LLM-powered OCR can tidy history while quietly changing it

Large language models can make historical OCR easier to read, but fluent corrections may alter spellings, names or passages that the page does not support. A new survey maps the main approaches and argues that preserving the source needs to be measured alongside transcription accuracy.

Watch Desk analysis

What happened

The survey reviews generative methods for correcting historical documents, including text-only prompting, fine-tuning, reference-guided correction and multimodal systems that use both OCR text and a page image. Its authors say no approach consistently wins: results vary with language, document type, initial OCR quality, training data, model configuration and the context available.

In a Hugging Face community article, survey author Emanuela Boros summarises the findings and points readers to the paper, OCR Post-Correction in the Age of LLMs: A Survey of Historical Document Processing. The article says the research covers methods, datasets, evaluation and reliability in work published between 2022 and 2026.

Key findings

• Fluency is not fidelity
  A smoother transcription can still modernise spelling, change rare names or add text absent from the document.
• No universal winner
  A method that works well on one collection may not transfer to another language, period or document type.
• Error rates are not the whole story
  Character and word error rates matter, but do not reveal every harmful edit or unsupported addition.
• Evaluation needs more dimensions
  The survey highlights historical spelling, visual grounding, layout, alignment and usefulness for later search or extraction.
• The evidence base has gaps
  The article says available resources are concentrated in European historical newspapers and better-resourced languages.

Why it matters

Historical archives are not ordinary text-cleaning jobs. An unfamiliar spelling or damaged line may be precisely the evidence a researcher needs, and a plausible invention can be harder to spot than a visibly garbled scan. The survey’s practical advice is to test systems on representative documents and assess whether they preserve the page, not just whether the output reads smoothly.

For archivists and developers, the choice is not only which model to use. It is also whether to correct existing OCR, use the image to guide correction, or transcribe directly from the image. The survey says that decision depends on factors including image availability, OCR quality, language, layout and operational cost.

Our read

This is a useful reminder that “better text” and “truer text” are not synonyms. Before deploying an OCR correction system, compare it with trusted examples from the collection and look specifically for changes to already-correct text, historical forms and unsupported additions. The model may be fluent; the archive still gets the final say.

What to watch

• Whether future evaluations compare text-only correction, image-guided correction and direct transcription on the same documents.
• Whether more multilingual datasets and varied archival material become available.
• Whether evaluation reports measure source fidelity and layout alongside character and word error rates.

Discussion spark: When evaluating AI-corrected historical OCR, should preserving the source take priority over readability, even when that leaves more errors visible?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.