Should sensitive archival images be shown? The case for applying AI to colonial archives

Abstract Should sensitive archival images be shown—and if so, how? This article examines the ethical, legal, and technical stakes of providing access to colonial-era photographs at scale for researchers, archivists, curators, and wider online publics, including members of source communities, arguing for qualified access supported by human-in-the-loop AI. Drawing on fifteen semi-structured interviews with professionals across GLAM institutions in the UK and continental Europe, we map current practice around sensitivity review, metadata enrichment, and collaboration with source communities. We complement these insights with a small pilot that applies a commercially available computer-vision tool (Google Cloud Vision SafeSearch) to twenty colonial photographs from the Archives Nationales d’Outre mer (France). This pilot shows that SafeSearch under-detects violence, particularly in monochrome, staged, or context-dependent scenes (a classic domain-shift problem). It performs comparatively better at flagging nudity within this small sample – an area that engages clear legal and safeguarding obligations for institutions within UK and EU frameworks (particularly where minors may be depicted). We outline a pragmatic pipeline for archival settings—combining SafeSearch-style triage, OCR/HTR of mounts and captions, authority-linked entities, semantic (vector) search, and constrained multimodal captioning—while acknowledging that these findings are based on a limited pilot, and emphasising that models inherit training biases, struggle with historical formats, and must never replace curatorial judgement. Interviewees converge on governance needs: clear takedown and escalation procedures; explicit labelling of AI-generated metadata; least-privilege technical setups; and early, sustained engagement with source communities. We propose a decision framework that aligns institutional risk (legal compliance, safeguarding, staff wellbeing) with public-interest goals (research access for scholars, students, and other users of digital collections; accountability to affected and source communities; and broader public engagement with histories of empire and violence), and that positions co-governance with source communities as a core condition of responsible access. The result is a practicable route for institutions working in similar legal and professional contexts to make difficult images more discoverable while minimising harm: AI can assist with scale; archivists and communities retain authority over what is surfaced, how it is described, and when warnings or access restrictions are required. By aligning institutional risk with public-interest goals, the framework also reflects institutional responsibilities and pressures, including legal compliance, reputational risk, and staff wellbeing, and therefore addresses not only user access but also the practical and ethical constraints under which GLAM professionals operate.