Дата публикации: 13.10.2026
Raimbek Maulenov
Author info: Bachelor’s Student, English and American Studies, Faculty of Humanities, University of Debrecen, Debrecen, Hungary
Author info: Bachelor’s Student, English and American Studies, Faculty of Humanities, University of Debrecen, Debrecen, Hungary
Abstract. This article examines the connection between artificial intelligence and low-resource and endangered languages, asking whether digitalization preserves or weakens them. It discusses the uneven quality of AI models across languages, using the examples of Bengali, Kurdish, Tamil, Arabic, Kazakh and Kyrgyz, and explains the role of limited digital data. The article concludes that AI can help preserve languages only if large amounts of quality material in those languages are regularly added to digital sources.
Key words: Artificial intelligence, endangered languages, low-resource languages, digitalization, language preservation.
In this topic, we will determine the connection and compatibility between AI and languages. We will try to answer the question of whether digitalization preserves or weakens low-resource languages. We will consider the impact of Artificial Intelligence on languages.
This issue can also be studied from two sides. There are many languages in the world. It is known that Artificial Intelligence cannot learn languages that are not available in its information base by itself. At the same time, Artificial Intelligence may increase linguistic inequality. For example, many AI models work well in English, Chinese, Spanish, and German. However, AI may not be able to work reliably with some languages that have little text and limited high-quality data. In low-resource languages, such models may produce incorrect translations and may not fully convey the specific features of the language. I think this may be a problem. To answer the question of what should be done in this situation, let us rely on researchers’ articles and studies.
There have been some problems and discussions in society related to this issue. For example, a person may ask an AI a question in another language, but because the AI cannot correctly identify the language, it may answer in another similar language. In 2023, journalists from Rest of World tested ChatGPT in Bengali, Kurdish, Tamil, and several other less widely used languages. The journalists found that in some languages, the AI created made-up words, made logical mistakes, and sometimes gave meaningless answers. At that time, this issue increased public discussion about differences in AI quality between languages and about AI’s limited ability to fully understand some languages. More specifically:
-
it created words that did not exist;
-
it made many grammatical mistakes;
-
it confused one language with another;
-
it gave completely incorrect answers to logical tasks;
-
it produced meaningless texts in that language even for simple questions.
This may lead to a loss of the value of low-resource languages. This is because many people are increasingly choosing to learn new languages through AI. If AI makes many grammatical mistakes when dealing with other languages, people who are learning those languages may accept these mistakes as correct.
In connection with this, I also looked at another study. In the Nature journal, Mohamed L. Seghier, in his article “ChatGPT: not all languages are equal,” asked the same questions in English, French, and Arabic, and the results were not the same. I learned that this raised a scientific discussion about whether AI works at the same level of reasoning in all languages.
ChatGPT:
for the question in English — a good answer,
for the question in French — a slightly weaker answer,
for the question in Arabic — an even weaker answer.
Then we cannot simply say:
“ChatGPT knows three languages well.”
For a user or a researcher, what matters is not only whether the model recognizes a language, but also whether it can provide high-quality information and reason logically in that language.
A similar situation, where AI could not distinguish between two languages, also occurred between Kazakh and Kyrgyz. When Kyrgyz people asked ChatGPT questions in their own language, the AI sometimes continued to answer in Kazakh. It is true that Kazakh and Kyrgyz are similar to each other, but they have many differences in grammar and word meanings. The inability of AI to distinguish between them and recognize Kyrgyz properly caused a strong public reaction. People who speak other low-resource languages also became part of this discussion. Kyrgyzstan is an independent country, so the fact that AI may have difficulty recognizing its language is indeed a problem. This raises another question: there are many languages in the world that belong to peoples without independent states, so does AI not recognize their languages at all?
The main reason for this is, of course, the small amount of digital information and data sources available for these languages. If there is not enough information about low-resource languages, if people do not upload enough data to digital sources, and if the speakers themselves do not use their language actively and fully, such problems will inevitably arise. There can also be confusion when the grammar of one’s own language is mixed with the grammar of another language or when its writing system is used incorrectly.
Currently, to improve the situation, work is actively being carried out in Central Asian countries, including Kyrgyzstan, to create local language models for Artificial Intelligence and collect open-source data. Efforts are being made to solve this problem.
Yes, of course, we can agree that AI has learned high-resource languages better. It can even help expand and preserve the use of these languages.
For example, in English, there are many:
-
scientific articles;
-
books;
-
news;
-
Wikipedia materials;
-
forums;
-
blogs;
-
educational materials.
This is because English has become an international language and its use continues to expand. There are also many grammar books and many films that help people learn and understand the language system. It is becoming useful not only for AI but also for people to learn English.
However, many low-resource languages have much less digital material and information of this kind.
I think AI can preserve languages. But under what conditions? To preserve low-resource languages, AI must be specially trained to understand these languages well from grammatical, lexical, and orthographic perspectives. I think it is necessary to create models that can store a large amount of information about these languages in their databases. If the people who speak that language also upload information to social media without mixing other languages and preserve the grammar of their language, AI can learn from these materials more effectively. Also, if you want Artificial Intelligence to answer in your own language, you should write your questions correctly, preserving the letters and orthography of your language. I think this method can also help.
Basically, everything is in the hands of humanity. Since humanity itself created such amazing technologies, people can also adapt them according to their needs. People are also responsible for languages that are disappearing. Today, many people do not speak their own languages on social media. As a result, the use and unique features of a population’s official language may decrease. The Internet and all digital sources are directly connected to each other, so they influence this process.
Conclusion
In conclusion, Artificial Intelligence and its digitalization system can directly affect both the preservation and weakening of languages. In order to preserve a language and help AI learn it fully, a large amount of information in that language should be regularly added to digital data sources, including books, educational materials, films, videos, interviews, and news in that language. If such conditions are not created, AI may not be able to learn or preserve languages whose use has become very limited.
References
-
Nature magazine Mohamed L. Seghier «ChatGPT: not all languages are equal»
-
Nick Petrić Howe “Why ChatGPT Can’t Handle Some Languages” — 2024, Nature.
-
UNESCO “Global Roadmap for Multilingualism in the Digital Era: Advancing the Role of Language Technologies” — 2025/2026
-
KyrgyzNLP: Challenges, Progress, and Future — 2024 Anton Alekseev, Timur Turatali (2024) “KyrgyzNLP: Challenges, Progress, and Future
-
Turatali, T., Turdubaeva, A., Izmailov, R., Alekseev, A. M., & Nikolenko, S. I. (2026). KyrgyzLLM-Bench: Benchmarking Kyrgyz Language Understanding
-
Deck, A. (2023, September 6). We tested ChatGPT in Bengali, Kurdish, and Tamil. It failed. Rest of World.
-
Madeleine North (2025) “Why generative AI needs to be trained on more languages”
-
Marleen Haboud (2019) “An Epidemic Is Threatening Indigenous Languages” — UNESCO.
-
UNESCO Ad Hoc Expert Group on Endangered Languages (2003) “Language Vitality and Endangerment”
-
UNESCO — “Towards World Atlas of Languages”