By Stuart Kerr, Technology Correspondent
Published: 31 July 2025
Last Updated: 31 July 2025
Contact: liveaiwire@gmail.com | Twitter: @LiveAIWire
Author Bio: About Stuart Kerr
As large language models continue to evolve, one question remains underexplored: whose voices do they truly represent? While AI promises borderless communication, it often silences the nuance and richness of global dialects, minority languages, and culturally embedded expressions.
Speaking in Majority
AI’s fluency is largely defined by the data it consumes. This poses a major issue: the internet—and by extension, training datasets—leans heavily toward dominant languages like English, Mandarin, and Spanish. As the Brookings Institution notes, this leads to generative models that excel in majority tongues but fail to grasp the syntax, semantics, or even the cultural relevance of under-resourced languages.
A simple prompt in a minority language might yield garbled or generic output, effectively denying entire communities meaningful access to AI’s potential. According to Language Magazine, this isn’t just about communication—it’s about identity and dignity.
Dialects, Dead Ends, and Data Deserts
Dialects, slang, and regional phrasing often confuse models trained on “standardised” forms. AI tends to flatten language, removing regional spice in favour of homogenised accuracy. For communities whose linguistic forms are rooted in oral tradition or colonial resistance, this technical limitation has social implications.
The Brookings Institution further calls for inclusive language technology frameworks. From indigenous groups in South America to isolated island communities, the absence of AI-literate tools in local languages represents not just a gap—but a barrier.
When Algorithms Translate the World
The problem extends beyond bias—it affects global understanding. As described in AI Digitising Cultural Heritage, algorithms can digitise but rarely contextualise. Translation engines, for example, often default to literal meanings, stripping folklore, idioms, and spiritual references of their depth.
Similarly, predictive text and speech-to-text systems may consistently ‘correct’ non-standard grammar into dominant language norms, reinforcing colonial hierarchies in digital form.
Academic Voices Weigh In
A 2025 arXiv paper on generative AI and endangered language preservation calls for diversified data pools and collaboration with linguistic anthropologists. The paper highlights tools already in use that help transcribe oral histories and automate dialect tagging—but warns that they are underfunded and overlooked.
Meanwhile, the Arab World English Journal (AWEJ) explores how educators must counterbalance algorithmic bias in classroom settings, especially in multilingual societies. AI, they argue, should reflect—not rewrite—human linguistic experience.
Repairing the Code
AI doesn’t have to erase linguistic diversity. It can, if designed ethically, become a catalyst for preservation. Projects like “language revitalisation through code” are gaining ground, combining NLP models with grassroots documentation efforts. As covered in AI Refugee Forecasting, coordination across humanitarian sectors can support cultural integrity alongside technical advancement.
In the arts, as seen in AI in Theatre, code has already begun capturing and expressing regional storytelling forms. There is precedent for inclusivity—if we’re willing to code for it.
Conclusion: Words Matter
Every language carries a worldview. When AI misrepresents or ignores a language, it erodes more than syntax—it chips away at identity. It’s time to move beyond “universal” models and start building ones that are polyphonic, context-aware, and locally rooted.
AI shouldn’t just learn to speak. It should learn to listen.
About the Author
Stuart Kerr is the Technology Correspondent for LiveAIWire. He writes about artificial intelligence, ethics, and how technology is reshaping everyday life. Read more