Sitting in rooms where a wrong word can derail a negotiation. Working as a translator at the SADC Secretariat, I learned one thing fast: in African institutional settings, close enough is never an option. The right word in Kinshasa lands differently in Johannesburg. A phrase that carries weight in Paris can fall completely flat in Lusaka. That lived reality is why I went to the people who have been doing this work for decades. What I found confirmed what I always suspected. AI can translate. It still cannot localise. Not here. Not yet.
“AI cannot understand humour. It cannot understand cultural nuances, and it does not live as a human among people.” Anonymous Lecturer, 20 years in translation and interpreting
The Gap Is Not the Same for Every Language
A university lecturer with nearly 20 years in translation, interpreting and teaching put it plainly. High-resource languages like English, French and German have enormous volumes of text for AI models to learn from. South African spoken languages do not. The AI has simply not seen enough of them to perform reliably.
Even Afrikaans, the most digitally resourced of South Africa’s spoken languages, still needs human review. For isiZulu, Sesotho and Setswana, the gap widens. Add internal language variation and the problem sharpens further. Setswana spoken in South Africa is not identical to Setswana spoken in Botswana. A human who has lived in those communities knows that immediately. An AI model without enough training data cannot.
FIELD REALITY Fewer than 5% of African languages have sufficient digital resources for effective AI language processing. (Google/WAXAL, 2026) [Source]
When the Machine Misses the Brief
The lecturer shared a translation that illustrates the problem clearly. A school commissioned him to translate their code of conduct with one specific brief: remove all gendered and patriarchal language. The school had replaced the word “uniform” with “multiform”, reflecting a policy where any student could wear any item of clothing regardless of gender. To understand why gender-neutral language in localisation matters, the context has to travel with the words.
The AI neutralized pronouns. It updated generic references. But it had no idea what a “multiform” was. It had no access to the school’s values or the thinking behind that word. The context lived in people, not in a training corpus. A human had to carry that meaning across.
He gave a second example. A newspaper article listed band names and song titles. The AI translated all of them, including names that should never have been touched. A human would have known those were fixed references. The AI did not. These are not rare edge cases. They are the routine cost of removing the human from the process.
“Lobola translated as ‘dowry’ is technically correct. It is also profoundly incomplete.”
The Balance Africa Must Get Right
Johan Botha, Director of Folio Online in Cape Town and co-founder of the Association of Language Companies in Africa (ALCA), speaks from nearly four decades in the industry. He is not anti-AI. He is clear-eyed about what it can and cannot do.
Africa largely skipped the neural machine translation era because the quality for African spoken languages was never good enough. Generative AI changed that conversation fast. ALCA’s recent survey found adoption above 80% among African language service providers. But the key detail is this: most of that adoption is in workflow management, not translation itself. Professionals use AI to handle admin, file formatting and quality checks so they can focus on the actual language decisions, which still require a human.
“Today will be the worst that AI will be,” Botha said. Each day it improves. However, he also named the future role clearly: translators will be caretakers between technology and language. Clients will bring a localization need and no idea how to meet it. The professional who holds both the technical and cultural knowledge will be the one they trust.
This balance is what top-ranking articles on AI translation consistently miss. They debate whether AI will replace translators in general. They rarely ask what happens in lower-resource, culturally layered languages spoken by communities with oral traditions that predate any written corpus. In those contexts, the human is still doing the real work.
What You Need to Bring to This Profession
The lecturer was direct. Do not enter language practice without deep knowledge of your languages and cultures. Not vocabulary lists. The living knowledge that tells you what sounds right to a community, what register suits a formal document, and what an AI will miss every time.
At the same time, refusing AI tools is professional self-sabotage. “If you don’t use AI as a translator, you’re an idiot,” he said. Learn your CAT tools, machine translation software and how to prompt AI well. Speed and output volume matter. But never confuse a fast draft with a finished localization.
For clients who believe AI translation of South African language content is ready to publish: it is not. Not for formal communication, not for public-facing content, not for any message a community needs to trust. At a minimum, get a human to review it. Preferably, one who has lived the culture.
Want to build the skills AI cannot replace? Explore the TCLoc Master’s Programme at the University of Strasbourg mastertcloc.unistra.fr/admissions/apply-now/
