Translation methods
LLM translation vs machine translation vs human translation.
Specialised translation models, LLMs and human translators solve different problems. The right choice depends on the material, language pair, deadline and cost of a mistake.
“AI translation” can mean a specialised translation model or a general-purpose LLM. They behave differently. Some workflows still need a person.
The basic distinction
Different methods, different trade-offs.
Neural machine translation (NMT) is built mainly to move text between languages. It is often chosen for predictable, high-volume work; the service and language pair still matter.
Large language models (LLMs) can translate while following instructions, reading supplied context and using reference material. They can also produce wording absent from the source, so fluent output still needs checking.
Human translators and interpreters can research an unfamiliar term, ask the speaker what they meant and explain a disputed choice. Their work still depends on subject knowledge, preparation, time and working conditions.
Grounding changes the workflow
A translation system can also be connected to a glossary, translation memory or approved source collection. When the speech contains a known quotation or fixed term, the system retrieves that record and carries its identifier into the output. A reviewer can then see where the wording came from.
Side by side
What each method is good for.
| Question | Specialised NMT | LLM translation | Human translation |
|---|---|---|---|
| Wider context | Usually works within the amount of context the product was designed to accept. | May read longer passages, follow instructions and use supplied reference material. | Can ask questions, research the subject and resolve ambiguity. |
| Consistency | Often predictable at high volume for established language pairs. | Wording can change with the prompt, surrounding context or model version. | A style guide helps, although different translators can still make different choices. |
| Culture and nuance | Performance depends heavily on the training coverage for that language and expression. | Can use context, yet recent studies still find problems with idioms, wordplay and applying cultural knowledge. | A translator with the right cultural experience can interpret intent and local expectations. |
| Speed and scale | Often selected for high-volume work where speed and cost matter. | Can be fast enough for many uses; latency varies by service and prompt. | Costs more and takes longer, especially for live or high-volume work. |
| Traceability | Traceability depends on the product keeping logs, glossaries or source records. | A generated explanation does not prove where the wording came from. Retrieval can provide a real source record. | The translator can record sources and explain a disputed choice. |
| High-stakes use | Use qualified review when an error could cause serious harm. | The reviewer still needs the relevant subject and language expertise. | |
Recent evidence
What recent studies found.
Longer inputs did not remove every problem.
A 2024 study found that LLM-based translation improved on long sentences and could handle documents of up to 512 words in its tests. Rare words and material outside the model's familiar domains still caused problems. The authors also identified efficiency, low-resource languages and human-aligned evaluation as open questions. Read the study.
Preparation changes the output.
The MAPS framework analysed keywords and topics before asking an LLM to translate. It also selected among candidate outputs. On the language directions and tests used in the paper, this reduced hallucinations, mistranslations, awkward phrasing and omissions. Read MAPS.
Recognising cultural knowledge did not ensure its correct use.
CanMT found a persistent gap between recognising culture-specific knowledge and applying it correctly in a translation. The translation strategy changed the result, and supplying a reference translation made the LLM-based evaluation more reliable. Read CanMT.
Idioms and puns scored lower.
“Be My Cheese?” asked native speakers to rate translations into 15 languages. Across the seven LLMs tested, the mean full-text score was 1.68 out of 3. Idioms averaged 1.65 and puns 1.45. A sentence can read smoothly while the joke or local meaning has already disappeared. Read the study.
Choosing a workflow
Ask what happens if the translation is wrong.
Reading for yourself
Automated translation may be enough for personal comprehension, a rough draft or live access where the listener knows mistakes are possible.
Publishing at scale
Use automation for the first pass, enforce the required terminology, then review the text before publication.
Quoted or fixed wording
Retrieve the approved text where possible, keep its identifier attached and check how the quotation fits the surrounding speech.
High-consequence content
Use a qualified reviewer with the right subject knowledge. Decide in advance when that person must correct or take over from the system.
Baian's current approach
How Baian handles this today.
Baian sends recent speech and the selected languages into the translation step instead of treating every new phrase on its own. For supported Qur'an workflows, likely Qur'anic Arabic can be matched against an indexed collection. A match can carry the surah and āyah reference and stored target-language wording where available.
Other authorised source collections would need their own records, permissions, matching rules and evaluation.
Baian's production source matching is currently limited to the Qur'an.
Sources
Sources for this guide.
- 01 Culture-Aware Machine Translation in Large Language Models: Benchmarking and Investigation (CanMT, 2026)
- 02 “Be My Cheese?”: Cultural Nuance Benchmarking for Machine Translation in Multilingual LLMs (2026)
- 03 Salute the Classic: Revisiting Challenges of Machine Translation in the Age of Large Language Models (2024)
- 04 Exploring Human-Like Translation Strategy with Large Language Models (MAPS, 2023)
About these notes
Written by Baian from the linked research. We have not rerun these papers’ experiments, and this is not an independent review of our product.
Need to discuss a translation workflow?
Send us the language pair, the setting and the wording that matters most.