The Language Barrier That Has Kept AI From Serving Most of Africa Just Got Smaller

Real-time speech translation across 70 languages, working in noisy environments, already on Android phones. The capability gap that made AI disproportionately useful to English speakers has narrowed significantly — and the governance and commercial implications for South Africa are immediate.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/the-language-barrier-that-has-kept?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

Twenty years ago, Google’s translation experiment began as one of its pioneering machine learning projects. Today, more than one trillion words are translated for billions of users across Google’s products every month. On 9 June 2026, Google took what it describes as its next step: the release of Gemini 3.5 Live Translate, an audio model that delivers near real-time speech-to-speech translation across more than 70 languages. Unlike older systems that wait for a speaker to finish before producing a translation, Gemini 3.5 Live Translate generates translated speech continuously, staying only a few seconds behind the original speaker while preserving their intonation, pacing, and pitch. It automatically detects the languages being spoken without requiring manual configuration. It works in noisy environments — cars, airports, open offices, classrooms, clinics, and meetings. And it is available today, at no additional cost, in the Google Translate app on Android and iOS.

For most of the world, this is a convenience upgrade. For Africa — a continent of 2,000 languages where the dominance of English in AI tools has functioned as a structural barrier to equitable access — it is something more consequential.

CONTEXT AND BACKGROUND

The language problem in AI has been present since the first large language models were trained predominantly on English-language text. The consequence was predictable and well-documented: AI tools worked better in English than in any other language, produced outputs of lower quality in African languages, and required English fluency as a practical prerequisite for accessing the most capable AI systems. For most Africans, this was not a minor inconvenience. It was the difference between AI being a useful professional tool and AI being a technology that served someone else.

I have previously written about one specific manifestation of this problem — the way AI detection tools flagged South African students writing in their second or third language at disproportionately high rates, penalising non-native English speakers for the same writing characteristics that made AI detection unreliable in the first place. That article was about the consequences of building AI systems with English as the assumed default. The Gemini 3.5 Live Translate release is the most concrete challenge to that default that any major AI company has produced at this scale and at this level of accessibility.

The previous version of speech translation in Google Meet supported five languages and could only translate to and from English — meaning every conversation was forced through an English intermediary, and any language pair that did not include English was simply not served. Gemini 3.5 Live Translate changes this structurally. It supports more than 70 languages and over 2,000 language combinations in a single meeting, meaning two participants can speak Swahili and Zulu, or Portuguese and Amharic, without either of them being required to switch to English as the translation medium. The constraint that made multilingual AI tools systematically less useful to African users has been removed at the architecture level, not worked around at the product level.

INSIGHT AND ANALYSIS

The commercial proof of concept for this capability at scale already exists, and it comes from a context more immediately relevant to Africa than most AI product announcements. Grab, the Southeast Asian ride-hailing and payments platform, is testing Gemini 3.5 Live Translate for communication between drivers and passengers at pickups. Grab users make more than 10 million voice calls per month through the app. The deployment addresses a genuinely difficult operational problem: drivers and passengers who do not share a common language need to communicate in real time, in noisy environments, under time pressure, without the ability to type. The AI translation layer removes the language requirement from the transaction entirely.

The South African parallel is direct and specific. South Africa has 11 official languages and a population in which a significant majority do not conduct their working lives, receive healthcare, or access public services primarily in English.

The implications of real-time AI translation for South African organisations are sector-specific and immediate. In healthcare, a nurse in a rural clinic who speaks Zulu communicating with a patient who speaks Sesotho, or a doctor in a township facility who trained in English conducting a consultation with a patient whose primary language is isiXhosa — these are daily operational realities whose quality of outcome is directly constrained by language barriers that AI translation can now meaningfully address. In public service delivery, the requirement to conduct official business in English systematically disadvantages citizens whose primary languages are one of the other ten official languages, and whose ability to access services, understand their rights, and navigate bureaucratic processes is constrained accordingly. In commerce, the multilingual workforce and customer base that most South African organisations manage requires communication infrastructure that real-time translation can now support in ways that were previously either unavailable or prohibitively expensive.

The SynthID watermarking that Google has applied to all Gemini 3.5 Live Translate audio output is worth noting for governance reasons. Every translation produced by the model carries an invisible watermark that identifies it as AI-generated audio. This matters because the EU AI Act’s Article 50, requiring synthetic content labelling, comes into force on 2 August 2026. Google’s implementation puts the tool ahead of that regulatory requirement for European users. For South African organisations deploying the tool, the watermarking also provides a basic provenance mechanism — a way to establish that a recorded translation was AI-generated rather than human-interpreted, which has implications for the evidentiary status of translated recordings and the accountability chain when a translation is disputed.

The absence of the English pivot is the most architecturally significant change in this release. Previous translation systems that required routing through English were not merely inconvenient for non-English speakers. They introduced a specific and measurable quality degradation for any language pair where neither language was English — because the model was performing two translations rather than one, compounding the error rates of each step. Removing the English pivot is not a product enhancement. It is a structural correction to a design choice that systematically disadvantaged non-English speakers, and it is the change that makes the tool genuinely useful in the multilingual contexts that describe most of Africa’s actual working environment.

IMPLICATIONS

For South African boards and executives, the Gemini 3.5 Live Translate release raises three specific questions that most current AI strategy frameworks have not yet addressed.

The first is about workplace AI inclusion. Most South African organisations have been evaluating AI tools primarily on capability and cost. The language dimension — whether the tool functions with equivalent quality for employees and customers across all 11 official languages — has rarely been a primary evaluation criterion. It should be. An AI tool that works well in English and poorly in Zulu, Sesotho, or Afrikaans is not a productivity tool for the organisation as a whole. It is a productivity tool for the English-speaking portion of the organisation. The Gemini 3.5 Live Translate release provides a benchmark for what language inclusivity in AI tools should look like — and it provides grounds for asking whether the other AI tools in the organisation’s stack meet the same standard.

The second is about public service and regulated industries. South African organisations operating in healthcare, financial services, and public administration have obligations under the Constitution, the Official Languages Act, and sector-specific regulations to serve South Africans in their preferred official languages. AI tools that support those obligations — that make it practically possible for a financial adviser to explain a complex product to a Sesotho-speaking client in real time, or for a healthcare worker to take an accurate patient history from a Tsonga-speaking patient — create compliance value as well as commercial value. The organisations that integrate AI translation into their service delivery frameworks will be better positioned than those that continue treating language access as a manual interpretation problem.

The third is about the governance framework for AI-generated translation in high-stakes contexts. When an AI translation system is used in a medical consultation, a legal proceeding, a financial advice session, or a public service interaction, the question of accountability for translation errors becomes specific and significant. The SynthID watermark establishes that the translation was AI-generated. It does not establish what standard of accuracy the AI translation is held to, who bears liability when a translation error produces a harmful outcome, or what the right of the affected party is to request a human interpreter. These are governance questions that South African organisations should be developing frameworks for now, before the deployment scale of AI translation makes them urgent.

CLOSING TAKEAWAY

The language barrier that has kept AI from serving most of Africa is not one barrier. It is a compound of many — the training data gap that made models worse at African languages, the interface design that assumed English fluency, the product architecture that routed multilingual conversations through English intermediaries, and the economic reality that professional human interpretation was too expensive to scale across the continent’s linguistic diversity. Gemini 3.5 Live Translate does not eliminate all of those barriers. It removes several of the most concrete ones at once, at a price point of zero for Android users, with a deployment breadth that includes Google Meet, Google Translate, and the Gemini Live API available to any developer who wants to build on it.

The governance questions that follow — about accuracy standards, liability, workplace inclusion, and the treatment of AI-generated translation in high-stakes contexts — are the ones that South African boards and executives need to be developing frameworks for now. The technology has arrived. The governance response should not wait until the deployment has scaled to the point where the absence of that framework becomes visible in the consequences.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net