The Misconception of Universal Language Proficiency in AI
When we use artificial intelligence (AI) for translating texts, answering questions, or drafting emails, we often assume it functions equally well across all languages. This seems logical: if AI is “intelligent,” it should handle all languages with equal ease. However, the reality is quite different.
AI models do not perform equally in English as they do in Spanish, or in Spanish compared to Euskera. Why is this the case? Is it an unavoidable technological limitation or a reflection of deeper digital world inequalities?
The Foundation: Data
To understand this, we must look at the foundation of these technologies: data. Language models like ChatGPT are trained on massive amounts of text, both original and created by those who fine-tuned them. However, the first significant asymmetry arises here: most written content on the web is in English.
Training Languages
OpenAI, the company behind ChatGPT, and other firms do not disclose exact percentages of each language’s weight in training, nor can the models calculate them with available data. Nevertheless, the trend is clear: English dominates this context, followed by other major global languages like Spanish, French, or German. With considerable distance, we find languages with limited digital presence like Catalan or Welsh. Even further away are minority languages with scant or no textual traces online.
With this distribution, the outcome is predictable: models perform better in languages with more data. It’s not about affinity but the opportunity for learning. When a model sees millions of examples in English, it learns its grammar, vocabulary, various registers, and cultural background better. Conversely, with few examples in a language, there’s less material to deduce reliable patterns.
Can This Gap Be Reduced?
Fortunately, modern AI doesn’t passively reproduce this inequality. Numerous strategies aim to mitigate the lack of data in less-common languages.
- Balanced Corpus: Adjusting the number of texts used for responses during training. Even if English is thousands of times more abundant, the model’s exposure to less-common languages can be increased, reducing English exposure.
- Multilingual Transfer Learning: Models don’t learn each language separately; they share internal representations. If the model learns Spanish, part of that knowledge benefits Portuguese or Italian. Similarly, German reinforces Dutch. This transfer aids less-data languages within the same language family.
Synthetic data generation through automatic translation or using multilingual parallel corpora like international organization documents or Wikipedia versions also helps. Later, human instructors native to the language correct unsuitable expressions, reinforce appropriate tones, and fine-tune cultural nuances that massive data misses.
Finally, techniques exist to prevent “catastrophic forgetting,” where the model continues training on dominant language data and inadvertently degrades knowledge of less-common languages. Regularization and continuous learning methods help maintain some balance.
AI’s Impact on Linguistic Diversity
Despite these technical resources, no resource can fully compensate for the lack of data in a language with little content renewal. Thus, English remains predominant, and the gap persists.
This raises an important question: can AI contribute to the loss of linguistic diversity? It’s a real risk. If it performs better in English, some may prefer using it in that language. Homogeneous text generation can influence institutional, academic, or media writing styles and displace local registers. Moreover, if a language barely appears online, it may be excluded from technologies shaping our communication.
Revitalizing Minority Languages
However, there’s an opposite potential: AI can revitalize minority languages. It can generate educational materials, aid in documenting vocabulary, serve as a learning interlocutor, or support digitalization projects. With political and cultural will, technology can be a valuable ally.
The uneven performance of AI across languages isn’t merely a technical issue; it mirrors real-world inequalities. It’s not about whether AI speaks some languages better than others—it does. The question is how we can build a future where technology reduces, rather than reproduces, linguistic gaps.
