When I sat down with bell hooks’ personal journals at an archive at Berea College in Kentucky, I expected an intimate peek into her private thoughts, her voice before the editing. What I got instead was frustration. Her handwriting was dense cursive, all loops that looked identical to my eye, and there were years of journals to go through. I found myself photographing pages and feeding them to ChatGPT just to read what she’d written. My tool of choice worked well, and it turns out I’m not the first person in an archive to have figured this out.

Getting computers to reliably read human handwriting, in all its variations, has challenged AI researchers since the earliest days of the field. Researchers in the 1960s predicted machines would soon simply devour handwritten text; instead the problem spawned decades of specialized research and entire commercial industries. Yann LeCun, who later went on to win the Turing Award for his contributions to deep learning, published landmark work on handwritten digit recognition in the 1980s that showed what was possible in narrow, controlled settings. Real archives were another matter.

Now that boundary is moving. General-purpose AI models are not perfect readers of every handwritten page, but they are now suddenly good enough to change what archives can do. Pages that once required paleography training, custom software, or weeks of squinting can produce usable transcriptions in seconds. Collections that were preserved but functionally hidden are becoming searchable, opening up the ability for scholars and families to ask questions they rarely had the time or money to ask before.

Scaling AI to Decipher Archival Handwriting

Mark Humphries had spent a decade wrestling with scale. A professor of history and coordinator of the applied generative AI program at Wilfrid Laurier University in Waterloo, Ontario, had digitized 10 million pages of World War I pension records in Canada. But with no index and no standardization, finding an individual pensioner meant going through files at random. The records were written by hundreds of different clerks, officers, and administrators, which ruled out the standard workaround of training a specialized model to recognize one person’s handwriting.

When OpenAI’s GPT-4 came out in 2023, Humphries started feeding it handwriting. The results were rough but better than any general tool he had tried before, and he wanted to know whether the trick would hold up. Humphries and his colleagues at Wilfrid Laurier spent two years systematically testing what these models could actually do. Their results, published in May 2025 in Historical Methods, backed up his anecdotal evidence. On a corpus of 50 English-language letters, legal records, and diary entries dating from the 18th and 19th centuries, large language models (LLMs) outperformed Transkribus, the specialized handwriting recognition software used by more than 150 major universities and archives, on accuracy, speed, and cost.

On documents it had not been trained on, Transkribus had character error rates of around 8 percent. Humphries’ best LLM-based approach pushed that below 2 percent, while completing the work 50 times as fast and at roughly 1/50th the cost. Transkribus, for its part, has announced it is integrating LLMs directly into its own platform.

“The dream was to have something like what we have now,” Humphries says.

Humphries has a theory on why. The AI researcher Richard Sutton argued in 2019 that general methods leveraging computation will always eventually outperform specialized ones. Humphries thinks that is exactly what is happening here. The general models have been trained on such a vast range of data that somewhere in that pile they absorbed the relationship between handwritten documents and their transcriptions without anyone explicitly teaching them to.

The practical consequences are already unfolding. Lianne Leddy, an associate professor of history and the Canada Research Chair in Indigenous Histories and Historical Practice who was one of Humphries’ coauthors, traces Indigenous women’s experiences across North America through fur trade post journals, baptismal records, and marriage registries scattered across archives all over Canada.

The records were almost all written by men working as clerks, priests, and post employees whose focus was rarely the Indigenous women around them. Surfacing these stories requires reading thousands of documents to find a handful of relevant details. The women’s names were often spelled phonetically—differently by French, English, and Scottish writers—or recorded only as someone’s wife.

“To build those stories up would take several careers doing things in the traditional way,” Leddy says. “This really changes the scale of what’s possible.”

AI Transcription in Historical Archives

The implications are already rippling through institutions. At the University of North Carolina at Chapel Hill, librarians are experimenting with AI transcription across their special collections material that gets heavy use from people tracing enslaved ancestors. The team found the models handled letters and diaries well and made a particular breakthrough with ledgers, which tend to have tabular structures that shift from page to page and had long been difficult to process.

“Gemini can do tables very, very well,” said Jackie Dean, one of the archivists leading the project. “For our use case, that was a major leap forward.”

It is not only universities paying attention. The Federal Reserve Bank of Philadelphia has been using LLMs to extract data from historical vehicle registrations and property deeds, which were previously too expensive and time-consuming to process at scale, opening new economic research questions.

Archive Pearl is an AI tool developed by researchers in Canada to transcribe handwritten documents in bulk. Here it shows a transcription of a leasing document from an archive in Quebec.Mark Humphries, Lianne C. Leddy, et al.

Benjamin Breen, a historian at the University of California, Santa Cruz who has been building his own AI tools for historical research, draws a distinction between who this helps most. Trained historians, he says, can already read the handwriting, meaning AI tools can augment their work but don’t transform it. The bigger change is for everyone else, like undergraduates and non-students trying to do family research. And beyond handwriting, the same models are unlocking texts that have been effectively inaccessible for different reasons entirely.

“There’s so much published in technical Latin and other archaic forms that no one reads anymore,” Breen says. “Books that you’d basically have to spend your whole life to understand.”

The Evolution of AI to Decipher Handwriting

The problem of getting computers to read human handwriting has a long history in AI. When Yann LeCun was working on it in the 1980s, neural networks were still a fringe idea, and he wasn’t even particularly interested in handwriting—he was after computer vision, but the computers weren’t powerful enough and the data wasn’t there. Handwriting was solvable, barely, because the post office had zip codes and the census had forms. “I was not particularly interested in character recognition,” he says. “It was something for which we had data.”

Since then the field has come a long way. The approach LeCun was sketching in the early 1990s—a neural network that reads an entire line of text, rather than chopping it into individual characters, and then using a language model to make sense of what the vision system sees—is essentially the blueprint that modern systems are built on.

LeCun considers the problem largely solved and has moved on to harder questions about machine intelligence with his new startup. Progress continues at the margins, however, and for specialized groups working with difficult historical documents that work still matters. “Even if the improvement is only a matter of speed, it still kind of makes new things possible that would have taken too long before,” he says. “But it’s more than speed—to be actually more reliable than people were doing.”

Humphries at Wilfrid Laurier is working on that reliability aspect. He has been building Archive Pearl, a not-for-profit tool currently in beta, designed to let researchers drag and drop hundreds of pages and get clean transcriptions back in minutes rather than weeks. The goal, he says, is democratization. “This should be a force for people,” he says.