Skip to main content
AI Data

Can You Train AI on Machine-Translated Data Instead of Collecting Native Data?

Lifewood Team12 min read

Short answer. Partly. Machine-translated data works as a volume supplement for pretraining, for classification and extraction, and for bootstrapping a first instruction-tuned checkpoint in data-poor languages. It weakens on open-ended generation, and it is absent by construction on local knowledge. It should never be used for evaluation: a benchmark built across 31 languages found locally written and English-translated versions of the same questions differing by more than 20 points for the best models. Translate for breadth, collect natively for behaviour, and always source test sets natively.

Key takeaways

  • Machine-translated training data is a legitimate supplement, not a substitute: workable for pretraining volume, classification and bootstrapping, weaker for open-ended generation, and absent by construction on local knowledge.
  • Translated data carries three distinct defects — translationese that models learn and reproduce, English-centric content that translation cannot localise, and machine translation errors that become training supervision.
  • The direct native-versus-translated fine-tuning study finds the difference shows up on native and generation benchmarks, and that regularisation helps on structured tasks but not generative ones.
  • Locally authored and translated test sets can differ by more than 20 points for the best-performing models, so a translated evaluation set can certify a model that native users find broken.
  • Translation quality propagates measurably: in a diagnostic study of machine-translated benchmarks, datasets with lower COMET scores showed a higher share of accuracy and mistranslation errors at span level.
  • The cheapest insurance in multilingual AI is a natively collected evaluation set. It is a small share of the data budget and the only part of the stack that can tell you the truth about the rest.

Quick comparison

Role in your stackTranslated dataWhy
Pretraining volume (data-poor languages)Acceptable supplementBreadth matters more than naturalness; curation still required
Classification and extraction fine-tuningOften workableLabels survive translation better than style does
Instruction tuning, first bootstrapWorkable, with narrow mitigationsRegularisation helps on structured tasks; the generation gap remains
Open-ended generation, dialogue, stylePoor substituteTranslationese and content mismatch dominate the output
Local-knowledge tasksFails by constructionThe source data never contained the target market's knowledge
Evaluation setsDo not useLocally written and translated tests disagree by more than 20 points in one 31-language benchmark

Why is this the first question in every multilingual scoping call?

Because the arithmetic looks irresistible. Machine translation into a dozen languages costs a small fraction of recruiting native speakers, writers and reviewers in a dozen markets. If translated data were equivalent to native data, native collection would be waste, so every team asks whether it can skip it.

You have a proven English dataset of instructions, conversations or labelled examples, and the question deserves a straight answer rather than a vendor reflex in either direction. Both extreme positions are wrong. "Translation is free multilingual data" ships products that embarrass you in market. "Only native data counts" prices smaller languages out of coverage entirely and ignores genuine research showing translated data pulling real weight in the right roles. Multilingual models are exposed to translated text at scale during training already, because native data at scale does not exist for most languages.

Two terms frame the decision. Machine-translated training data is text produced by translating existing data, usually English, into a target language with an automatic translation system. Native data collection is the sourcing of text, speech or labels directly from speakers of the target language, written or recorded in that language from the start.

This article is the decision, not the slogan: what translation preserves, what it destroys, and which parts of your data stack can tolerate the difference. It is the production-method companion to our piece on why local knowledge matters, Beyond Translation: Why AI Needs Culturally Relevant Data, which covers the content side of the same divide.

What actually goes wrong with translated training data?

Three distinct defects: translationese, English-centric content, and injected translation errors. It pays to separate them, because they respond to different fixes and each hits different tasks differently.

Translationese is the measurable difference between translated text and natively written text, including source-language sentence structures, unnatural collocations and a flattened style. It appears in human and machine translation alike, and it is not a quality bug that a better engine removes; it is a property of translation itself, documented across decades of corpus research. Models trained on it learn it. Work on whether large language models have an English "accent" measures the related failure directly: multilingual models produce non-English text carrying English-centric patterns in both vocabulary and grammar, which is what native users hear as a product that sounds foreign in their own language even when every word is correct.

English-centric content is the second defect. Translation moves the words and keeps the world. A translated instruction set asks about American holidays, Anglophone institutions, imperial measurements and English-language cultural references, now rendered in fluent Thai or Swahili. The knowledge your market actually asks about — local entities, local procedures, local assumptions — is simply absent, and no translation step can add what the source never contained.

Error injection at scale is the third. Machine translation makes mistakes, unevenly by language pair and domain, and in a training pipeline those mistakes become supervision. The clearest measurement of this comes from quality-assurance work on translated benchmarks: a diagnostic study of the EU20 benchmark suite found that datasets with lower COMET scores exhibit a higher share of accuracy and mistranslation errors at span level, with the trend consistent across the suite. What corrupts an evaluation set corrupts a training set by the same mechanism, just less visibly.

The three defects stack differently by task. A classification label survives translationese fine; an open-ended generation task absorbs all three at once.

What does the research say about training on translated data?

More nuanced than either camp admits, which is why the actual comparisons are worth reading. Translated data can come close to native data on some task types, while the difference opens up on generation and on any test that is itself native.

The most direct study is Is It Good Data for Multilingual Instruction Tuning or Just Bad Multilingual Evaluation for Large Language Models? by Pinzhen Chen, Simon Yu, Zhicheng Guo and Barry Haddow. It fine-tuned models on native versus machine-translated instruction data and evaluated both ways, and its findings draw the usable map:

  • The difference is real but task-dependent. Native or generation benchmarks reveal a notable difference between native and translated instruction data. On other task types the gap is modest, which is why translation keeps earning a place in bootstrapping.
  • Your choice of benchmark decides what you see. The study's framing — good data, or merely bad evaluation? — is the warning itself: a translated test set is the wrong instrument for answering this question, because it shares the artefacts of translated training data.
  • Mitigations are narrower than people hope. The paper reports that regularisation is beneficial for bridging the gap on structured but not generative tasks. It does not offer a fix for the generation case.
  • Round-trip construction is a diagnostic, not a remedy. Its comparison of round-trip against single-pass translation points to the importance of knowledge from language-native resources rather than to a translation technique that substitutes for it.

Zoom out to pretraining and the picture is permissive but conditional. Translated text adds volume in languages with thin corpora, and that is a real use. But multilingual web-data audits are a standing warning about what uncurated scale buys: one manual audit of 205 web-crawled language corpora found entire datasets without usable text, a result we unpack in Can More Data Make AI Worse? Translation multiplies whatever quality you feed it, in both directions.

Why is translated evaluation data a different, worse problem?

Because training on translated data costs you quality, while testing on it costs you the ability to see the quality loss. A measurement instrument that flatters the thing it measures is worse than no instrument at all.

An evaluation set is a held-out collection of examples used to measure how well a model performs, kept separate from training data. The evidence here is hard to argue with. MultiLoKo, a local-knowledge benchmark built across 31 languages with both locally authored and English-translated versions of its questions, reports differences of more than 20 points for the best performing models between the two. A translated benchmark is therefore not a noisy version of the real test; it is a different test, measuring translation-shaped competence.

A WMT 2025 empirical analysis of machine translation for expanding multilingual benchmarks reaches a complementary conclusion from the dataset-design side: current language selections in multilingual datasets tend to overestimate large language model performance on low-resource languages. Published in the Tenth Conference on Machine Translation proceedings, it is the peer-reviewed anchor for a claim that otherwise circulates as folklore — that the multilingual numbers in a model card are optimistic in exactly the languages where you most need them to be honest.

The practical rule follows directly, and it is absolute in a way the training-side rule is not: whatever you decide about translated training data, your evaluation sets must be natively sourced. They are the smallest dataset in the stack and the one place the economics of translation were never compelling anyway. Our guide to building multilingual evaluation sets for LLMs reaches the same conclusion from the buyer's side: an English-built evaluation stack applied to other languages scores well only because it never tested them properly.

Where does the translate-versus-collect line actually sit?

It sits by task role and market stakes, not by budget line. Translate for volume, classification and bootstrapping; collect natively for generation, dialogue, local knowledge, regulated domains, and always for evaluation.

Translate when: you need pretraining-scale volume in a language with thin corpora; you are bootstrapping a first instruction-tuned checkpoint to have something to iterate on; the task is classification or extraction where the label, not the prose, is the product; or the language is in a later rollout wave and translated data buys a usable interim while native collection runs. Apply the mitigations that are actually evidenced: quality-filter the output with translation-quality scores, since error rates track those scores measurably, and document which data is translated so it can be weighted down or retired later.

Collect natively when: the task is open-ended generation, dialogue or anything users read as voice; the domain is regulated or safety-adjacent, where fluent error is the expensive failure; the content requires local knowledge the English source never had; the language is strategically core rather than long-tail coverage; or, always and without exception, the dataset is for evaluation. Native collection is also where speech lives almost entirely — accents, acoustics and code-switching do not translate, which is why speech programmes run as collection from the start, in the managed-pipeline shape described in How a Speech Data Collection Programme Actually Runs.

Plan the ratio to shift. Healthy multilingual programmes typically start translation-heavy for breadth and move native-heavy as each language proves its market value, so language prioritisation and the translate-to-native schedule should be decided together. Translated data is scaffolding. The mistake is not using scaffolding; it is forgetting to build the building.

How does Lifewood draw the line in client programmes?

Lifewood sits on the native-collection side of this market, so the honest way to use it in this decision is for the parts of the stack where the research says native data is non-negotiable: evaluation sets first, then generation-facing fine-tuning data, then speech.

Lifewood Data Technology offers multilingual data collection across 100+ languages with native contributors through managed delivery centres. In practice, scoping conversations frequently reduce the native-collection ask a client arrives with. Where a language is long-tail coverage and the task is classification, translated data plus a natively collected evaluation set is the defensible configuration, because an evaluation set collected natively will show precisely when the translated training data stops being good enough. That measurement is worth more to the relationship than the larger invoice. Buyers comparing vendors on this basis can use our guide to choosing a multilingual data collection partner.

The translation side of the split has its own mature supply: machine translation providers and the language-service industry handle translated-data production at scale, and the academic groups cited throughout supply the independent evidence no vendor on either side should be grading alone. Whoever you buy from, the configuration to insist on is the same — translated data clearly labelled as such in your pipeline, and a native evaluation set standing outside it, holding the ruler.

What should you do if your budget only covers translation?

Protect the evaluation set first, then filter, label and schedule the translated data for replacement. A constrained budget changes the order of spending, not the principle that tests must be native.

  • Protect the evaluation set anyway. It is typically a small fraction of total data volume. Collect it natively even if everything else is translated; it is the difference between a cheap programme and a blind one.
  • Quality-filter the output with automated translation-quality estimation and a native-speaker spot-check, and discard the bottom band rather than training on it. The error profile tracks the quality score closely enough to act on.
  • Keep translated data labelled in your pipeline so it can be down-weighted or replaced per language without archaeology.
  • Spend the first native-collection money on the highest-variance tasks — open generation and dialogue — rather than spreading it evenly.
  • Do not count on clever translation pipelines to close the generation gap. The evidence points to native resources rather than to a translation technique, so treat elaborate round-tripping as a way to use native seeds well, not as a substitute for having them.
  • Re-test on the native evaluation set after every model update. The translated-training gap moves with model generations; measure it rather than assume it.
  • Schedule the graduation. Write into the roadmap which languages convert from translated to native data at which trigger — usage threshold, revenue, error rate — so scaffolding does not become structure by default.

Frequently asked questions

Safer on error rate, not on the structural problems. Translationese appears in human translation too, and English-centric content survives any translator. Human translation narrows the quality gap at a cost that often approaches native authoring, at which point native authoring usually wins because it also brings local knowledge.

Post-editing removes errors and softens the worst translationese, and it is a reasonable middle rung. It cannot add local knowledge the source lacked, and heavily post-edited data at scale costs enough that you should price native creation against it honestly before committing to either route.

That is a different trade with different risks, including the recursive-training degradation covered in Is It Safe to Train AI Models on AI-Generated Data? Generating natively in-language avoids translationese, which is why some teams prefer it for instruction data, but it still needs native human review.

Native-speaker review of a random sample is the reliable method. Classifier-based detection exists in the research but tends to overfit to its training data. If reviewers consistently describe the text as "correct but foreign-sounding", you have your answer, and you should label that data as translated in your pipeline.

Translated text can script speech collection prompts, but the audio itself must be natively recorded, because accent, acoustics and code-switching are the product. Text-to-speech from translated scripts compounds both problems at once, producing audio that is neither natural nor representative of real speakers.

Small in money, fixed in priority. Native evaluation sets are small datasets, so for most programmes they cost a low share of the data budget. Fund them first, before the training-data split is even argued, because they are the only part of the stack that measures the rest.

Lifewood Data Technology, founded in 2004, runs multilingual data collection across 100+ languages with native contributors through managed delivery centres, covering text, speech and annotation. Its scoping practice is to protect the native evaluation set first, then generation-facing fine-tuning data, then speech.

Sources and further reading

  1. Is It Good Data for Multilingual Instruction Tuning or Just Bad Multilingual Evaluation for Large Language Models? (arXiv:2406.12822) — Chen, Yu, Guo and Haddow's systematic native-versus-translated fine-tuning comparison: native and generation benchmarks reveal a notable difference, regularisation helps on structured but not generative tasks, and the round-trip comparison points to the importance of language-native resources
  2. MultiLoKo: a multilingual local knowledge benchmark for LLMs spanning 31 languages (arXiv:2504.10356) — Hupkes and Bogoychev, reporting differences of more than 20 points for the best performing models between locally authored and English-translated test data
  3. An Empirical Analysis of Machine Translation for Expanding Multilingual Benchmarks (WMT 2025) — Rajaee, Choenni, Shutova and Monz, in the Tenth Conference on Machine Translation proceedings: current language selections in multilingual datasets tend to overestimate model performance on low-resource languages
  4. Do Large Language Models Have an English "Accent"? Evaluating and Improving the Naturalness of Multilingual LLMs (arXiv:2410.15956) — Guo and colleagues on English-centric patterns in the vocabulary and grammar of multilingual models' non-English output
  5. Diagnosing Translated Benchmarks: An Automated Quality Assurance Study of the EU20 Benchmark Suite (arXiv:2604.01957) — Thellmann, Stadler and Färber, finding that datasets with lower COMET scores show a higher share of accuracy and mistranslation errors at span level

Have an AI or visibility project in mind?

From AI evaluation and human-in-the-loop review to GEO and AEO strategy, our team can help you deploy with confidence and get found in the AI search era.

Talk to our team