Spell-check AI output without code-word false positives
Language models occasionally produce typos, dropped letters or invented
spellings — and ordinary spell checkers are useless on AI text because they
flag every technical term as an error. This checker runs entirely in your
browser and combines a base English dictionary with a built-in technical
word list (ML, AI and programming vocabulary) so it catches genuine
misspellings while leaving words like tokenizer, async and embeddings
alone.
Why LLM output still needs spell-checking
A common assumption is that language models are perfect spellers. In practice they are very good but not infallible. Specific situations where typos and misspellings slip through:
- Low-frequency proper nouns. When a model generates a person’s name, a product name, or a place name it encountered rarely in training data, it sometimes transposes letters or applies a plausible but incorrect spelling.
- Technical neologisms. Newly coined terms (model names, framework names, recent acronyms) may not appear enough in training data for the model to have a reliable spelling, and it will produce phonetically reasonable but wrong variants.
- Long generations under high temperature. As generation continues over thousands of tokens, the probability of an individual token error accumulates. A single malformed token becomes a misspelled word.
- Transliterated terms. Words from non-Latin scripts that have multiple valid English transliterations (names, technical terms borrowed from Chinese, Arabic, Russian) are inconsistently handled.
- Copy-editing context. If you are pasting existing text into a model for rewriting and the model slightly alters the spelling of a term while paraphrasing, the change is easy to miss in a read-through.
How it works
The text is split into word tokens. Anything that looks like code — tokens with digits, underscores, dots or slashes, camelCase identifiers, content inside backticks, and URLs — is skipped, because those are not natural-language words and would only create noise. Each remaining word is lowercased and checked against two sets: a common-English base list and a curated technical dictionary. Words found in neither are flagged. You can extend the technical list with your own custom terms (product names, acronyms, jargon) so they pass on every subsequent check.
Building a useful custom word list
The custom allow-list is what makes this checker valuable for repeated use on the same project. Add:
- Your product and company names
- Internal acronyms and jargon
- Framework and library names specific to your stack
- Proper nouns (names of people, places, organizations) that appear regularly in your content
- Domain-specific vocabulary the base dictionary does not cover
Once a term is in your custom list, it passes silently on every subsequent check without generating noise.
Tips and notes
Use this as a final proofreading pass on generated documentation, marketing copy or release notes before publishing — it is fast to scan a flagged-word list and confirm whether each is a real typo. Because the base dictionary is intentionally compact (it favours precision over coverage), rare but valid English words may be flagged; add them to your custom list when that happens. This tool checks spelling only, so pair it with a grammar pass if you need full editorial review.