Public views on the use of LLMs in identifying privacy risks in sensitive free-text data: findings from deliberative workshops to inform the STAR-TRE project

STAR-TRE (Safe and Trustworthy Assessment of Risk in Trusted Research Environments) is a new project led by the University of Edinburgh working in collaboration with the Scottish Safe Haven Network and Brighton and Sussex Medical School. The project is exploring how different tools and technologies could be used to further address risks to confidentiality in free-text data before these are made available for research. The project is driven by several Trusted Research Environments (TREs) including DataLoch, a partnership between the University of Edinburgh and NHS Lothian. The STAR-TRE project is focussed on free-text (unstructured) administrative data. Free-text data contains rich contextual information which can help understand complex societal situations and outcomes. However, free-text data is rarely shared for research purposes due to the unpredictable presence of indirect identifiers. When these indirect identifiers – such as recognisable locations, social circumstances, and familial relationships – appear individually, the possible privacy risk is relatively small, but when combined they may reveal someone’s identity. Technology, such as Large Language Models (LLMs), could offer opportunities to support the assessment of privacy risks within free-text extracts. Since LLMs are still relatively new, the STAR-TRE project team wanted to understand how the public felt about their use for this purpose. DataLoch commissioned Ipsos to design and carry out a programme of deliberative engagement to understand public perspectives on the use of LLMs in identifying privacy risks within sensitive free-text data and the types of safeguards they would like to see in place. Deliberative engagement was carried out in November and December 2025 with a group of 38 participants from all four UK nations, delivered in two stages. It started with an introductory online learning session with all participants to introduce them to key topics. This was followed by four in-person sessions, in Edinburgh, Belfast, London and Cardiff, which reconvened participants based on their location and focussed on participants’ deliberation and conclusion-forming. Across the online and in-person sessions, participants worked to address the following overarching questions: 1. What are the benefits of using LLMs to identify the privacy risks in free-text data?2. What concerns do you have about using LLMs for this task?3. What measures should be in place to ensure that the use of LLMs is safe and trustworthy?