Generative AI for Historical Research: Augmenting, Not Automating

Custom GPTs, built upon powerful language models like GPT-4, offer historians a unique opportunity to tailor AI for specific research needs. Breen introduces his own creation, the "Historian's Friend," as an example of how these agents can be configured to assist with tasks ranging from transcribing and translating historical texts to analyzing visual media. The author emphasizes that the goal is not to replace the historian's critical thinking or interpretive skills, but to provide tools that can handle laborious aspects of research, freeing up human intellect for higher-level analysis and synthesis. This approach shifts the focus from AI as a potential threat to AI as an indispensable partner in the pursuit of historical knowledge.

The second case study focuses on an 18th-century Catalan drug manual, testing the AI's capabilities in paleography (transcribing old handwriting) and translation. While AI language models are generally good at translation, Breen found that GPT-4 made transcription errors in the handwritten Catalan text, leading to flawed translations. For instance, it mistook "sal eſſencial" for "Mencíal Salt" and struggled with compound words due to line breaks. Despite these transcription issues, the AI was useful for "getting the gist" of the text, aiding Breen in understanding a language he could only partially read. More significantly, when asked to interpret an accompanying image of a sugar refinery, the AI provided a detailed and insightful analysis of the "Anatomy of the Engraving," correctly identifying the sugar mill, processing areas, and the likely labor dynamics, including the implication of enslaved labor. This demonstrated AI's strength in analyzing complex historical imagery, even when the accompanying text has transcription errors.

The final case study involved feeding pages from a challenging 1749 Portuguese-Brazilian medical treatise, "Prodigiosa lagoa descuberta nas Congonhas," into the custom GPT. This book is a crucial resource for understanding the health perspectives of enslaved Africans. Despite some minor transcription errors (marked in red by Breen), the overall translation was accurate. A particularly interesting error occurred when the AI translated "quigilia" (a complex ailment with roots in Central West African cosmologies, often diagnosed in people of African descent) as "gangrene." When questioned, the AI admitted its error. Breen notes that this type of error, while seemingly minor, can silently elide the most interesting aspects of a historical text. However, he also recognizes that the AI's persistent struggle with this unfamiliar term was precisely what led him to investigate it further, ultimately uncovering Júnia Ferreira Furtado's insightful research on the term's etymology and cultural significance. This experience underscores how AI's errors can sometimes serendipitously guide researchers to deeper discoveries.

One of the most compelling potential impacts of generative AI on historical research, as highlighted by Breen, is its ability to democratize the field. By lowering the barriers to entry for tasks such as transcription, translation, and initial source analysis, AI can empower a wider range of individuals, including students and non-experts, to engage with historical materials. The iterative, question-and-answer format facilitated by AI tools like the "Historian's Friend" can help newcomers quickly learn the intricacies of historical subjects. Breen suggests that undergraduate history majors, equipped with such tools and working in tandem with experienced professionals, could conduct significant original research. This potential for broader participation promises to enrich the historical discourse and bring new perspectives to light.