The writing is on the wall for handwriting recognition

Real‑world performance and limits

• Several commenters report being “blown away” by current OCR/LLM capabilities compared to the 1990s, especially on messy modern handwriting and personal notes.
• Others find results “hit and miss”: mixed-language diaries, bad handwriting, and non-English text often degrade performance.
• Users working through family letters say models are impressive for transcription and summarization, but still miss lines, hallucinate phrases, and require full human verification.

Historical documents and non‑English scripts

• Historical hands (secretary hand, Carolingian minuscule, Roman cursive, cuneiform, Gothic/Danish, 18th‑century Dutch, fraktur/blackletter) are seen as far from “solved,” largely due to scarce training data.
• Russian cursive becomes a test case: models do surprisingly well even on “doctor’s cursive,” but still misread key medical phrases and diagnoses; older church records quickly expose limitations, especially with names and locations.
• Some specialized systems (e.g., for Japanese manuscripts or Russian archives) achieve low character error rates using large, targeted datasets.

LLM vs “pure” OCR and hallucinations

• A recurring concern: LLMs don’t just recognize characters, they rewrite text, substituting plausible words instead of faithfully transcribing—unacceptable for archival or scholarly use.
• One commenter traces the continuum from character models to language models: as context windows expand (pairs, words, sentences), you inevitably drift into language modeling.

Training data, contamination, and confidence

• Suspicion that famous historical letters were part of model training; others counter that models also do well on private, never-digitized material.
• Discussion of token-level confidence: with downloadable models you can use low-confidence markers to focus manual review; commercial APIs often hide logprobs.
• A workaround is to ask the model to flag low-confidence words, with mixed expectations about reliability.

Open‑source and self‑hosted options

• People seek local, trainable solutions for private notebooks. Suggestions include Tesseract, TrOCR (with tricky version pinning), surya‑v2, nougat, and various vision-capable LLM weights used in ensemble fashion.
• For difficult historical handwriting, several commenters say Gemini 3 is the first general model to give “decent” results.

Future of handwriting and cognition

• Debate over whether handwriting itself is dying vs. protected by the “Lindy effect.”
• One side cites research claiming handwriting engages more brain regions and improves memory and idea formation; others say the main effect is higher cognitive load that can hurt comprehension during note-taking.
• Some imagine an ideal future of writing freely on paper with near‑perfect digitization; others point out keyboards are still faster.

Cultural and societal reflections

• Nostalgia for beautiful 19th‑century penmanship and concern that modern signatures show declining personality and care.
• Broader thread about whether AI productivity gains will free people for “thinking and walks” or just intensify competition and work, with references to education shortcuts, mental laziness, and capitalism’s incentives.