What AI Translation Leaves Out

2026, Nov 17 . AI Translation

What AI Translation Leaves Out

The failure mode worth worrying about in AI translation is not the one that looks wrong. It is the one that looks right.

Here is an ordinary sentence from a maintenance manual:

"Wear gloves, eye protection and a respirator when handling the compound."

Here is what came back:

"Wear gloves and eye protection when handling the compound."

Read the second sentence again and try to find the defect. There isn't one. It is grammatical, idiomatic, correctly punctuated, and the terminology is right. A two-item list is a perfectly normal thing for a sentence to contain. Nothing in the output invites suspicion.

This is what makes omission different from every other translation error. Most errors leave evidence. A truncated segment ends mid-clause. A dropped subject breaks agreement. A missing connector leaves an orphan "however" hanging at the start of a paragraph. Those are visible, and a reviewer catches them without ever looking at the source.

Omission in a modern model leaves no evidence, because the model's objective is fluent, well-formed output. When something falls out, the sentence is rebuilt around the gap. The seam closes. What you receive is not a damaged sentence. It is a different, intact sentence that happens to say less than the original.

Now consider what your quality controls actually test.

Spellcheck passes, obviously. Grammar check passes. Terminology check passes, because every term that survived is correct. Length ratio checks are the interesting one: English into most European languages expands anyway, so a segment that quietly lost twenty percent of its content can land comfortably inside normal variance. And a post-editor reading only the target passes it too, because there is genuinely nothing to notice.

Every routine control most workflows run is a fluency control. Omission is an adequacy failure. Fluency controls cannot detect adequacy failures. That is not a criticism of anyone's process, it is a structural mismatch, and it is worth naming because most quality conversations skip straight past it.

Then there is the part that gets uncomfortable when the request is multilingual.

Most companies checking AI output check the languages someone in the building can read. Usually English, often Spanish or French. Those come back clean, and the conclusion becomes "the model is good." That confidence is then extended, silently, to Greek, Romanian, Finnish, Hungarian.

Two problems with that. The first is that those are precisely the directions where models are weakest, because that is where the training data is thinnest. Research covering more than a hundred language pairs found hallucination markedly worse in low-resource directions, and worse again when translating out of English. The second problem is that "we found no problems" was never what happened. What happened is that nobody looked.

This is not a fringe research concern, incidentally. Meta's HalOmi work built a manually annotated benchmark for hallucination and omission across eighteen translation directions, and its central finding is one every buyer of multilingual content should absorb: conclusions drawn from a single language pair largely do not hold at multilingual scale. Your Spanish sample tells you about your Spanish. It tells you nothing about your Greek.

So what actually helps?

Check adequacy, not fluency. That means someone reads the source and target together, counts what the source asserts, and confirms each assertion survived. It is a different task from reading for style, and it has to be briefed as one.

Automate the things that carry meaning: numbers, units, negations, conditionals, list item counts, cross-references. Cheap to implement, and it catches a real share of this.

Treat each language as its own risk rather than a sample from one pool. Coverage is per language, or it is not coverage.

And accept the honest version of the trade. AI translation works. It genuinely does, and anyone telling you otherwise is not paying attention. It works as a draft. The remaining cost is not polishing the language. It is verifying that the meaning is all still there, in every language you shipped, including the ones you cannot read.

That cost is real; it is smaller than most people fear, and it keeps getting left out of the estimate.

This article is part of our series on where AI translation fails in regulated content. For the full map, see The Risks of AI Translation in Regulated Content. To see how we combine AI speed with professional review, see Machine Translation with Human Post-editing.

References

Dale, Voita, Lam, Hansanti, Ropers, Kalbassi, Gao, Barrault & Costa-jussà, HalOmi: A Manually Annotated Benchmark for Multilingual Hallucination and Omission Detection in Machine Translation, EMNLP 2023. https://aclanthology.org/2023.emnlp-main.42/

Guerreiro, Alves, Waldendorf, Haddow, Birch, Colombo & Martins, Hallucinations in Large Multilingual Translation Models, TACL 2023. https://aclanthology.org/2023.tacl-1.85/

Share:

Speaking to the world with confidence is easy!

1. Request a quote 
2. Validate the terms of the proposal 
3. Receive accurate translations 
You are using an outdated browser. Please upgrade your browser to improve your experience.