Archivists Turn to LLMs to Decipher Handwriting at Scale (spectrum.ieee.org)

Archivists are increasingly turning to large language models (LLMs) like OpenAI's GPT-4 to effectively decipher handwritten documents at scale, significantly advancing the accessibility of archival materials. Historically, reading diverse handwritten texts has posed a challenge for AI; however, recent developments show that general-purpose models are achieving higher accuracy and efficiency than traditional specialized software. For instance, research led by Mark Humphries revealed that LLMs achieved a character error rate of below 2%, compared to approximately 8% for Transkribus, the industry standard, while being 50 times faster and significantly more cost-effective. This shift has profound implications for researchers and historians, democratizing access to materials that were once labor-intensive to process. With AI adeptly handling historical documents written in various scripts and dialects, previously inaccessible stories—such as those of Indigenous women in archival records—can now be uncovered by scholars and the public alike. Institutions ranging from universities to financial entities are already leveraging these advancements, paving the way for a new era of research that enables a broader array of users, including students and non-specialists, to engage with historical texts that were once out of reach.