Language diversity in LLMs: the position of Dutch

By Ivo Donker - 6 August 2026

The rise of large language models (LLMs) has fundamentally changed the way text is generated, analyzed and processed. Although modern AI systems appear at first sight to command several languages impressively, a considerable linguistic inequality lies behind the user interface. Language capability in LLMs is far from evenly distributed across the thousands of spoken languages in the world. The underlying architecture of these models is basically language-independent, but the outcomes are determined directly by the size, diversity and structure of the training data on the global web.

Within this field of forces, Dutch occupies a peculiar and sometimes uncomfortable middle position. Dutch is neither a dominant world language such as English or Spanish, nor a sparsely documented language with a marginal digital archive. As a mid-sized language, Dutch has enough digital material to let language models function reasonably, but its capability lags structurally behind the primarily English-language core in which many of these AI systems are rooted. In this analysis we explore the mechanics behind this imbalance, the hidden costs of tokenization, the specific substantive bottlenecks in practice, and the strategic choices for Dutch organizations.

The imbalance in training data on the global web

The fundamental cause of the linguistic imbalance in large language models lies in the sources AI developers draw on. Large language models are trained on gigantic volumes of text crawled from public web pages, digital libraries, news archives and discussion forums. The public internet, however, is no proportional mirror of the human world population or its linguistic diversity.

English covers by far the largest share of available digital documentation. It is not only the base language of internet infrastructure but also the international standard for scientific publications, software documentation and global business information. When developers gather web pages to assemble training corpora, the outcome reflects this historical imbalance in the digital ecosystem.

Moreover, it is not only about the physical volume of words but equally about the substantive variation and depth of the text sources. A language amply represented with a rich palette of specialist domains — legal dossiers, medical literature, technical manuals, philosophical treatises and literary works — lets a language model build deep conceptual relationships. In languages with less substantive diversity on the web, the model mainly learns everyday colloquial language or repetitive commercial text, and struggles with complex reasoning in that specific language.

The status of Dutch: a mid-sized language

In the global landscape of AI training data, Dutch falls into the category of well-documented mid-sized languages. With tens of millions of native speakers in the Netherlands and Flanders, combined with high digital penetration and an active online culture, a solid base of digital material exists. Government documents, newspaper archives, encyclopedic articles and public debate provide a continuous stream of Dutch-language text on the network.

As a result, most commercial and open-source language models can communicate convincingly in Dutch. A user asking an everyday question in Dutch usually gets a grammatically correct and comprehensible answer. This apparent flawlessness conceals, however, that Dutch often rides along in the pre-training phase as a byproduct of the English-oriented data stream.

The challenge for Dutch lies not in basic grammar but in the substantive depth of its representation. Compared with English, the supply of high-quality, specialized and current Dutch-language data is limited. This means that for abstract concepts, legal nuances or specific domain knowledge, a model often has to fall back on representations formed primarily in English and converted to Dutch during generation. Dutch therefore performs adequately for generic applications but falls short as soon as accuracy, cultural context or specific domain terminology is required.

Tokenization as a hidden disadvantage for Dutch

A crucial mechanism through which language inequality expresses itself is the step preceding processing by the neural network: tokenization. Language models process neither individual letters nor complete words, but break text into small fragments called tokens. A token can be a whole word, but also a syllable, a prefix or a loose sequence of characters.

The way a tokenizer cuts up text is determined by the frequency of character combinations in the tokenizer's own training corpus. Because most tokenizers are trained on predominantly English-language corpora, their vocabulary tables are optimized for English word structures. Common English words are stored whole as a single token.

For Dutch this has adverse consequences. Dutch words are often split by a largely English-oriented tokenizer into several separate fragments. A simple Dutch word can therefore consist of two, three or even more tokens, while its English translation takes up only one. This mechanism causes three concrete disadvantages:

• Higher financial costs: Since API providers' rates are almost without exception computed per token processed, processing Dutch-language text is on average more expensive than processing exactly the same content in English.
• Lower processing speed: A model's processing time scales with the number of tokens generated. Because Dutch sentences break into more tokens, generating an answer physically takes longer.
• Reduction of effective context length: While the newest model generations support gigantic context windows, less efficient tokenization limits how much factual information fits in that window. In the ongoing race for context , Dutch therefore loses effective processing capacity relative to English.

For a deeper technical explanation of how tokenizers assign fragments and which algorithmic choices underlie this, consult the background guide on tokenization explained on the learning platform.

Where practice chafes: typical shortcomings

In daily practice, the imbalance in training data and the dependence on English concepts leads to specific linguistic and substantive errors. These shortcomings range from subtle style errors to serious substantive distortions.

One of the most striking ailments is so-called translationese. The model formulates sentences that are not flatly ungrammatical, but whose sentence structure, word choice and idiomatic turns directly reflect English constructions. The text consequently feels unnatural, stiff and mechanical to a native speaker.

LLMs also show specific shortcomings in the following areas:

• Forms of address and register: Dutch has an explicit distinction between formal ('u') and informal ('je/jij'). Models regularly jump at random from the formal to the informal form within a single paragraph, because English 'you' carries no built-in grammatical distinction.
• Regional variation: Dutch has considerable differences between the Dutch used in the Netherlands and Belgian Dutch (Flemish). Models mix Flemish vocabulary and sentence constructions haphazardly with Netherlandic Dutch, which is undesirable for audience-targeted communication.
• Writing compounds as one word: In Dutch, compound nouns are written as one word (for example 'taalmodellen' or 'kwaliteitscontrole'). Under English influence, where words stand apart, LLMs frequently commit the 'English disease' by splitting compounds with spaces.
• Idiomatic expressions: Dutch proverbs and figurative expressions are regularly translated literally from English, or Dutch proverbs are given an English sentence structure.
• Specialized professional language: In the legal and medical domain, a misused term can have major consequences. Dutch statutory articles, case law and medical guidelines use specific terminology that differs from literal translations of Anglo-Saxon sources.

Overview of common error categories

To make these bottlenecks visible, the table below gives an overview of the most common error categories in Dutch-language LLM output and the concrete points to test during validation.

Routes to improving performance in Dutch

To bridge the gap for Dutch in language models, various strategies are applied in practice and in research. There is no single solution, but rather a combination of approaches intervening at different levels.

One obvious step is expanding and cleaning Dutch-language training data. Through targeted collection of high-quality sources — digitized government archives, scientific journals and quality media — the language's share in the pre-training phase can be increased. Where authentic data is scarce, attention also turns to generating synthetic data to supplement specific linguistic patterns and sentence structures.

Another route is fine-tuning existing base models on specific Dutch-language instruction sets and domain knowledge. This lets developers adjust a model's behavior and push back translationese without having to repeat expensive pre-training. Here small language models play an increasingly important role: more compact models trained specifically on quality Dutch-language data can outperform enormous generic models on well-defined tasks.

Technical innovation also focuses on adapting the tokenizer itself. By using a tokenizer explicitly trained on a multilingual corpus in which Dutch is better represented when building models, processing costs fall and processing speed rises. This trend is clearly visible within broader open source LLM trends, where the community increasingly tinkers with language-specific adaptations of international base architectures.

The challenge of Dutch-language evaluation and benchmarks

A persistent problem in improving Dutch-language LLMs is the scarcity of reliable evaluation methods. To determine whether a new model performs better than its predecessor, the AI industry relies on standardized benchmarks. The best-known and most-used evaluation sets — testing logical reasoning, reading comprehension and domain knowledge — are, however, almost without exception developed in English.

To solve this, researchers often reach for machine-translated versions of these English benchmarks. While this is a quick way to create a Dutch-language test set, this approach introduces fundamental problems:

• Translation errors in the test set: If the machine translation of a question or answer option contains a subtle error or ambiguity, the benchmark becomes unreliable.
• Cultural and institutional bias: Questions resting on the American legal system, the American education system or specific Anglo-Saxon cultural expressions stay irrelevant to Dutch reality even after translation.
• Measuring translation capability: A translated benchmark primarily measures how well the model can reconstruct translated English-like logic, rather than how well it can reason within the character of the Dutch language.

Developing authentically Dutch-language benchmarks — built from the ground up by native speakers with questions about Dutch legislation, history, culture and language use — is therefore an essential but labor-intensive necessity.

Practical advice for Dutch organizations

For Dutch companies, governments and institutions building or implementing AI applications, the current state of the art means acting soberly and strategically. Blind trust in generic claims from international model vendors is unwise when it comes to critical Dutch-language processes.

The following advice offers a handhold in developing and selecting AI solutions:

• Always test on your own representative material: Do not rely on general leaderboards. Assemble an internal test set with authentic documents from your own organization and validate the outcomes manually with native speakers and domain experts.
• Steer tightly through prompting: State explicitly in the instructions which form of address is wanted (formal or informal), instruct the model to write compounds correctly as one word, and forbid the use of known anglicisms.
• Account for higher token costs: Factor in that Dutch-language texts consume more tokens than comparable English projects, which affects operational costs and the processing capacity required.
• Safeguard accuracy in critical domains: In applications in healthcare, the judiciary or the financial sector, it is wise to work with strict grounding through retrieval-augmented generation (RAG) using verified Dutch-language sources.

Looking ahead at the language landscape

The position of Dutch in the landscape of large language models is expected to improve step by step, but it will not draw level with English by itself. As long as global AI development is driven primarily by English-language investment and data infrastructures, English's technological lead will persist.

At the same time, the growth of multilingual open-source models, better tokenizers and initiatives for local data collection are making functionality in Dutch steadily more mature. By investing deliberately in quality datasets, sharp evaluation methods and practical adaptations, Dutch organizations can keep building reliable and culturally aware AI applications.