special category data is used for model development, an Article 9 condition for processing should be identified. However, the ICO accepted that there are only limited circumstances in which an Article 9 condition will permit processing, and has recommended that where no condition applies, developers should implement measures to prevent the collection of special category data or filter it out.

Models as personal data

One of the most significant aspects of the report is the ICO's position as to whether AI models themselves can contain personal data. The ICO's position is that when training datasets include personal data, the resulting embeddings can still be personal data.

In December 2024, the European Data Protection Board (EDPB) 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's view is that answering the question of whether a specific model contains personal data requires a case-by-case assessment of the model, the detail of which is set out further in the report. To mitigate the risk that a model contains personal data, organisations should test whether the model can single out an individual across records, and whether the same person (or group of people) can be identified by combining multiple records.

This has a significant impact in a commercial context, as considerations around model hosting and transparency become important and complex where data protection law applies.

As an interesting point of reference, the ICO also stated that they are monitoring the complex field of "machine unlearning", to understand the maturity of relevant approaches to removing individuals' data from models.

What are the key takeaways?

There are several areas worth reviewing in light of this report, if your organisation is using personal data to fine-tune or develop AI models:

1. LIAs: If your organisation is using personal data to fine-tune or develop AI models, review whether your LIA is sufficiently specific, evidence-based and robust. You should ensure that broad justifications are backed up with evidence. You should also consider whether less intrusive alternatives are available (e.g. use of synthetic data).
2. Transparency: Review your privacy notices to ensure they clearly explain, in accessible terms where possible, how personal data is being used in AI training. Ensure that the information is clearly drafted, and is not buried across multiple documents or interfaces. If you are using data scraped from the web, make it clear which transparency information applies to this data as opposed to data received directly from users.
3. Data usage: If you use prompts or uploaded data to train AI models, ensure this is clearly communicated to users, that it is compatible with the purpose for which the data was originally collected, and that available options to opt out of model training are communicated to users.
4. Data subject rights handling: Provide clear routes for submitting requests, information about any technical limitations that may affect the outcome of a request, and clearly justify any decision not to fulfil a request.
5. Special category data: Do not assume that because you did not intend to collect special category data, that you are not processing it. Technical mitigation measures (e.g. filtering) can help to mitigate this risk.

We expect the standards applied here to be extended to any organisation that uses personal data to train, fine-tune or deploy AI. In particular, we can expect to see these standards further clarified through upcoming ICO developments such as the statutory code of practice on AI and automated decision-making (ADM), as well as any guidance on agentic AI arising from the recently published call for evidence.

Developers should therefore ensure that their governance frameworks remain sufficiently robust and are adapted in response to guidance produced by the ICO, so that they are well placed to respond to these developments as they occur.  At TLT, we are advising clients across multiple sectors on AI governance frameworks and data protection compliance in light of regulatory developments. If you would like to discuss or review your existing AI arrangements, please do get in touch.

This publication is intended for general guidance and represents our understanding of the relevant law and practice as at October 2026. Specific advice should be sought for specific cases. For more information see our terms & conditions.