AI Reads Old Documents Well — But Sometimes Invents Names

Computer programs that read old scanned documents — called OCR, or optical character recognition — have gotten a big upgrade lately. Newer AI models, known as vision-language models, can look at a page image and type out what it says, much like a person would. Researchers in Uruguay tested these AI readers against older, traditional software on a tough collection: dictatorship-era documents scanned from microfilm, where the ink is faint and the paper is damaged. By the usual scorecards, the AI won clearly, making fewer letter and word mistakes.

Then the researchers read the transcripts themselves, line by line. What they found was unsettling. The AI often "cleaned up" old-fashioned spelling, added words that were never on the page, and swapped in similar-sounding words that kept sentences reading smoothly while changing the meaning. Errors on names and places were the most dangerous, because they barely move the accuracy score while completely changing who or what a document is about.

Why does this matter outside a research lab? Historical archives are how societies prove what happened. Governments, courts, historians, and families searching for disappeared relatives all rely on these records. If an AI transcription reads smoothly and scores well but quietly renames a person or alters a date, the error can sit undetected for years. Traditional software tends to fail visibly — garbled nonsense you notice. AI failures are fluent, which makes them harder to catch.

The study's honest limitation: it examined one dataset, in one language, from one country's history. But the lesson travels. The researchers argue that accuracy scores based on individual letters are not enough, and that anyone using AI to digitize records needs human review of names, dates, and key facts — especially before publishing archives or using them as evidence.