Integrating AI into historical research: a conversation with Benoît Majerus about his ERC project on the history of offshore shell companies

When Benoît Majerus began his Letterbox project, AI was not on his radar. But he’s now relying on it as he embarks on a related project, DIGSHELL, recently awarded a European Research Council (ERC) Advanced Grant.

It was in the middle of the Letterbox project, 18 months after it kicked off, that members of the C²DH’s Digital Research Infrastructure team suggested that Benoît Majerus could use AI for his research. The aim of Letterbox was to analyse historical data linked to Luxembourg shell companies, going beyond bankers in a bid to shed light on the connections between other players in the financial sector such as lawyers and notaries.

As a historian specialising in the history of psychiatry in the 20th century and the history of offshore financial centres, Majerus was not an expert in digital history, but he had been finding it challenging to extract data from sources like Luxembourg’s business registry (Registre de commerce et des sociétés).

Harmonisation across various formats and languages

The team wanted to find a simpler, quicker way to extract names and addresses which were often written in different ways from one document to the next. The multilingual nature of Luxembourg also further complicated things. Majerus saw two major challenges in using AI for such tasks. “The first is that if we work with these digital methods and AI, the answers may not always be correct, but they are more likely to be so,” he says. “The second problem is how we cite the results.”

As he points out, historians who are referred back to a source can normally check the original themselves, but when it comes to AI there can be issues with reproducibility. “Sometimes if you ask the same question to the same AI model, you’ll get different results,” he notes. “How will historians deal with that?” Majerus adds that some AI models are black boxes and it can be hard to know how the results are produced.

‟ Sometimes if you ask the same question to the same AI model, you’ll get different results. How will historians deal with that?”

Prof Benoît MAJERUS

Full professor/Chief scientist 1 in European history, Social history, History from below 20th century

View profile

For the DIGSHELL project (short for “Digital archaeology of shell companies: infrastructures of global connections and local networks”), Majerus and his team will receive around €3.3 million in the form of an ERC Advanced Grant. This will allow them to widen their focus to include Panama, the British Virgin Islands and Singapore, but they will have to manage even more data and languages. In Singapore, for instance, the data Majerus is seeking exists only in PDF format, whereas in Panama the records are handwritten, complicating extraction in a different way. In the British Virgin Islands, records are found in newspapers and will have to be digitised by the team.

AI should also be able to assist the researchers in linking names and address. For instance, in the documents for Singapore, there are some Chinese names that come up hundreds of times but don’t refer to the same person, so the researchers need to clarify which individual is linked to which address. Meanwhile, in Panama, names tend to be longer or double barrelled, and here AI can also come in handy. “AI will just help us identify networks, the people who are most central,” Majerus adds, “but then we will have to do the conventional archival part and look at other sources.”

Two sides of the coin

Majerus describes his view of AI as being both emotional and rational. While he isn’t fully comfortable with certain aspects of AI – biases or transparency issues, for instance –, rationally he knows that it will save a lot of time and effort for the team.

Recently, as he was writing an article for the Journal of Digital History, he had to include some probabilistic findings in his footnotes – something he’d never done before and which he says felt rather “bizarre”. He’s still learning how to cite these references and has plenty of questions about the level of granularity that needs to be included in such findings.

“For me, AI is a tool, and I want to participate in that debate,” he concludes, “but what really interests me is the offshore history. And to be able to tell that story, I need AI.”