Thanks to British bureaucracy, convicts sent to Australia in the 18th and 19th centuries left behind a trail of data that may yield new meanings when fed through artificial intelligence (AI).
University of New England (UNE) History researchers are investigating whether AI can turn old records into narrative biographies that recognise vanished convicts as more than a series of datapoints.
The UNE research team’s experiment looked at whether historical data could be converted by AI into short-form encyclopedic sketches – think of Wikipedia entries – for anyone requiring quick and accessible biographical sketches of convicts or other recorded historical figures.
Initial results showed that AI has promise in this role, but the researchers found that the technology currently has a prominent flaw as a curator of history. Trained on interactions between modern people, it picks up cultural mythologies and weaves them into its interpretations as fact.
The “AI as biographer” experiment was led by PhD candidate Mark Mclean. The results have been published in the Taylor & Francis journal, Historical Methods: A Journal of Quantitative and Interdisciplinary History.
The pilot used a dataset of records from the short-lived Norfolk Island settlement, which operated between 1788–1814. The data has been curated over the last decade by UNE historian Professor David Andrew Roberts and UNE Master of History graduate, Cathy Dunn.
Mr McLean said the experiment aimed to extract “the ghosts from the machine” by converting the data of colonial administrative records into historical life-course biographies of individuals.
Image: AI "biography" of convict Denis Mohair generated from digitised colonial records.
The results were mixed, but encouraging.
The work involved in preparing the data – originally handwritten into ledgers – was considerable. “Before the data becomes ready for the machine, it has be carefully curated,“ Mr McLean says. ”In that there is still no substitute for human intelligence and intuition. AI platforms feed on immense hours of human labour.”
Then the project team found that AI, rather than provide an objective rendering of a subject’s life, drew on its training data to weave some cultural myth through its interpretations.
“Because the language models drew heavily from the web, the biographies became infused with populist narratives about colonial pioneers and heroes,” Mr McLean says. “That does not sit well with serious scholars. In fact, our database contains some very surly characters, many of them convicts under sentence for ugly crimes.”
This was no surprise to Professor Roberts, who teaches Australian history.
“This is the old convict-made-good narrative,” he observes. “It’s how white Australians have excused and marketed their past for over a century. It’s a one-dimensional nationalist myth, shallow and often badly inaccurate.”
“Because this is a persistent myth in Australian culture, it was picked up by AI and treated as an element of the biographical narratives it created.”
Yet overall, the project team found the results were effective. AI proved capable of turning datasheets into life-course descriptions, with some limitations.
“LLMs helped to accomplish what we intended,” says Mr McLean, pictured, “but AI was only one step in the larger research process. You still have to create the dots that AI will join.”
In their journal article, the research team proposed some solutions to overcome gaps and bias, including tampering with the language model to provide additional context from contemporary documents like letters and diaries.
Professor Martin Gibbs, a UNE archaeologist who works closely with industry partners including Norfolk Island’s Kingston and Arthur’s Vale Historic Area, says that AI still has considerable potential to revolutionise historical and cultural heritage research.
“Our historical datasets have become immense,” Prof. Gibbs says. “They make good sense to us, but we need new ways of mining and presenting them so they become accessible to our end-users, and to a wider audience of history lovers. The heritage and history organisations we work with don’t have the tools or time to read our database the way we do as researchers.”
The recent experiment indicated that AI shows promise for this work, but Prof. Gibbs believes that for researchers, the results are still more notable for flaws than for practical usefulness.
Prof. Roberts adds that AI’s canned outputs “are no substitute for the very human enterprise of writing biography”.
Experimenting with AI is part of the years-long adoption of a ‘digital Humanities’ approach to research and teaching by UNE’s History and Archaeology departments.
“We’ve been converting historical documents into tabulated data for years now,” says Prof Roberts. "These days a lot of our research publications are based on record linkage, life-course studies, and network analysis. Tabulated data also allows us to more readily convert historical records into something that can be used by our students and a variety of non-specialist end-users. Our work in AI is a progression of this use of digital technologies to make these vast datasets intelligible to others.”
The research is part of a long-running agenda, funded in stages by the Australian Research Council, which tailors historical and archaeological research to the needs of cultural heritage industry partners.
UNE’s expertise in this field recently culminated in an ARC Linkage project, Making Crime Pay, with National Trust Tasmania. This project applies interdisciplinary approaches and digital techniques to the preservation of local heritage sites in Tasmania, where heritage tourism generates over $2 billion annually to the local economy.
Paper: Ghosts and the machine: testing the use of Artificial Intelligence to deliver historical life course biographies from big data. (Mark McLean, David Roberts & Martin Gibbs) - https://www.tandfonline.com/doi/full/10.1080/01615440.2024.2398455
