Most "multilingual AI" is a translation layer. You write a prompt in English, get an answer in English, and a translation step renders it into the user's language at the end. It works well enough to look impressive in a demo and badly enough to be useless for the thing it claims to do.
This post is about why, and about the pipeline we built instead.
Why the translate-at-the-end approach fails
Consider what a sentence carries. "That market is a bubble" is six words and a claim. It assumes the reader knows what a bubble is, that markets can have them, that the metaphor comes from soap, and — crucially — that the speaker is making a warning rather than a description.
Translation preserves the words. It does not preserve the assumptions. A reader who has never been handed the soap-bubble metaphor receives a sentence about a market and a bubble, and has to do the metaphor work themselves, in their second language, while also following the argument.
The cost of that is not zero and it is not small. It is cognitive load: the mental effort spent decoding the surface rather than understanding the content. Load that a native reader never pays. This is the mechanism behind a familiar experience — reading a page in a second language, understanding every word, and finishing with no idea what the argument was.
So the failure is not that the translation is wrong. It is that the translation is correct and still doesn't land.
The four stages
Our pipeline has four stages, and the useful thing about naming them is that each has its own distinct failure mode.
1. Concept extraction
Before anything is explained, the underlying claim has to be separated from the language it arrived in. What is actually being asserted? What does it depend on? What would make it false?
The failure mode here is inheriting the source's framing. If the English explanation was organised around an example the reader has no access to, extracting the "concept" while keeping that example produces a re-explanation that is just as foreign as the translation was.
2. Cultural mapping
The concept gets mapped onto something the reader already has — a metaphor, an institution, a familiar situation. Where the source says "a margin call", the re-explanation may need to start from a situation the reader has actually lived through.
The failure mode is the wrong metaphor. A mapping that is culturally adjacent but structurally different is worse than no mapping, because it teaches the wrong thing fluently. A loan metaphor that maps onto a moneylender rather than a bank produces an explanation of a different financial instrument.
3. Re-explanation
The mapped concept is written out in the reader's language, at the depth the original had.
The failure mode is depth collapse. Re-explanation is genuinely harder than translation, and the pressure to make it readable pushes toward simplification. The output gets clearer and shallower at the same time, and because it reads better, nobody flags it.
4. Depth verification
Which is why there is a fourth stage. The output is checked against the source for what it lost: dropped nuance, hedged claims made absolute, conditions silently removed.
The failure mode is verification that only checks fluency. A grader that reads the Marathi output and asks "does this read well?" will pass a hollow answer every time. The question has to be "does this say the same thing, with the same conditions, and would a domain expert accept it?"
What changed when we measured it
Two things surprised us, and both are worth stating because they contradict the intuition most teams start with.
Longer is not deeper. Our first version of depth verification rewarded completeness, and the model responded by padding. Word count went up; information content did not. The metric had to be rewritten to score what a domain expert would call load-bearing — the specific claims, conditions and exceptions that the answer depends on.
Fluency and faithfulness diverge, and fluency wins if you let it. In side-by-side evaluations, reviewers consistently preferred the more fluent of two answers before being shown the source. Shown the source, they often reversed. That is an ordering effect, and it means any evaluation where the grader sees only the output is measuring the wrong thing.
What we do not claim
Two honest limits.
First, cultural grounding is not a solved problem, and it is not solved by us. The mapping stage is where the residual risk lives, and a wrong metaphor is harder for a reader to detect than a bad translation, because it reads naturally.
Second, we do not have native-speaker coverage for every language we accept. Where we do, the re-explanation is checked by someone who thinks in that language. Where we do not, the output is a best effort and we would rather say so than imply otherwise.
Three objections worth taking seriously
"Isn't this just translation with extra steps?"
No, and the difference is testable. Take an answer, translate it, and separately re-explain it from the concept. Then ask a native speaker of the target language, who does not read English, which one they understood — and ask a domain expert whether either one lost a condition. Translation tends to pass the first test and fail the second. Re-explanation fails the first test more often than we would like and passes the second. That trade is deliberate: an answer nobody trusts is worse than an answer that takes slightly more work to follow.
"Doesn't simplifying for a local audience lose precision?"
It can, which is exactly why depth verification is a separate stage rather than a hope. The rule we work to is that precision lives in the claims, conditions and exceptions — not in the jargon. You can rebuild the explanation around a moneylender instead of a bank and keep every condition intact. You cannot drop the condition and compensate by sounding more confident.
"Won't a good model eventually handle all of this automatically?"
Partly, and that is the goal. But "eventually" is doing a lot of work in that sentence, and the people who need the tool now are not served by waiting. The more useful framing is that the evaluation has to exist before the capability can be claimed. A model that produces fluent Marathi nobody has graded is not a solved problem, it is an unmeasured one.
The principle underneath it
Translation asks: how do I say this in your language?
Grounding asks: what would have to be true for this to make sense to you, and do you already have that?
The second question is harder, slower, and more expensive. It is also the only version that actually removes the barrier rather than relocating it to the reader's working memory.
Found this useful?
Share it with someone who reads better in their own language.