Language as a Barrier and Facilitator for Artificial Intelligence

With Artificial Intelligence (AI) permeating every aspect of our livelihood from chatbots to voice assistants, language is central to the development, use, and in the social impact of AI. With the English language being the default mode of communication with most of these systems considered, the question arises regarding its inclusivity and accessibility for non - English speaking groups, which is the majority of the population in the world. Language here isn't merely a modem of communication, it determines who can access said technology and who gets to be included not only in its use but also in its development. Technology is meant to be inclusive and accessible, and communities cannot benefit from technologies that don't speak or understand them, thereby making language an incredibly important feature.  

Language can either be a barrier that prevents people and the system from understanding one another or be used as a facilitator making AI inclusive, accessible, useful and culturally sensitive. Although the world has over 7000 languages, AI tools are available only for a small fraction of them. UNESCO’s report shows that most AI systems with their large language models are only trained on a few languages, thereby limiting accessibility of these models to indigenous and underrepresented languages. Currently nearly 90% of the training data available for generative AI systems stems from English thereby excluding a large portion of the world from accessing it. This dominance of the English language in these data models stems from colonial and neocolonial legacies of the English speaking countries rather than any sort of inherent linguistic superiority. AI therefore in its current form, tends to create a black box effect of replicating the same biased stereotypes and information that users feed into them. 

At the same time, AI offers important possibilities ranging from translation systems, multilingual speech interfaces, language-learning applications, and automated transcription that can  not only help individuals communicate across linguistic boundaries but also prove to be an incredibly useful tool for people with speech or hearing disabilities. Alongside that, it can aid language preservation which has been a longstanding concern. With AI being used to superspeed communication, responsible ethical and mutually beneficial use of the system could provide aid and do more good than harm. 

The barriers to AI in terms of language include limited training data, linguistic diversity, unequal access to AI technologies, translation difficulties and bias. This paper will explore two of the facets that are more pertinent in that of data unavailability and speech nuance. This paper will also assess how language can facilitate AI to be useful for society in terms of its use in natural language interaction and the fields of education, translation for accessibility and the cultural preservation of language. This paper will assess the barriers and avenues that language can provide for AI and where AI can be challenged to be made better for an inclusive use of it. 

2. Language as a Barrier for AI 

2.1 Digital Divide and Limited availability of Training Data in LLMs

Most large language models (LLMS) and training datasets are built primarily in English, creating a digital divide for non-English speakers. Languages vary widely in resource, quantity, quality, and diversity of textual data to train language models and there is a large unavailability of data on indigenous languages. Most major LLMs thereby underperform on non-English and low resource languages, specifically on languages with non-Latin scripts and are therefore not accessible in parts of the Global South countries. Studies show that amongst AI platforms that included African, East Asian, and European languages performed 50% worse in those languages than they did in English. Even with medium to high resource languages like Amharic, Bengali, Zulu and Chinese, accuracy levels dropped by an average of nearly 60%. 

There lies a quantity gap as well as a quality gap when it comes to labeled and unlabeled data on low resource languages with lacking usable text and showing less than 50% acceptable texts (Kreutzer et al., 2022). For instance, most data that is scraped for non- English languages are often machine translated or scanned from an image which tends to cause huge margins of errors(Dodge et al., 2021). Furthermore, researchers note a “token tax” where some languages  require more tokens to represent the same content compared to English. This gap drives up cost and latency.AI systems in low-resource languages are found to be up to five times more expensive to source and process compared to the English-language equivalents. This further manifests as lower performance in low resource languages and creates  systemic barriers in accessing the foundational infrastructure needed for AI development such as computing resources, the limited availability of high-skilled technical talent and as well as biased and uneven model performance across different linguistic and cultural contexts. Access to AI capacity is already limited by high cost and a lack of infrastructure in low-income countries but even for those with technical access, language can prove to be a major barrier. 

2.2 Dialect/accents, Ambiguity and Cultural Contexts 

Large language models are trained on extensive collections of human-produced text. However, fluency does not necessarily mean genuine human-like understanding. AI systems process relationships among symbols and patterns. Whereas language carries the essence of human experience, nuanced with cultural context and emotions which often gets lost in these models. This nuance can also get lost when it comes to translation causing difficulties. Neural machine translation requires large quantities of language data, which are abundant for languages such as English, Chinese, French and Spanish but insufficient for many African, Asian, Creole, Indigenous, and minority languages. This inherent imbalance causes to reinforce existing inequalities in society. Then there’s the larger issues surrounding dialects, accents, pidgin, nuance, sarcasm, cultural tones etc. that need to be accounted for even in a singular English language, let alone the other languages that need to be inputted in the system. AI systems often struggle with nuances that give language its depth and richness which can be hard to navigate around. Sarcasm, idioms, and metaphorical elements tend to get lost in communication leading to misunderstandings and misinterpretations. 

3. Language as a Facilitator for AI 

3.1 Education and Inclusivity 

Although language creates aforementioned barriers it is also one of the most important mechanisms through which AI has become accessible to people. From translating documents to summarising ideas and analysing text and speech patterns, it has become a model through which people have been gaining knowledge and  information. Digital assistants, automatically generated video captions and image descriptions, sign language avatars can enable people with disabilities to access information and education. Real-Time Translation, Speech-to-text transcription and real-time captioning which convert spoken audio into readable text can empower users with hearing impairments or reading difficulties allowing them to bridge the communication gap. This development can broaden access to digital tools. However, it is inclusive only when systems understand a wide range of languages, accents, dialects, literacy levels, and communication styles and are participatory in nature. 

3.2 Economic Empowerment, Language preservation and revitalization

Language serves as raw data to be input to provide an inclusive space for people to access and utilise. While doing so, it can create avenues of job opportunity for people from indigenous communities from lower income backgrounds with underrepresented languages and digital barriers. Multilinguistic AI models can therefore prove to be an inclusive strategy to not only build said AI models but also facilitate social growth and provide dignified digital work and income opportunities for economically disadvantaged communities. This creates a two fold effect in not only uplifting people out but also preserving language and thereby culture.

AI can support endangered and minority languages by helping communities document vocabulary, grammar, oral histories, pronunciation, and traditional knowledge. In New Zealand, the Maori language has experienced a revival in recent years with AI playing a supportive role. By producing new educational resources through text generating systems, and machine learning systems researchers have found a way to transcribe audio recordings of Maori oral traditions which has further aided in the preservation of its cultural heritage.

Speech recognition can transcribe recordings, while language models can assist with dictionaries, educational materials, translation tools, and language-learning applications.

This work is valuable because many languages are underrepresented in digital environments. If a language is absent from keyboards, search engines, translation platforms, and educational software, younger speakers may find it less useful in modern life. The UNESCO Global Roadmap for Multilingualism in the Digital Era recognizes this and has stated that there needs to be equitable representation of languages in the digital space. This includes an active community participation, data sovereignty, capacity building, multilingual standards, and stronger support for low-resource and endangered languages whilst also providing for communities who are inputting this data. However, it is to be noted that while language can be preserved digitally for knowledge consumption, language can only be truly preserved through actively using it and passing it along. 

4. Language-Inclusive & Responsible AI for the Future

While AI in its current form, unfortunately isn't as inclusive, but with strategic investment in low-resource languages, high-quality language data through participatory research with local communities, and equitable data ownership frameworks, there can be venues for language inclusive and responsible AI for the future. This also allows communities to engage and participate in the process of developing a tool that can be useful for them. Language is not simply an input that AI processes, it is also a record of human history, identity, social relationships, and collective knowledge. When AI systems support a language, they can strengthen its digital presence. When ignored or misrepresented, they may contribute to further marginalization. Involving language communities in the development and governance of AI resources helps create AI systems that are accessible, accurate, culturally respectful, and socially inclusive, while protecting the rights and contributions of the people whose languages and data make these very systems possible. 

Author: Swadhi Karthikeyan

Date: 2026