At the Precipice of Possibility: Reflections on the Generative AI and Historical Research Workshop

By Dr Bronagh Ann McShane

On 7 May 2025, more than fifty individuals gathered at Trinity College Dublin for the workshop, ‘Generative AI and Historical Research’. Convened as part of the ERC-funded VOICES project and generously supported by the Faculty of Arts, Humanities and Social Sciences Events Fund at Trinity College Dublin, the workshop brought together historians, computer scientists, digital humanists, and archivists to explore how AI might shape our understanding of the past, and what responsibilities we bear in the process of engaging with this technology.

In my opening remarks, I characterised the day as ‘A Precipice of Possibility’, a phrase chosen to encapsulate the spirit of our gathering: One defined by curiosity, critical reflection, and the sense of standing at the threshold of something both uncertain and full of promise.

The day began with a welcome address by Professor Jane Ohlmeyer, Principal Investigator of the European Research Council Advanced Grant Project,
VOICES, who reminded us of both the potential and the limits of digital tools in recovering the lives of early modern women. She emphasised the need for interoperability across digital silos and made a compelling case for collaboration between historians and computer scientists to achieve research goals that would otherwise remain out of reach.

Our keynote speaker, Professor John Kelleher, Director of Research Ireland’s ADAPT Centre (https://www.adaptcentre.ie/), offered a comprehensive and jargon-free introduction to how Large Language Models (LLMs) work and how they can be applied to historical research. His demonstration of entity extraction from historical texts (names, dates, places), highlighted how AI can aid information retrieval at scale. Yet his talk also served as a cautionary tale. LLMs are not databases; they are probabilistic generators trained on patterns, not facts. Their capacity to hallucinate plausible but inaccurate content raises fundamental concerns about trust, provenance, and scholarly rigour.

We were especially honoured to welcome Dr Delfi Nieto-Isabel (Lecturer at the School of History and Fellow at the Institute for the Humanities and Social Sciences, Queen Mary University of London). In her reflections, Dr Nieto-Isabel described AI as the next paradigm shift in historical methodology, akin to the transition from oral to written culture, famously critiqued by Socrates. She stressed that AI is not intelligence in any true sense, and that the misnomer does historians no favours. Crucially, she reminded us that bias is not new: archives and languages have always encoded cultural values. What matters now is traceability, transparency, and the development of institutional strategies that enable scholars to work across disciplinary boundaries. She also encouraged us to look beyond plagiarism fears and consider AI a powerful, if flawed, pedagogical and research tool.

A series of case studies grounded these abstract debates in practical realities. Dr Madina Kurmangaliyeva (Research Fellow, VOICES) presented her work on the 1641 Depositions, showing how AI can dramatically reduce the time needed to extract structured data from historical testimonies. Her rigorous methodology (cross-checking AI outputs with manually coded data) underscored the importance of human oversight, even in highly automated workflows.

Dr Marvin Suesse (Associate Professor in Economics, Trinity College Dublin) introduced an ambitious project using GPT-4 to structure biographical data on political dissidents in Tsarist Russia. His emphasis on iterative prompt engineering, data cleaning, and transparency illustrated the complex but promising integration of LLMs into socio-economic history. Similarly, Dr David Brown (Senior Research Fellow,  Irish Research Council Advanced Laureate Empire project) explored how Retrieval-Augmented Generation (RAG) might offer a more reliable model for producing historical narratives grounded in verifiable sources.

In the afternoon, the team behind EyeCR, a prototype tool developed at the Virtual Record Treasury of Ireland (https://virtualtreasury.ie/) and led by Dr Peter Crooks (Associate Professor in Medieval History, Trinity College Dublin), presented a historian-facing interface that connects digitised documents directly to structured outputs using OCR and generative AI technologies. As Dr Ciaran Wallace (Keeper of the Virtual Record Treasury of Ireland) discussed, their stated aim is to transform historical images into usable, citable data in a transparent and reproducible way (or to spin gold from straw). Crucially, EyeCR underscores the importance of human intervention and domain-specific configuration, reinforcing the central message of the day: AI can assist, but not replace, historical interpretation.

The roundtable discussion that concluded the workshop drew together key themes and raised urgent new questions. Participants voiced concerns about linguistic and cultural bias in English-trained LLMs, the ethics of including (or excluding) hate speech from training data, and the increasing pressure from funding bodies to produce quantitative research. There were calls for institutions to equip students with the critical skills needed to navigate AI and for historians to proactively engage with public policy in this evolving space.

A particularly memorable contribution came from Professor Declan O’Sullivan (Professor in Computer Science, Trinity College Dublin), who proposed rethinking the acronym ‘AI’ not as ‘Artificial Intelligence’but as ‘Appearing Intelligent’, a term that captures the illusion of understanding these models can create. This playful reframing served as a timely reminder of the need for precision in our language and caution in our interpretations. O’Sullivan also highlighted the value of socialising the experience of using AI: sharing tools, methodologies, and even failures within the humanities disciplines, and across disciplines and publics.

As workshop convenor, I left the day both invigorated by Generative AI’s possibilities and sobered by its responsibilities. This technology is not a passing trend; it is a paradigm shift. But scholars are not powerless in the face of this change. We can ask hard questions. We can design better tools. We can demand transparency, accountability and traceability. And we can use these technologies not to flatten historical complexity, but to help recover the lives and voices long obscured in the archival record.