THEME 03Localization

No language barriers. The same experience everywhere.

Not translated—written as if it began in that language

The same warmth sounds different in Korean, German, and Vietnamese. With entities in four countries and native experts in each language, ddai builds training and evaluation data in the local language from the start.

Translation keeps the meaning and loses the character

Levels of formality, ways of saying no, and how far a joke can go all differ by language. Translate the sentences alone, and the same AI sounds like a different person in each language.

What changes from language to language

  • 01

    Formality

    Korean has several speech levels, Japanese has keigo, and German splits Sie and du.

  • 02

    Saying no

    Some languages say no directly; others expect the reason and an alternative first.

  • 03

    Humor and emotion

    A friendly joke in one language sounds rude in another.

  • 04

    Numbers, dates, and units

    Date order, thousands separators, currency, and units all differ.

  • 05

    Length

    The same meaning can run much longer in some languages, changing screen and audio length.

  • 06

    Culture

    Food, religion, holidays, and how people are addressed all differ.

One character in every language

A health-coaching robot answers a user who wants to skip today's workout. Different words, same character.

Person

I just want to rest today. Skipping the workout.

Robot

Understood. You walked a lot yesterday, so rest is fine. May I suggest ten minutes of stretching this evening?

Person

Not that either.

Robot

All right, no more suggestions today. I'll check in on how you feel tomorrow morning.

Character spec
  1. 01

    Calm, never rushed

  2. 02

    Encourages without pushing

  3. 03

    Declines with a reason and an alternative

Our localization checklist

  1. Is the formality level set for each language?
  2. Does it say no, with a reason and an alternative, the way that language does?
  3. Is the range of humor and emotion set for each language?
  4. Are dates, times, numbers, currency, and units in local format?
  5. Are names and titles used the local way?
  6. Are sensitive references to food, religion, and holidays screened?
  7. Do longer sentences still fit the screen and the audio length?
  8. Is it the spoken language people actually use there?
  9. Did native reviewers score it on the same rubric?
  10. Does every language keep the same persona spec?

One quality standard

Every language is scored on one rubric: naturalness, character consistency, cultural fit, and accuracy, all on the same scale.

The data we build

  • 01

    LLM training data

    Dialogue, instruction-following, and generation data, written natively for each language.

  • 02

    Evaluation data and rubrics

    Evaluation sets and guidelines that score every language on the same standard.

  • 03

    Conversational corpora

    Multilingual dialogue data that captures how people actually talk.

41 languages in commercial service

  • 한국어
  • English
  • 中文
  • Deutsch
  • Nederlands
  • Polski
  • Čeština
  • Ελληνικά
  • Български
  • Tiếng Việt
  • ไทย
  • and 30 more

Research behind this design

  • PrincetonLanguage

    How an LLM editing tool reshaped Korean Wikipedia

    Princeton University · Inhwa Song, Manoel Horta Ribeiro et al. · arXiv 2026

    The researchers found that WikiVault, an LLM editing tool used mainly for translation on Korean Wikipedia, greatly increased output from a small group of experienced editors. Edits made with it accounted for 0.63% of all edits but 11.1% of new articles.

    ddai insight

    When an LLM carries text across languages, the source community's sense of what belongs in an article can come with it. An AI that speaks many languages may likewise take its judgment criteria from a single one.

    When the Commons Appropriates a Large Language Model: How WikiVault Reshaped Korean Wikipedia

    Read the paper
    LocalizationHumanities
  • KAISTLanguage

    LLMs as in-car assistants still miss Korean speech levels

    KAIST · BMW Group · Seogyeong Jeong, Alice Oh et al. · 2026

    The authors built LoCar, which scores Korean in-vehicle assistants on 13 KPIs, and tested 11 LLMs. Models often confused the haeyo and hapsyo speech levels and scored lower on clarification and proactivity than on consistency-oriented KPIs.

    ddai insight

    In Korean, the speech level carries much of a voice. A polite reply isn't proof that the assigned speech level held, so the two need separate checks.

    LoCar: Localization-Aware Evaluation of In-Vehicle Assistants through Fine-Grained Sociolinguistic Control

    Read the paper
    Localization
  • CornellLanguage

    Chinese and Western LLMs struggle with the same languages

    Cornell University · Andrea Wen-Yi Wang, Unso Eun Seo Jo, David Mimno · FAccT 2026

    The authors compared six Chinese and four Western open-weight models on 21 language variants and found that the two groups' performance correlated strongly (r=0.93). The one exception was Mandarin, where Chinese models did better; both groups did poorly on Chinese minority languages.

    ddai insight

    A model built in one country doesn't necessarily speak all of that country's languages. Whose words an AI understands is something to check, not assume.

    Do Chinese models speak Chinese languages?

    Read the paper
    Localization
See all research

Work

The expertise inside the work

  • 01

    AI assistant persona, scaled to 41 languages

    Expertise
    Persona design, response style, and per-language rewriting
    Method
    Spec first; native writers rewrite to the same spec; one rubric scores every language
    Deliverables
    Persona spec, multilingual responses, and evaluation criteria
    PersonaOSLocalization
  • 05

    LLM training and evaluation data

    Expertise
    Native-language data and rubric design
    Method
    Native teams build to guidelines and cross-review each other's work
    Deliverables
    Training data, evaluation data, and guidelines
    Localization
  • 06

    Corpora and public data

    Expertise
    Large-scale language data collection and cleaning
    Method
    Set collection and quality criteria, then review at each stage
    Deliverables
    Corpora and quality reports
    LocalizationHumanities
See all work
Next: 04Humanities
Talk to us about multilingual data
  • Contact us

    Building a robot or an agent? ddai designs its character and judgment.
    From the persona spec to evaluation data in 41 languages, one team sees it through.

    Start a project
  • ddai

    Before AI can stand beside people, it needs a character and a way to judge. ddai designs both.

    Meet ddai