ICO consults on draft guidance on anonymisation and pseudonymisation for research

On 24 August 2026, the Information Commissioner’s Office (ICO) launched a consultation on draft guidance on anonymisation and pseudonymisation for research, archiving and statistical purposes (“RAS”). The consultation remains open until 19 October 2026.

The draft guidance will be of particular interest to universities, healthcare and life sciences organisations, public sector bodies and technology providers that rely on data-driven research and innovation.

Whilst much of the draft guidance consolidates existing ICO guidance on anonymisation and pseudonymisation, it also provides insight into how the ICO expects organisations to apply these principles in the context of modern research practices, including AI development, trusted research environments (TREs), trusted third parties (TTPs), federated analysis models and synthetic data. The guidance also sheds light on how the ICO expects organisations to approach the new safeguards for RAS purposes introduced by the Data (Use and Access) Act 2025 (“DUAA”).

Below we highlight the key takeaways that organisations involved in research are likely to find particularly relevant.

Organisations must apply safeguards for RAS processing

The DUAA introduced new safeguards for processing personal data for RAS purposes, requiring organisations to implement appropriate technical and organisational measures to support data minimisation.

Whilst the ICO’s broader guidance on the RAS provisions remains under review following a separate consultation, this draft guidance provides additional insight into how the safeguards may be applied in practice. In particular, the ICO emphasises the importance of assessing identifiability throughout the lifecycle of a research project, including whether anonymisation or pseudonymisation can be used at different stages, and what technical and organisational measures may be required to manage re-identification risks.

Whilst these concepts are not new, the guidance suggests that organisations should approach anonymisation, pseudonymisation and data minimisation as part of a broader governance framework rather than as isolated technical exercises. For organisations relying on the RAS provisions, the practical focus is therefore likely to be on documenting how the safeguards have been considered and implemented throughout the research lifecycle.

The ICO highlights the Five Safes framework (safe people, safe projects, safe settings, safe data and safe outputs) as an example of a governance framework that may help organisations demonstrate that appropriate safeguards are in place.

TREs, TTPs and synthetic data can reduce privacy risks, but they are not anonymisation techniques

The draft guidance contains detailed commentary on TREs, TTPs and federated access approaches, reflecting the growing role these models play in research.

In simple terms, these tools are designed to make research safer by limiting who can access data and how it can be used. TREs, for example, allow researchers to analyse data in a secure environment without receiving a copy of the underlying dataset, whilst a TTP can help separate identifiers from research data or link datasets on behalf of researchers.

However, the ICO is clear that these arrangements do not automatically make data anonymous. Rather, they can help manage identifiability and reduce privacy risks, but organisations must still assess whether individuals remain identifiable in the circumstances.

The same principle applies to synthetic data. The ICO recognises its growing use in research, testing and AI development, but emphasises that synthetic data generated from personal data is not automatically anonymous. Organisations must still assess identifiability and re-identification risks, including whether the synthetic dataset could enable individuals to be identified directly or indirectly. The ICO also notes that organisations should consider the appropriate release model when making synthetic data available, as identifiability risks may vary depending on how the dataset is accessed and by whom.

The practical takeaway is that organisations should view TREs, TTPs and federated access models as tools that can support compliance and reduce privacy risks, but they do not automatically render data anonymous.

Take particular care when working with unstructured or higher-risk datasets

The draft guidance highlights the challenges of applying data minimisation principles to certain types of data, including free-text data, speech data, images, video recordings and sensor data. The guidance notes that organisations should take particular care when working with unstructured or higher-risk datasets, where personal data may be embedded within the content itself rather than contained in obvious identifiers such as names or email addresses.

The examples provided by the ICO illustrate the limitations of traditional de-identification approaches. For textual data, the content itself may identify an individual even where direct identifiers have been removed. For speech data, both the information being communicated and the characteristics of the speaker's voice may be relevant. Images, video and sensor data can similarly contain distinctive features, behavioural patterns, location information or other attributes that increase identifiability risks.

This is particularly relevant in the context of AI development, where models are often trained or tested using large volumes of unstructured data. The guidance reflects the reality that datasets commonly used for AI, such as clinical notes, call recordings, images and behavioural data, may contain personal data in less obvious ways, meaning that removing direct identifiers may be insufficient.

The broader message is that organisations should consider data minimisation at the design stage and throughout the research lifecycle, rather than treating it as a final redaction exercise.

What should organisations do now?

Although the guidance remains in draft form, it provides a useful indication of how the ICO expects organisations to approach identifiability, data minimisation and governance for RAS processing.

Organisations relying on the RAS provisions should be able to demonstrate how they have considered identifiability throughout the research lifecycle, including whether research objectives could be achieved using anonymised or pseudonymised information and what safeguards are needed to support data minimisation. Where directly identifiable information is used, organisations should be able to explain why that is necessary and proportionate for the research purpose and why anonymised or pseudonymised information would not achieve that purpose.

In practice, this may involve reviewing existing DPIAs, data-sharing arrangements, access controls and governance frameworks, as well as revisiting assumptions about whether information is genuinely anonymous. 

With the consultation open until 19 October 2026, organisations involved in research, AI and data-driven innovation should monitor developments closely. The final guidance is likely to provide an important reference point for organisations relying on the RAS provisions, particularly when assessing identifiability, selecting appropriate safeguards and determining whether anonymised, pseudonymised or identifiable data is justified for a research purpose.

For queries or advice on the content of this article, please contact Hamish Corner, Lucy Pegler, Amanda Leiu or a member of Burges Salmon's Commercial & Technology team. 

This article was written by Fraser Campbell and Amanda Leiu.