Translation and communication with communities: speed, quality and accountability

1Machine translation today

Claude and Gemini translate very well between languages widely present on the Internet (English, French, Spanish, Portuguese, Modern Standard Arabic, and many others) for everyday texts: emails, procedures, reports, terms of reference. They can also adapt the register (formal, simple, technical).

Quality drops markedly:

• for less-resourced languages: many national and local languages of Africa, Asia and Latin America, some of them very widely spoken;
• for dialect varieties (spoken Arabic dialects, creoles, regional varieties);
• for very technical texts (medical, legal) or very cultural ones (proverbs, humour, taboo topics);
• for proper names and local terms.

The main danger: a fluent translation looks correct even when it contains a mistranslation. The more fluent the text, the less we think of checking it.

Fluency is not proof of accuracy. Any translation intended for affected people is reviewed by a native speaker who knows the context.

2Review proportionate to risk

Not all translations require the same level of control. A simple grid:

• Low risk (internal understanding, reading a document in a foreign language, informal email between colleagues): translation by AI, quick review by the user.
• Medium risk (reports, internal procedures, correspondence with partners or donors): review by a person fluent in both languages.
• High risk (health messages, informed consent, information on rights and access to aid, complaints mechanisms, legal documents, security messages): professional translation or in-depth review by a qualified native speaker, and testing with community members.

A useful method: back-translation. Another person (or another AI conversation, as a complement) translates the text back into the original language; the two are compared to spot differences in meaning. It helps spot mistranslations, but guarantees neither naturalness nor cultural appropriateness.

3Glossaries and terminological consistency

Humanitarian terms have precise and sometimes sensitive equivalents: "internally displaced person", "refugee", "asylum seeker", "survivor", "household", "multipurpose cash transfer". Inconsistent translation creates confusion and sometimes legal or protection risks.

• Build a validated bilingual or multilingual glossary (starting from existing agency and cluster glossaries where possible).
• Give it to the assistant: "You must use this glossary; if a term is not in it, flag it."
• Place it in a Claude Project or a Gem so that it serves the whole team.
• Update it with feedback from translators and communities.

Example. In Arabic, the translation of "internally displaced person" must remain constant across all documents; in French, "survivante" is preferred to "victime" in programmes to combat gender-based violence, according to the guidelines used by the organisation.

4Writing for communities

Communication with communities and accountability to affected populations (Core Humanitarian Standard commitments) require messages that are understandable, useful and respectful. AI helps to:

• simplify a technical text into plain language (short sentences, common words, one idea per sentence);
• adapt the format: poster, 30-second radio message, voice message, 160-character SMS, script for facilitators;
• propose answers to frequently asked questions ("Who is eligible? When? Where can I complain?").

But only communities can say whether a message is understood and acceptable. Essential steps:

• test the message with a few people of different profiles (women, men, young people, older people, people with disabilities);
• check preferred channels and literacy rates (assessment data, existing language profiles);
• provide non-written formats (audio, images, oral presentation);
• always state how to ask a question or make a complaint.

5Interpretation, voice and images

The Claude and Gemini mobile apps offer voice modes, and other tools provide real-time interpretation. They can help in a simple conversation, but:

• for a protection interview, a medical consultation, a registration interview or a complaint, a trained human interpreter is needed (confidentiality, ethics, nuance);
• the voice and content of the conversation are processed by the provider: the same data rules apply as for text;
• synthetic voices and generated images for community messages must be labelled as such and validated for cultural acceptability.

For community radio, a script prepared with AI, reviewed by native speakers and recorded by a human voice from the community generally inspires more trust.

6Rumours and false information in communities

During epidemics and crises, rumours spread fast ("vaccines make you sterile", "aid is reserved for one group"). Rumour tracking mechanisms collect this information; AI can help to:

• group reported rumours by theme (from anonymised data);
• prepare factual answers based on official information (Ministry of Health, WHO), in simple language;
• propose adapted formats (radio questions and answers, messages for community relays).

Each answer is validated by the technical specialist (health, protection) and tested. A clumsy answer can reinforce a rumour. Week 10 covers online disinformation and fake content.