On Tuesday 22 July 2025, a one-day workshop titled Government Records and AI was held at the Mercure Piccadilly Hotel, Manchester, as part of the LUSTRE/GLOW initiative led by Professor Lise Jaillant and supported by Loughborough University in collaboration with the UK Cabinet Office. GLOW builds on the work of the LUSTRE project and brings together government professionals, archivists, technologists, and researchers from the UK and US to explore how AI technologies can improve access to government records while addressing issues of sensitivity, metadata, and trust.
The workshop provided a focused platform for participants to discuss current practices, identify shared challenges, and explore responsible, collaborative approaches to using AI across institutional and national contexts.
The presentation slides and abstracts from this workshop can be accessed HERE.
The workshop began with a keynote by Professor Victoria Lemieux of the University of British Columbia, who presented “Clio X: A Web3 Solution for Privacy Preserving AI Access to Digital Archives.” Lemieux introduced Clio X as a decentralised, privacy first Web3 solution that embeds privacy enhancing technologies into archival workflows so institutions can enable AI supported reference and access while maintaining control of sensitive content. Drawing on a user evaluation of a medium fidelity prototype, she discussed adoption barriers around trust, system opacity, costs, and governance. She alsooutlined a participatory, Decentralized Autonomous Organization (DAO) based approach to ethical and transparent uses of AI in cultural heritage and public records contexts.
In the Q& A after the keynote, discussion centred on where sensitivity review should sit and how to keep access workable for researchers, with Professor Victoria Lemieux pointing to privacy-enhancing processing that can reveal useful patterns while withholding identifiers. She also noted the need to run in secure environments with vetted toolsets and controlled access, so archives can support research and appraisal without compromising protection.
The first session Advancing Security and Scalability in Digital Collections, began with Dr James Lappin from the Department for Science, Innovation and Technology, who spoke on “Managing the Digital Heap: Intuitions from Records Management and from Data Science.” that services must scale alongside human-in-the-loop controls and explanations, and that effective practice will draw on both records management and data science to address the full variety, systems, and lifecycles of digital content.
He was followed by Professor Georgina Cosma from Loughborough University, who presented “Preserving and Accessing: Secure AI Solutions for Sensitive Information Systems.” Cosma outlined a Retrieval Augmented Generation framework designed to operate entirely on curated internal sources, avoiding external dependencies. She described how semantic search, classification, and question answering can be delivered in secure environments. She also surveyed agentic RAG, multimodal processing, and the practical challenge of unlearning so that information can be removed from AI systems while preserving transparency and accountability in sensitive settings.
In the second session Evolving User Communities and Institutional Approaches, Dr Adam Nix of the University of Birmingham presented “Understanding Methodological Diversity in Born-Digital User Communities,” mapping four user types, the aggregator, the synthesiser, the fact finder, and the narrator, each with distinct motivations and access needs. Heargued that designing with methodological diversity in mind prevents systems and services from privileging one research approach over others and leads to more equitable, effective access.
Following it, Elizabeth (Lizzie) Hatfield and Andrew Bonnie from the Bodleian Libraries delivered “The Future of AI and Libraries: Opportunities and Challenges at the Bodleian,” situating their exploration of AI in the Bodleian’s mission to enable access for learning and research, outlining the scale, services, and digital strategy that frame the work, and setting out practical challenges they are addressing, including metadata quality, trust, copyright, academic integrity, environmental impact, and infrastructure modernisation as part of a wider digital transformation.
After the break, in the third session Policy Innovation and Practice in Digital Heritage, Dr Katherine McDonough from Lancaster University presented “Historical Maps and the Future of Land Use,” emphasising that historical maps are not mirrors of the past and showing how carefully designed computer vision methods can interrogate very large map collections while remaining sensitive to cartographic genres, an approach that supports policy and land use work without flattening source complexity.
That session continued with Mahin Ali of SVGC presenting “Navigating the Digital Heap with AI,” explaining how machine learning classifiers are being used to accelerate transfer of digital collections to The National Archives while mitigating risk, he outlined a semi-automated triage approach paired with human in the loop sampling and guardrails based on training data sufficiency to increase throughput and deliver public value within agreed risk controls and archival policy.
With those themes in view, the afternoon Plenary Discussion, Collaboration, Access, and Institutional Trust in the Age of AI brought the room together around eight prompts, with each table taking two or three in turn, so the conversation moved from enhancing access and overcoming barriers to ethics, archival principles, future practice, cross sector collaboration, sensitive content detection, and the evolution of work, and it kept circling back to practical workflows, data readiness, transparency, and the need to keep humans in the loop as institutions test and learn.
1. Enhancing Access with AI
Participants tied potential gains in identification and discoverability to very practical steps in existing workflows. They stressed that outputs should be checked for being legally sensible and actionable, then proceed to digital scanning and quality assurance. Any improvement in access should move together with assurances about authenticity, custody, and public confidence, and in some cases, systems may need to maintain or display multiple versions.
2. Overcoming Barriers to AI Adoption
Groups highlighted data readiness and organisational culture as early obstacles. One table cautioned against assuming AI can clean the mess, arguing the foundations must be put in order first or pilots risk seeding mistrust. Concerns also included environmental impact and anxiety about job change among information professionals, with a push to equip staff to use precise technical language rather than umbrella terms. Another table added money, time, confidence, legitimacy, change fatigue, data protection, and the wider political economy, noting that fast tech cultures can collide with archival practice where moving fast and breaking things is not an option.
3. Ethics and Responsibility in AI Use
Participants kept the human central, asked for visible accuracy levels, and called for keeping a record of model changes rather than overwriting them. They discussed bias, hallucinations, and loss of context, including an example where an image classifier began labelling a toy robot as a rock after small changes, which was used to underline the need to educate users about dataset bias. Regulatory frameworks were offered as an analogy for progress with guardrails.
4. Reconciling AI with Archival Principles
The discussion revisited authenticity, provenance, and original order, with suggestions to make the steps that ensure authenticity more explicit and transparent. Several speakers observed that institutions often frame ethical assessment through a risk mitigation lens, which shapes how AI work is evaluated.
5. Looking Ahead: AI and the Future of Archival Practice
Looking five to ten years ahead, groups connected changing record containers and classification practices with risks from synthetic information. One contribution warned about a drift from original meaning if language models repeatedly generate and then consume institutional text, so controls and anchoring in original sources remain important as workflows modernise.
6. Cross-Sector Collaboration for AI Innovation
Collaboration was described as people centred and inevitably slow, because partners work at different speeds and under different incentives. Participants called for shared understanding and shared values across archives, libraries, records management, and academia, and suggested that professional bodies could convene joint efforts instead of working in silos
7. AI for Sensitive Content Detection and Risk Management
Tables distinguished generative systems from other techniques and often preferred trainable classifiers, natural language processing, ontologies, and knowledge graphs for sensitivity review. They stressed that context and risk should guide tool choice and noted that some collections may be too sensitive for AI. Temporal change also matters, for example a classification shift after Brexit, so human-in-the loop supervision remains necessary and prior human decisions can seed training libraries.
8. AI and the Evolution of Archival Work
Participants anticipated shifts in competencies toward clearer technical vocabulary and better judgement about when a particular tool is fit for purpose. They discussed translating from paper-based sensitivity review to digital sensitivity review by building on existing frameworks rather than reinventing processes, and they observed that information is increasingly stored around people rather than in rigid file plans, with implications for future appraisal and description.
After the breakout group session, in Keynote and Closing Remarks, David Canning MBE of the Cabinet Office presented “Everything, Everywhere, All at Once,” reflecting on the constant challenges that archivists and records managers face in keeping up with and adapting to new and emergent technologies. He outlined how we can keep focused through the hype, and some of the steps already being taken by the United Kingdom Government to protect and preserve the official record.
The workshop concluded with a Q&A following David Canning’s keynote, drawing together the day’s discussion. Topics included the shift from paper to fully digital workflows and an unbroken chain of information, how to treat Teams chats, video meetings and consumer messaging alongside email as formal records, and the need for clear retention and privacy rules to guide practice.
Overall, the GLOW workshop offered practical insights into privacy by design, secure and scalable access, user-centered discovery, and the evolution of record keeping, and it highlighted the importance of continued cross-sector collaboration to turn these conversations into shared guidance and pilots.