AI's "Why It Decided" Explanations Are Broken for Arabic and Hebrew Readers
If AI can't explain itself in your language, you can't argue back.
When an AI decides something about you — flags your email as spam, rejects your loan application, hides your post — software like SHAP and LIME draws a little picture showing which words pushed it that way. Think of it as a receipt for the AI's thinking. These receipts were built for English and other left-to-right languages, and that turns out to matter more than anyone expected.
Applied to Arabic, Hebrew, Urdu or Persian, the maths behind the explanation stays correct, but the picture falls apart. Letters that should join together break apart, words appear out of order, and the layout runs backwards from how people actually read. The researchers measured this with a computer-vision read-back test on 200 words per language. Default charts scored character error rates of 0.82 to 0.98 — essentially unreadable. A popular fix-it trick made Urdu worse, at 0.998, and a newer charting library flipped the problem again.
Their solution, SHAP-RTL, is a rendering layer that corrects reading direction and letter shapes while leaving the AI's original reasoning, word order and scores untouched. It picks the right font per language and can also rewrite the same explanation as short sentences in the reader's own language.
One honest caveat: the study only checked whether the pictures are readable, not whether the explanations underneath are actually good or fair. That's left for future work. Still, the core point is simple. There are hundreds of millions of Arabic, Hebrew, Urdu and Persian speakers. If an automated system denies you something, an explanation you cannot read is no explanation at all — and this work is a step toward giving it back.
- SHAP and LIME are AI "explainers" that highlight which words drove a decision — but their charts only display correctly in left-to-right languages like English.
- In tests on Urdu, Arabic, Hebrew and Persian, default charts had 82–98% character errors, and a common workaround made Urdu hit 99.8% — worse than doing nothing.
- The new SHAP-RTL fix keeps the AI's original reasoning intact while repairing fonts, letter shapes and reading direction, and can restate the explanation in plain sentences.
Why It Matters
If you can't read why an AI rejected you, you can't appeal it — this restores that for millions.