LLM translation validator
When an LLM translates copy, you usually cannot read the target language well enough to know whether it stayed faithful. Back-translation is the classic trick: translate to the target language, then translate the result back to the source. If the round trip drifts, the forward translation probably did too. This tool runs both legs with your own OpenAI key and shows them side by side.
What back-translation catches
Back-translation is not perfect, but it reliably surfaces the most common failure modes in machine translation:
Semantic drift. A sentence that changes meaning in the forward translation will produce a back-translation that says something different from the original. For example, “we look forward to your arrival” may come back as “we await your appearance” — fine — or “we are expecting you to arrive” — a subtle flattening of the warmth.
Dropped clauses. A conditional or qualification — “provided that the payment clears” — is sometimes omitted in the forward translation. The back-translation returns a sentence without it, making the omission visible.
Hallucinated additions. A model sometimes adds context to “help” the translation read naturally in the target language. These additions come back in the back-translation as new claims not present in the source.
Idiom literalisation. English idioms that do not have direct equivalents are either literalised (producing nonsense) or localized (changing the image used). A back-translated idiom that reads literally (“to bite the bullet” back from German might become “to take the pain”) is a signal to check the forward translation.
How it works
On run, the tool sends two sequential requests to
https://api.openai.com/v1/chat/completions. The first asks the model to
translate your source text into the chosen target language and return only the
translation. The second feeds that output back in and asks for a translation
back into the source language. Both calls use temperature 0 for stable,
repeatable results, and everything happens directly between your browser and
OpenAI — there is no Gera backend in the loop.
Limitations of back-translation
Back-translation can miss errors that cancel out: a mistranslation in the forward direction that is repeated symmetrically in the reverse direction will produce a back-translation that matches the source even though the forward translation is wrong. This is uncommon but real — back-translation narrows the problem space rather than eliminating it. For legally or commercially important translations, it should be paired with a native-speaker review.
Tips and notes
- Read for meaning, not wording. Synonyms and reordered clauses in the back-translation are fine. Watch for changed numbers, dropped clauses, or new claims that were not in your original.
- Test idioms and product names. These are where translation most often breaks; a back-translation that literalises an idiom is a clear warning.
- Short segments isolate problems. Validating one sentence at a time makes it obvious which line drifted, instead of one long blurry paragraph.
- Pair with a native reviewer for anything shipping. Back-translation narrows where to look but does not replace a human who reads the target language.