It looks like a gold mine
Every time someone fixes a typo on Wikipedia, a pair of data gets saved: before and after, free, at scale, complete with a timestamp. For anyone who needs "wrong word -> right word" examples to train a spell checker, that sounds like a ready-made data source.
We mined it. Got 744 correction candidates. Then checked them one by one.
That's all that survived. Nine out of ten candidates turned out not to be real corrections.
Why nine out of ten failed
Our initial assumption was wrong: we thought "someone changed A to B" was enough of a signal for a correction. It wasn't.
Most were just style edits, one word swapped for another equally correct one. Some were cases where both forms are valid, and the editor simply happened to prefer one. Others actually introduced an error, which was only fixed a few edits later. Mined raw, we'd have learned the mistake, not the fix. The slipperiest category: proper names that genuinely have two accepted spellings. Changing one isn't a correction, it's just preference.
We read each candidate alongside its original sentence, and decided one thing only: whether the "after" form was genuinely more correct, not merely different. The bar was set high on purpose, since anything that survives becomes a permanent reference for judging whether a future code fix is actually a fix. A wrong reference is more dangerous than having no reference at all, because it locks in the mistake and makes a genuinely correct fix look like a regression.
Why this matters beyond Wikipedia
This is the part we think is most useful for people to know, since it often gets left out of conversations about automation: mining data is cheap and fast, but deciding what's actually correct is expensive and slow. It demands someone who knows the language, reads the context, and makes a decision they can stand behind.
AI can help filter the first layer, narrowing 744 down to fewer items that need a human check. But it can't yet be trusted to make the final call alone, because if it's wrong, no one finds out until the mistake is already locked in as a reference.
This 5.8 percent figure isn't a report card for Indonesian Wikipedia. What we tested is how fit its edit history is as material for automatic correction, and the answer: fit, provided 94 percent of what's mined gets discarded first.
How we checked each candidate, and all the raw data, are in the technical version below.
Behind this finding
The technical version has the raw numbers, the test setup, and everything that cannot be concluded from them.