Why One-to-One Translation Is Dying — And What Replaces It

For 70 years, translation worked the same way: take content in English, map it word-by-word (or phrase-by-phrase) to another language, deliver the result. It was a mechanical problem with linguistic rules—one source language, one target, one direction of flow.

That paradigm is collapsing. Olga Beregovaya, VP of AI at Smartling with 27 years in NLP, watches companies still clinging to source-target mapping while the technology moves on. “Source, target, source, target… that’s not going to live for long,” she says. Instead of translating from English, forward-thinking companies are now generating content directly in each target language from scratch—a shift that fundamentally rewires how global content works.

The Source-Target Assumption Is Broken

The old workflow made sense when English was the command center. Headquarters in the US created one version of a product, marketing collateral, or documentation in English. Then that version rippled outward: translate to Spanish, to Mandarin, to Japanese. The source was king. Everything else was a derivative.

This model created two problems. First, literal translation lost cultural relevance—hence the invention of “transcreation,” where translators adapt content creatively to resonate in the target culture. But transcreation broke the database: if a one-sentence English source became three sentences in the target language, the entire translation management system fell apart. You lost the alignment between source and target chunks, which meant slower updates, more maintenance burden, and constant engineering friction.

Second, translation inherited a core assumption: accuracy means fidelity to the source. But what if the source wasn’t the best version to begin with? “Why would I want to rely on English language phenomena when the model can actually generate content that’s relevant for me in my target language?” Olga asks. If you’re selling a product in Russia, why force Russian content through English first?

Direct Generation Changes the Source

Generative AI makes the old workflow optional. Instead of translating from English, companies now feed a brief to a language model—“Marketing collateral for Russian-speaking women aged 25-40 who are time-poor professionals”—and the model generates relevant content directly in Russian. No English intermediary. No translation. No loss of cultural nuance.

This requires a mental shift. The “source” is no longer a published document in English. It’s a description: a marketing brief, a set of personas, a product specification, maybe a few examples. The model then generates from that source in each target language independently.

“What if I just get a creative brief from marketing saying my marketing collateral should reflect brand, should reflect personas?” Olga walks through the new workflow. “I feed it into the model, throw in a couple of target language examples, throw in persona information… Why would I not just generate for that persona as opposed to taking a target language translation of the English?”

The implication is radical: you’re not translating anymore. You’re creating. Each version is native-born to its language and culture, not derived from English.

The Hard Transition for Enterprises

This sounds simple until you start running a company on translation infrastructure. Most enterprises have 15-20 years of content locked in English-first systems. Their databases, their review workflows, their QA processes—all built around source-target pairs. Switching to direct generation means rethinking data architecture, retraining teams on new processes, and accepting that the “single source of truth” no longer exists.

Companies that have made the leap report better results faster. But the cost is real. “You’re dealing with 7,200 languages, and if you want to be serious about that, you need different approaches,” Olga notes. Direct generation scales to languages that were uneconomical to translate before—low-resource languages with tiny speaker populations now get content tailored to them, not cobbled from English.

But there’s a catch: the model has to be accurate and culturally aware. That’s where hallucinations creep in—especially for long-tail languages where the training data is thin. And where your source brief isn’t clear, the model fills the gaps with whatever it thinks you wanted.

What Translation Becomes

Olga is clear: “I would never say that global content will die. That’s not going to happen. The content will always be there.” But the discipline of translation—the work of mapping one language to another—is becoming obsolete in the generative era.

What remains? Two things. First, interpretation—human speech, real-time, voice-to-voice. That’s where humans will thrive longest because people still want to hear their mother tongue in real conversation. Second, quality assurance—someone has to catch when the model generates fluent-sounding nonsense or when cultural nuance gets lost. The translator’s job isn’t disappearing; it’s morphing into editor, cultural reviewer, and guardrail engineer.

The companies winning this transition aren’t trying to preserve translation as it was. They’re building systems where a brief becomes a hundred pieces of native content in a hundred languages, each one generated for its audience, each one reviewed by someone who speaks that language fluently. Translation dies. Multilingual generation lives.

FAQ

What’s wrong with translating from English as the pivot language?

Translation from English creates two problems: it forces cultural adaptation through a foreign lens (resulting in content that’s less relevant to the target audience), and it breaks data management systems when one-to-one mapping fails (a two-word English phrase becomes a five-word Spanish sentence). Direct generation skips English entirely, generating content natively in each language from the start.

How do companies transition from translation to direct generation?

Instead of handing translators a finished English document, you give the model a brief: product specs, marketing persona, tone guidelines, and a few examples in the target language. The model generates from that brief directly in the local language. This requires new data infrastructure and team retraining, but pays off in speed and cultural relevance.

Can generative models replace human translators?

For marketing and user-generated content, yes — generative models work well. For brand-critical content, legal documents, or languages where the training data is thin, the model generates a first draft but humans still review. The translator role is shifting from “producer” to “editor and quality gatekeeper.”

What happens to translation jobs?

Translation jobs in high-resource languages (English-Spanish, English-Mandarin) are shrinking. But new roles are expanding: prompt engineers who write effective briefs, cultural consultants who validate output, and QA specialists who catch hallucinations. The work is moving up the value chain.

Why does language representation in AI training data matter so much?

Models trained on billions of English tokens but only millions of, say, Swahili tokens will generate lower-quality content in Swahili. The imbalance means more hallucinations, weaker vocabulary coverage, and cultural skew toward English-speaking assumptions. Long-tail languages suffer most.

Can you mix translation and direct generation in the same workflow?

Yes. For marketing briefs and new content, generate directly. For legacy content that’s already published in English, you might still translate because rewriting from scratch is slower. Most enterprises will use both approaches in parallel during the transition.

What does “transcreation” mean and is it still relevant?

Transcreation means creative adaptation rather than literal translation—adapting content to resonate culturally in the target audience. It’s less relevant for direct generation (where you generate culture-native from the start) but still valuable when you’re adapting existing English content that wasn’t designed for global audiences.

How do you measure quality when each language gets unique generated content?

You can’t compare it to the source (there is no source anymore). Instead, you measure: does the generated Russian version perform as well as the generated English version in its market? Is the tone consistent with brand guidelines? Are there factual errors? Quality shifts from “fidelity to source” to “effectiveness in market” and “brand consistency.”

Which languages are winning this transition fastest?

High-resource languages (English, Spanish, Mandarin, German) where there’s abundant training data. Long-tail languages are lagging because models don’t have enough data to generate reliably. The cost of direct generation is lower than translation, but quality is inconsistent for languages with little representation in the training corpus.

What’s the biggest risk of direct generation?

Hallucinations, where the model sounds fluent and confident but is completely inaccurate. The risk is highest in languages with less training data and in domains (legal, medical) where errors have consequences. This is why human review—not human generation—becomes critical.