Correction engine·Empirical

We Added AI to Our Spell Checker and Accuracy Dropped 11 Points

The idea sounded reasonable. Let AI read the sentence, not just count letters, then pick the correction that fits the meaning best. We added it and tested it against an answer key set in advance, and the result was the opposite of what we expected.

This is a summary and another perspective on the research article.

The idea sounded good

Imagine someone typing kanus. That word does not exist in Indonesian. The spell checker has to guess what was meant, and there is more than one plausible guess if all you look at is the shape of the letters. kamus (dictionary) is one letter away, and so is kaus (t-shirt).

What can actually decide is the sentence:


Saya pinjam kanus di perpustakaan.   ->  almost certainly "kamus" (dictionary)

Saya pakai kanus warna biru.         ->  almost certainly "kaus" (t-shirt)

Our correction engine originally just counted letter shapes, blind to meaning. So we added AI at the second step, a language model called IndoBERT whose job is to read the sentence and pick whichever candidate makes the most sense. On paper the pairing works, because one side knows form and the other knows meaning.

We tested it against an answer key set in advance

To keep the result honest, we corrupted 1,500 correct words in a controlled way, so the right answer was known before the engine ever saw the data. Then we ran the engine twice with a single switch, AI on and AI off. Apart from that switch, everything else was identical.

In other words, a teacher seals the answer key before the exam sheets are handed out, so the key cannot be nudged to fit what students wrote.

With AI Without AI
Correct (overall)
77.1
88.3
Swapped letters
63.4
91.4
Deleted letters
71
79
Substitution
82
89
Doubled letters
91
94
Suggestion accuracy, with AI vs without AI

In every error class, accuracy with AI on was worse than with AI off. The worst damage was in the most mechanical class, letters swapped in place, like kamsu from kamus, which needs no understanding of meaning at all. In that class accuracy fell 28 points.

11.2 points
the drop in suggestion accuracy once AI was switched on, not a gain

The right candidate was already there, but the AI picked wrong

One extra comparison convinced us. We checked the rawest candidate list, before the AI touches it, and it was identical under both conditions at 88.5%.

Once the AI was switched on and given the authority to choose from that exact same candidate list, what reached the user dropped to 77.1%. The raw material was the same. It was the AI's choice that got worse.

It was also expensive, not just wrong

We also measured how long the process took, on 10,000 news sentences. Switching the AI off made the engine 67 times faster, with the number of successfully flagged errors staying identical under both conditions. The AI spent 98.5 percent of the entire analysis time choosing answers that turned out more often wrong than without it.

Why this went unnoticed for so long

This part is the most useful for other people, because the cause is not about Indonesian, and it is not about the AI being dumb.

We already had a monitoring tool, and it reported all clear. The AI changed the top answer in nearly half of the cases, but almost all of those changes happened between candidates that looked alike in form, so they were assumed valid. That conclusion was wrong. Our monitoring tool could only see how often the AI changed an answer, not whether the change was correct. Without an answer key set in advance, those two things cannot be told apart, and the first is far easier to make look safe.

The same direction had actually been recorded seven months earlier, in a code comment, and never followed up on. This time we measured it against an answer key set in advance, so the result could not be ignored.

What we changed afterwards

The AI step that chooses spelling corrections is now switched off by default. The switch is still there, so the decision can be re-measured at any time without touching the code.

How we corrupted the words, and all the raw data, are in the technical version of this article. We also tested the same correction engine on real writing in Passed the Exam, Failed on Real Writing.

Behind this finding

The technical version has the raw numbers, the test setup, and everything that cannot be concluded from them.

Read the technical version