Large language models process text in units called tokens. English maps efficiently onto tokens. Arabic does not.
This is not incidental. Tokenisers are trained on available text. The internet is predominantly English. The result is that a word or phrase which costs one or two tokens in English can cost four to six times as many in Arabic. The same semantic content. A different price.
Most commercial AI systems charge per token. The pricing model treats Arabic as computationally expensive. It does not say this explicitly. The arithmetic says it.
Arabic has over four hundred million native speakers. It is an official language of twenty-two countries and one of the six official languages of the United Nations. It is not a minority language. The AI industry's pricing structure treats it as one.
The consequence runs through the incentive chain. A developer building a legal aid tool for Arabic-speaking users pays a multiple of what an equivalent English-language tool would cost. A company deciding which markets to build for faces unit economics that consistently favour English. An individual using an AI assistant in Arabic receives less for the same price. None of these decisions require intent. The structure produces them.
The people most exposed to this are not difficult to identify. Asylum seekers, refugees, and migrants navigating legal systems in countries where they do not speak the dominant language are precisely the population for whom AI assistance could reduce the gap between institutional complexity and individual comprehension. They are also, in many cases, speakers of languages that the token pricing model treats as expensive.
The standard response to this is that it is a technical problem, not a moral one. Tokenisers reflect training data. Training data reflects the history of digital text production. No one decided that Arabic should cost more. This is accurate. It does not address the consequences of the decision that was made, which is that the tokeniser would be optimised for English and the pricing model would charge by the unit.
Structural disadvantage does not require an architect.
There is a further question that the technical framing avoids. When a technology that is becoming foundational infrastructure charges a consistent premium for interaction in particular languages, the long-term pressure is not only economic. It is linguistic. The rational response for an individual, a business, or a developer facing this premium is to use English. Over time, across millions of interactions, that pressure accumulates. The question of which language is efficient to think in is not only a technical question.
Arabic costs more to use in AI systems because the systems were not built with Arabic in mind. The pricing model converts that design decision into an ongoing charge. The people paying it are not the people who made it.