AI and history

Uncovering evidence of historical theories and identifying patterns in past events has long been held back by the time-consuming process of inputting data from artifacts and handwritten records.

The introduction of artificial intelligence and machine learning techniques accelerates such research and draws attention to information that may be overlooked. That said, history, as a humanities discipline, is struggling to compete with other applications that are more future-oriented. For example, start-ups that aim to analyze Facebook will receive better funding than those that aim to study ancient Greek artifacts.

Nevertheless, applying algorithmic methods to historical research could improve AI’s capabilities, as recovering or recreating archaeological artifacts is a challenging task for computer vision models. Until now, neural networks have been extensively trained to re-assemble photos and process documents, but there has been no training to account for the degradation of fragments over time, blurred images, or inaccurate detail fitting.

Before the AI model can reconstruct artifacts, it must learn to reverse erosion and predict how the original slices look. The machine must then check that the fragments fit together.
Last year, a research paper was published on a deep learning model designed to fill in the missing text in ancient Greek inscriptions. Not only did this research help develop a tool for historians, but it also solved a big artificial intelligence problem: understanding algorithmic decision-making.

Researchers are increasingly dealing with digitized information rather than primary sources stored in archives. And without adequate adaptation of sourcing and analysis methods, their theories and conclusions will be distorted. There is a risk that individual letters and words can be misidentified and may not be present in the documents at all. At a more fundamental level, historical data processed algorithmically can mislead historians.

US professor Connelly recently pointed out that machine learning algorithms can overestimate the historical significance of some documents and overlook others that turn out to be important. Because of this, some historical data can be destroyed.

Therefore, AI is best suited to identifying gaps in historical records and recovering artifacts. But historians should be critical of the use of scientific methods when dealing with data.