How Is Artificial Intelligence Changing Historiography?
From handwriting recognition to network analysis: where artificial intelligence touches the historian's method, and the traps to watch out for.
Translated from the Turkish original: Türkçesini oku
IN SHORT
Artificial intelligence speeds up the historian's most time-consuming tasks: reading handwriting, searching large corpora for patterns and mapping networks of correspondence. But interpretation still belongs to the historian. Because language models can invent sources, their output must also pass through source criticism.
Historiography has always been a problem of sources. To talk about the past, you first need to reach the traces it left behind and be able to read them. The trouble is that most of these traces are still unread: uncatalogued boxes, untranscribed manuscripts, unindexed registers. There is more out there than a historian could reach in an entire career, and access itself is the bottleneck.
Artificial intelligence has brought several tools that touch this bottleneck at once.
Handwritten text recognition (HTR)
How does artificial intelligence read handwritten documents?
The success of optical character recognition (OCR) with printed text has reached manuscripts only in recent years. Platforms such as Transkribus let historians train models on their own archives: a sample transcription of a few dozen to a few hundred pages becomes a reading machine for the remaining millions.
This matters especially in fields such as the Ottoman archives, where manuscripts in Arabic script have not received the attention they deserve. The number of experts who can read the old script is limited; machine support promises to move their eyes and their years to more productive work. But accuracy varies greatly with the regularity of the handwriting, the language and the abbreviations, and every text that has passed through the machine still needs a human eye.
Long series and big data
What do millions of books add to historical research?
The culturomics study, published at the end of 2010 and based on about 5 million books scanned by Google, could chart how often words were used between 1800 and 2000. This approach is related to what the literary scholar Franco Moretti called distant reading in 2000: instead of reading individual texts one by one, searching for patterns across huge corpora.
Quantitative history was an early forerunner of this, with its economic series, price registers and population records. What artificial intelligence changes is scale and flexibility: we can now turn non-numerical texts (court records, newspaper clippings, letters) into series as well.
Network analysis
What do networks of correspondence tell the historian?
One of the most visual methods historians have adopted in recent years is network analysis. Stanford’s Mapping the Republic of Letters project mapped whom Voltaire or Franklin wrote to, and when, as nodes and edges. From networks of letters one can reconstruct the geography of intellectual communities, the direction of influence and the people who carried it.
The same method can be used for pious foundation deeds, trading partnerships or court appointments. But every network graph is an interpretation: you decide what counts as a node and which tie counts as “important”, and that decision quietly shapes the result.
Language models: translator and assistant, but not witness
Can language models be trusted in historical research?
Large language models have become quick assistants for working with sources: drafting a translation from an old language, summarising a long register, listing the period meanings of a term. These are valuable gains.
The trap is this: the model is fluent, but it is not a witness. It can even invent sources. A 2023 study found that more than half of the bibliographic citations produced by GPT-3.5 were fabricated. When a model invents a source, it does something no different from the gravest professional sin in history writing. Its output also needs source criticism: who produced it (which model), with what data, when, and in answer to which question? If the answers do not hold up, the output cannot even count as a “secondary source”; it can only be treated as a hypothesis.
Tools change, the question remains
Can artificial intelligence replace the historian?
I would sum up the contribution of artificial intelligence to historiography like this: it speeds up access and processing, but it cannot replace interpretation. It is still the historian who decides which document matters, what a pattern means and which bridge to build between the past and the present.
One could go further: new kinds of sources, such as log files and social media archives, will give rise to a new tradition of source criticism in the coming years. What we call the digital humanities is precisely the effort to ask old questions of new sources.
References
- Edelstein, D., Findlen, P., Ceserani, G., Winterer, C., & Coleman, N. (2017). Historical research in a digital age: Reflections from the Mapping the Republic of Letters project. American Historical Review, 122(2), 400–424. https://doi.org/10.1093/ahr/122.2.400
- Michel, J.-B., Shen, Y. K., Aiden, A. P., Veres, A., Gray, M. K., The Google Books Team, Pickett, J. P., Hoiberg, D., Clancy, D., Norvig, P., Orwant, J., Pinker, S., Nowak, M. A., & Aiden, E. L. (2011). Quantitative analysis of culture using millions of digitized books. Science, 331(6014), 176–182. https://doi.org/10.1126/science.1199644
- Moretti, F. (2000). Conjectures on world literature. New Left Review, 1, 54–68.
- Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5
