BETA This site is in beta. Information is still being added and reviewed.
LINC Reach mental health care in your language
English
What you can actually do with AI
3

Working across languages day to day

AI can help you move between languages for everyday, low-risk tasks, with a qualified person always finishing the job. Use it to understand and to draft, and let a human own anything that matters.

The rule

Machine translation for the gist, a person for what counts

AI is fine for the gist

Reading an incoming email, drafting a routine notice, or getting the general sense of a document, always with a bilingual person checking before it reaches a client.

A qualified interpreter for what matters

Consent, rights, clinical or legal advice, safety planning, and any high-stakes conversation are never left to machine translation.

Machine translation helps staff move faster on low-stakes text; the moment the meaning carries consequences for the client, a qualified human owns it.
  • Get a gist translation of an email, a notice, or a flyer so you know what it says and can decide what to do next.
  • Draft a routine, non-legal message in a client's language, then have a qualified bilingual person check it before it goes out.
  • Ask the tool to flag any part of a translation where the meaning might not carry over, and treat those flags as places to slow down.
  • Expect the quality to be uneven across languages, and expect it to be weakest for the languages with the least text online, which are often the ones your clients speak.
  • Keep AI translation away from anything that carries legal weight or medical risk, where a small error changes what a person agrees to.
  • Treat a machine translation as a starting point, and remember a qualified interpreter is the client's right for anything consequential.
  • Watch names, numbers, dates, and negation closely, since these are common places a translation flips or garbles the meaning.

How it works, and where it breaks

The tool reads text as tokens, and languages are not charged the same

Before a model can process your text, it breaks it into tokens, small chunks that are often word fragments. English text tends to split into few tokens because these tools were built mostly on English. The same sentence in many other languages splits into far more tokens for the identical meaning, up to roughly fifteen times as many for some writing systems (Petrov and colleagues, 2023). More tokens per sentence means these languages fit less into the tool's working memory, cost more on per-use pricing, and get processed with less of the fluency the model learned from English.

Less training text means weaker, and less safe, output

A model's skill in a language tracks how much text in that language it learned from. For languages with less text online, quality drops in measurable ways. In a cross-lingual test of healthcare questions, a model gave fewer complete answers and was several times more likely to be incorrect in non-English, with the gap widening for lower-resource languages (Jin and colleagues, 2024). Its safety checks are weaker there too, so unsafe or wrong content slips through more often in exactly the languages many immigrant clients speak (Wang and colleagues, 2023). Understanding a document is a safe use of this. Anything the client acts on needs a qualified human.

Fluent output can hide a meaning error

A translation can read smoothly in the target language and still carry the wrong meaning, because smoothness and accuracy are separate. A flipped negation, a swapped date, or a mistranslated legal or medical term reads just as fluently as a correct one. This is why a qualified bilingual person is what tells you a translation is safe to send, since fluency on its own does not.

A flyer in another language The gist in English
Get the gist of a flyer written in another language

Tell me in plain English what this flyer is about, who it is for, what it offers, and any dates or contact details it lists. Note anything you are unsure about. [paste or describe the flyer text]

Guardrail. Use this to understand the flyer, and have a qualified bilingual person confirm anything you act on.

A routine message in English A draft in the client's language
Draft a routine message in a client's language for a bilingual check

Translate this routine, non-legal message into the language I name, at a simple reading level. Keep the tone warm and clear, and flag any part where the meaning might not carry over. [paste your message and name the language, no client identifiers]

Guardrail. A qualified bilingual person checks it before it goes out, and never use this for consent, legal, or medical wording.

A translation you already have A flagged check for the risky spots
Have the tool flag the risky parts of a translation for a human to check

Here is a short, non-legal message and a translation of it into the language I name. Point out any place where the meaning may not carry over, any name, number, or date that could be wrong, and any word that has no clean equivalent. List these so a bilingual reviewer knows exactly where to look. [paste both versions and name the language, no client identifiers]

Guardrail. This narrows where a reviewer looks; a qualified bilingual person still confirms the message, and never use AI for consent, legal, or medical wording. See Module 3.

A phrase that will not translate cleanly Plain options a bilingual reviewer can choose from
Get plain wording options for a term that has no clean equivalent

The English phrase I name may not have a clean equivalent in the language I name. Give me two or three plainer ways to say the same idea, explain what each one implies, and note any that could sound formal, cold, or confusing. I will have a bilingual reviewer choose. [name the phrase and the language]

Guardrail. A qualified bilingual person makes the final choice, since the tool is weakest in exactly the lower-resource languages this matters most for. See Module 3.

Worked example

The same flyer, two languages, uneven quality

  1. You paste a community flyer and ask for a plain-English gist. The tool returns an accurate summary of who it is for, what it offers, and the contact details.
  2. You then ask it to draft the flyer in a client's language that has far less text online. The draft reads smoothly, so it looks finished.
  3. A bilingual colleague reviews it and finds the date shifted by a day and a key eligibility phrase turned into something that no longer matches the program. The output was fluent and still wrong.
  4. You fix the two errors with the colleague and send the corrected version. Gisting into English was a safe use; producing text a client relies on in a lower-resource language is where a human review is not optional.

Guardrail. Never use AI or machine translation for consent, legal, or medical wording. Use a free qualified interpreter, which is the client's right. See Module 3.

Sources
  1. Petrov, A., La Malfa, E., Torr, P. H. S., & Bibi, A. (2023). Language model tokenizers introduce unfairness between languages. In Advances in Neural Information Processing Systems 36. https://arxiv.org/abs/2305.15425
  2. Jin, Y., Chandra, M., Verma, G., Hu, Y., De Choudhury, M., & Kumar, S. (2024). Better to ask in English: Cross-lingual evaluation of large language models for healthcare queries. In Proceedings of the ACM Web Conference 2024. Association for Computing Machinery. https://doi.org/10.1145/3589334.3645643
  3. Wang, W., Tu, Z., Chen, C., Yuan, Y., Huang, J., & Jiao, W. (2023). All languages matter: On the multilingual safety of large language models. arXiv. https://arxiv.org/abs/2310.00905