The Cost of Thinking in Spanish

Three years ago I was tasked with recruiting a team of AI engineers, and the candidate pool could come from either Mexico or India. I asked the same questions in every interview, and without a doubt, the most striking difference between candidates from the two countries was their command of English, a difference that, unfortunately, ruled out almost all of the Mexican candidates in that group.

In Mexico, people who speak English are estimated to earn between 19% and 28% more, depending on the study, and they also find jobs more easily and move up faster within their companies. Even so, the EF index ranks Mexico among the countries with the lowest English proficiency: 103rd out of 123 in 2025, the lowest in Latin America.

In the software era this gap grew, since most of the documentation for new technologies on the internet is written in English. That means people who speak the language can absorb new techniques and tools more easily. I tell my own students to learn the terms in English, because it is much easier to find resources online when you search that way.

In the AI era, the disadvantage of not speaking English will only grow, through three different channels.

The first is cost. Models process text in tokens, and their tokenizers were built from data that is mostly in English. Several studies have shown that the same task in Spanish takes between 30% and 60% more tokens, depending on the model. That means a programmer using AI in Spanish will run out of tokens faster than one using it in English.

The second is local knowledge. As I wrote in this space in May 2025, LLMs understand our local context poorly, and in general they lack much of the knowledge that lies outside the English-speaking world, which limits their usefulness in other countries. AI knows more about a small town in Delaware than about Querétaro or Barranquilla.

The third is reasoning. Reasoning models tend to think in English even when you ask them in Spanish, and when they are forced to reason in the user’s language, their accuracy drops. Today, the best reasoning happens in English.

The reality is that there is no immediate solution. There are no models trained primarily in Spanish, or in any language other than English or Chinese, that can compete with Claude or GPT. Even Chinese models such as DeepSeek or Kimi revolve around English and Chinese. And retraining the models may not be enough to remove these biases: the tokenizer is set before training, and studies show that forcing a model to reason in another language costs it accuracy.

Since the arrival of the internet, English became its lingua franca, with most of the content and platforms in that language. Today that advantage is turning into a measurable disadvantage for any company that uses AI in another language: a company whose employees speak English will spend less on AI than an identical company whose employees do not.

This could push the wage premium for speaking English even higher, with multiple effects on Mexico’s economy. According to Mexicanos Primero’s Sorry study (2015), the few public middle school graduates who reached an elementary level of English did so thanks to private lessons. That deepens an inequality rooted in the most basic levels of schooling.

And I have not seen the AI plans proposed in Mexico address this issue, or explain how they plan to solve it. That leaves society adrift, trying to close these gaps with its own tools. As long as no one takes this seriously, the gap will keep growing, and a very capable Mexican will keep losing out to a candidate from India for not speaking English.