The Translation Extinction

How AI Could Silence Languages Without Ever Erasing Their Words

I. When Translation No Longer Requires a Translator

A Kenyan teacher told a conference that she had been hopeful about an AI learning application an international non-governmental organization brought to her school. Then she asked it to help explain a local proverb about trees that her grandmother had taught her, and it had no answer. The account comes from researcher James Maisiri, writing in The UNESCO Courier. The episode illustrates how a system may handle language while missing the cultural knowledge that gives a proverb its meaning.

Meanwhile, the technology keeps advancing. On September 4, 2026, Google announced that Android users can keep live translation running in the background and that iPhone users can hear it through the earpiece, across more than 70 languages. In May 2026, OpenAI published a developer guide for GPT-Realtime-Translate, a speech-to-speech translation model that accepts more than 70 input languages and supports 13 output languages. For supported language combinations, live speech translation can help two people converse without sharing a language, although the available input and output languages are not the same. If people can communicate without learning one another’s languages, what will persuade the next generation to learn and use a smaller language?

Three phenomena are often conflated. Here, “translation extinction” is a working term for the possible displacement of routine human translation, not an established scholarly label for language loss. Linguistic marginalization is the declining status, visibility or practical use of a language in settings such as education, technology and public administration. A language is extinct when it has no remaining speakers; later revival efforts may seek to restore its use. The loss of intergenerational transmission is a major warning sign, though the timing and circumstances of language shift vary.

The evidence examined here does not show that AI is causing a global wave of language extinction. In US data, machine translation is associated with slower growth in translator employment and lower demand for foreign-language skills, while digital representation remains uneven. Neither strand directly measures whether families stop transmitting a smaller language; together, they identify a risk that requires testing rather than an outcome already demonstrated. If investment and reliable tools concentrate in major languages, smaller-language communities may face stronger pressure to use languages that technology serves better.

II. The Languages Missing from the Machine

Effortless translation does not mean equal language support. The gap between linguistic diversity and digital representation predates generative AI. But as translation and AI tools become part of everyday services, languages missing from those tools risk being left out of more than conversation. UNESCO’s Global Roadmap for Multilingualism in the Digital Era, published in 2025, highlights a stark imbalance: more than 7,000 languages are spoken worldwide, but only about 1,000 are represented online.

A language’s survival does not depend on dictionary size or the existence of a translation app. Under UNESCO’s vitality framework, intergenerational language transmission is the first factor listed, alongside shifts in domains of use, responses to new domains and media, and other indicators of language vitality.

Language coverage is not the same as language equality. Even when a model accepts speech in a language, its output options may not include that language. A supported-language list alone cannot establish how well a system performs across accents, dialects, code-switching or noisy environments. If educational and workplace tools work more reliably in dominant languages, those languages may become the easier choice in settings where access matters. Whether this convenience changes the languages children learn or speak at home remains an open research question.

III. The Hidden Economics of Linguistic Dependence

Uneven language support also raises a question about economic value, and one part of it is already measurable: the effect of machine translation on translators. In a 2025 study summarized on VoxEU, economists Carl Benedikt Frey and Pedro Llanos-Paredes examined the rollout of Google Translate’s mobile app across US local labor markets. Using Google Trends searches for Google Translate as a proxy for local uptake, the study estimated that a 1-percentage-point increase in its measure of machine-translation use was associated with about 0.7 percentage points slower growth in translator employment. The authors estimated that this effect corresponded to roughly 28,000 fewer new translator positions over 2010–2023. They also report reduced demand for foreign-language skills. The study does not test whether machine translation changes the languages families use at home. Its findings therefore cannot establish whether easier translation affects intergenerational language transmission.

The employment outlook is not a picture of a profession disappearing. The US Bureau of Labor Statistics reports 73,900 interpreter and translator jobs in 2025 and projects 2% growth by 2035, slower than the average for all occupations. These figures do not explain how the work may change, but they caution against treating slower growth as proof that the occupation is vanishing.

One possible downstream risk is that weaker demand for language services could reduce the resources available to people and institutions that maintain written materials in smaller languages. The labor-market study does not test this possibility, so it remains a question for future research.

IV. Five Countries, Three Cases and Two Open Questions

The labor evidence is measurable, but what these tools mean for speakers differs by country and language. Each country below contributes a different kind of evidence, and not all of it is equally strong.

India: public-service access. The Government of India’s February 2026 document reports that the BHASHINI platform supports text services in 36 languages, hosts more than 350 AI models and datasets, and has completed more than 4 billion language inferences. These figures show the scale of deployment, not how well the systems perform for each language or dialect. The unresolved question is whether broader access will translate into dependable service for speakers of smaller languages.

Nigeria: resource and performance gaps. In a February 2025 survey of research on Hausa, Yorùbá and Igbo, languages spoken by over 175 million people, Isa Inuwa-Dutse found that only about 25.1% of reviewed studies contributed new linguistic resources. A May 2025 study at the LoResMT workshop by Ebelechukwu Nwafor and Minh Phuc Nguyen created benchmark datasets for Igbo-English and Nigerian Pidgin-English translation. On the tasks tested, specialist low-resource translation models outperformed the GPT-based models they evaluated. The findings point to a practical gap: stronger support for Nigerian languages will require not only access to general-purpose models, but also local data and evaluation that can show how well those models perform.

The United States: labor-market substitution. The US case supplies the clearest measured economic evidence in this comparison. It does not establish whether easier translation changes which languages students learn or families use at home.

China and Indonesia: evidence gap. The available sources do not support a comparable assessment of language-level AI performance or changes in everyday language use in either country. These cases remain research questions, not evidence of a demonstrated effect.

V. A Language Can Survive on Paper and Die in Practice

Deployment figures cannot tell us whether a system understands a culture or whether a community still speaks its language. Two sources illuminate different limits: one concerns cultural context, the other the language of translated text. Neither demonstrates that AI has silenced a community.

The first concerns cultural mismatch. Fluent output can still miss cultural context. In the Courier article, Maisiri reports that researchers asked ChatGPT and Google Gemini how many seasons exist. The systems answered “four,” an answer that did not reflect the wet-and-dry seasonal framework described in the article’s West African example. He also notes that elders and community members co-teach in some schools in Kenya, Namibia and South Africa, sharing proverbs and oral tradition.

The second concerns translated text. In an April 2024 corpus study of Chinese-to-English texts published in Humanities and Social Sciences Communications, Jiang Niu and Yue Jiang compared machine translations, human translations and English originals across three genres. In that corpus, machine-translated texts had lower lexical diversity than human translations and English originals across the three genres tested. Genre also influenced the results. The study does not establish whether these textual differences change how communities speak or whether they affect language vitality over time. Low-resource languages can also receive weaker support when training data and reliable evaluation benchmarks are limited, a concern that the systematic review of low-resource translation and UNESCO’s roadmap both address.

These examples reveal limits in cultural understanding and translated writing. Their implications for language vitality remain an open empirical question.

VI. The Machines That Could Save What They Threaten

These risks are not inevitable. AI systems that reduce the need for some translation tasks could also help document, teach and support endangered languages. Whether they do so depends on the data, design and priorities behind the tools. The design goal matters. A substitution system asks how a tourist, executive or student can operate without learning another tongue. A preservation system asks how speakers can document, learn and use their own language across civic and economic life.

In an August 2026 systematic review in Technologies, Sweeta Agrawal and Abayomi O. Agbeyangi reviewed 63 studies selected from 1,696 records retrieved across Scopus, Web of Science and Google Scholar on neural machine translation and large language models for low-resource languages. The review identifies five major approaches: data augmentation, back-translation, transfer learning, pre-training and parameter-efficient fine-tuning. These methods can help researchers work around scarce training data, although their effectiveness still needs to be assessed for each language and task.

Data work is under way. According to the Courier, language specialists with the African Next Voices project have digitized 9,000 hours of spoken languages from Kenya, South Africa and Nigeria, creating a dataset intended to support the development of AI tools for local languages. Its practical value will depend on the languages represented, the data’s terms of access and how well resulting systems perform. A larger dataset is an input, not a preservation outcome. Its value depends on whether speakers can use the resulting tools for real tasks, whether outputs are accurate in local contexts, and whether the community retains meaningful control over the data.

Capability also raises a governance question: who owns the digitized voice of a community? In 2025, an Independent Expert Group convened by UNESCO published recommendations on AI, cultural rights and diversity, including the need for governance that protects cultural sovereignty. For language data, the sources cited here do not settle questions of consent and control. A defensible principle is community governance: speakers should have a meaningful role in setting data-access terms, validating transcriptions and deciding which recordings may be shared or withdrawn.

VII. Who Will Speak for the World’s Languages?

Technical capability is not enough without community authority and everyday use. A company’s count of supported languages cannot show whether communities are gaining or losing the ability to use their languages. That requires longitudinal research on school language choices and household transmission, language-specific accuracy testing, and transparent audits of how language data are licensed and governed.

The evidence examined here does not establish that AI translation causes languages to die. If investment continues to favor a small set of dominant languages while tools for low-resource languages remain less reliable or less sensitive to cultural context, smaller-language communities may face greater pressure to use languages that technology serves better.

A language can survive in recordings and still retreat from daily life. AI could help bring it into classrooms, public services and new forms of communication, but only if the tools serve the people who speak it. The decisive test is not whether a machine can translate a language. It is whether the next generation still has reasons and opportunities to speak it.

References

1. Maisiri, James. “African languages, the blind spot of AI.” The UNESCO Courier, April 2, 2026. https://courier.unesco.org/en/articles/african-languages-blind-spot-ai
2. Kapur, Sasha. “Google Translate rolls out new upgrades for iOS and Android.” Google, September 4, 2026. https://blog.google/products-and-platforms/products/translate/google-translate-ios-android-upgrades/
3. Kettleson, Erika. “Build Live Translation Apps with gpt-realtime-translate.” OpenAI Cookbook, May 7, 2026. https://developers.openai.com/cookbook/examples/voice_solutions/realtime_translation_guide
4. UNESCO. Global Roadmap for Multilingualism in the Digital Era: Advancing the Role of Language Technologies. 2025. https://unesdoc.unesco.org/ark:/48223/pf0000393920
5. UNESCO Ad Hoc Expert Group on Endangered Languages. Language Vitality and Endangerment. 2003. https://ich.unesco.org/doc/src/00120-EN.pdf
6. Frey, Carl Benedikt, and Pedro Llanos-Paredes. “Lost in translation: AI’s impact on translators and foreign language skills.” VoxEU/CEPR, March 22, 2025. https://cepr.org/voxeu/columns/lost-translation-ais-impact-translators-and-foreign-language-skills
7. US Bureau of Labor Statistics. “Interpreters and Translators.” Occupational Outlook Handbook. https://www.bls.gov/ooh/media-and-communication/interpreters-and-translators.htm. Accessed October 19, 2026.
8. PIB Research Unit, Press Information Bureau, Government of India. “Artificial Intelligence for Culture and Languages.” February 9, 2026. https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/feb/doc202629784101.pdf
9. Inuwa-Dutse, Isa. “NaijaNLP: A Survey of Nigerian Low-Resource Languages.” arXiv, February 27, 2025 (v1). https://arxiv.org/abs/2502.19784v1
10. Nwafor, Ebelechukwu, and Minh Phuc Nguyen. “Fostering Digital Inclusion for Low-Resource Nigerian Languages: A Case Study of Igbo and Nigerian Pidgin.” Proceedings of the Eighth Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2025), May 2025, pp. 44–53. https://aclanthology.org/2025.loresmt-1.6/
11. Niu, Jiang, and Yue Jiang. “Does simplification hold true for machine translations? A corpus-based analysis of lexical diversity in text varieties across genres.” Humanities and Social Sciences Communications, 11, article 480, April 2024. https://doi.org/10.1057/s41599-024-02986-7
12. Agrawal, Sweeta, and Abayomi O. Agbeyangi. “Bridging the AI Language Divide: A Systematic Review of NMT and LLMs in Low-Resource Translation.” Technologies, 14(8), 518, August 21, 2026. https://doi.org/10.3390/technologies14080518
13. UNESCO Independent Expert Group on Artificial Intelligence and Culture. Report of the Independent Expert Group on Artificial Intelligence and Culture. 2025. https://www.unesco.org/en/articles/new-expert-report-explores-how-ai-transforming-culture
    PDF:
    https://www.unesco.org/sites/default/files/medias/fichiers/2025/09/CULTAI_Report%20of%20the%20Independent%20Expert%20Group%20on%20Artificial%20Intelligence%20and%20Culture%20%28final%20online%20version%29%201.pdf