National genomics initiatives collect and link genomic data with health records and other sensitive information to support research and improve healthcare. AI offers new ways to analyse these resources, but also raises questions about protecting participant data and controlling how trained models are used. 

The researchers reviewed publicly available information and found that out of 90 initiatives, 32 initiatives (36%) reported using AI or planning to use it. Howeever, publicly available AI-specific policies were identified for only three of the 32: All of Us, Genomics England and UK Biobank. 

Across these policies, the study identified three safeguards in these policies: 

• Secure computing environments: requiring researchers to analyse participant data within controlled platforms. 

• Controls on exporting trained AI models: recognising that models can retain or reveal sensitive information about the people whose data was used to train them. 

• Restrictions on external generative AI services: limiting the transfer of participant data to third-party tools. 

The review also found limited policy guidance on assessing AI bias and fairness, informing participants about AI use, and governing commercial AI development. Most AI guidance appeared in researcher-facing FAQs or separate documents, rather than core data-use policies or participant information. 

The review recommends to incorporate AI safeguards into data access policies, consent processes and communication with participants.