The AI That Reads Mexican Spanish, Cuban Spanish, and Argentinian Spanish — And Knows the Difference

Like many of these things, it began with a minor annoyance. When a linguist in Madrid asked a well-known chatbot to write something in “Spanish,” it nearly always responded in Spanish, with clipped consonants, the vosotros form, and vocabulary that would sound a little stiff in Mexico City and a little strange in Buenos Aires. Not exactly incorrect. Simply narrow. As though 600 million people had decided to speak in unison somewhere offstage. She didn’t.

Even though the dataset that resulted from that annoyance is tiny by AI standards—just thirty questions—it has accomplished something that more sophisticated systems haven’t quite been able to. In simple terms, it asks the models if they can distinguish between four regional varieties of Spanish, including Mexican, Cuban, and Argentinian. Then it waits for them to fail, almost courteously.

A lot of them do. Some people achieve success in peculiar, partial ways, either by correctly answering lexical questions but struggling with grammar, or vice versa. Based on the preliminary findings, it appears that these models have learned a lot of Spanish without actually understanding that it is plural.

The questions themselves are surprisingly easy. The question is whether “Llegas tarde, vávete y corre” or “Llegas tarde, võete y córrele” sounds more natural. The second is unmistakable to a Mexican ear; that tiny le tucked onto the verb is like a fingerprint, the kind of detail you grow up with but never give much thought to. It may appear to be a typo to a model that was primarily trained on Peninsular text.

Whether you “levantarse” or “pararse” to stand up is another question. Pararse sounds like you’ve stopped in Madrid. It’s how you get out of a chair in Mexico City or Havana. It’s not trivia. They are the places where identity resides, the tiny joints of a language.

The test was developed slowly by the team. The questions were written by one expert, reviewed by two others, and revised until they were clear enough to be understood in every area they addressed. The questions were then passed through a number of LLMs, sometimes asking the model directly and other times telling it to respond as though it were born in Montevideo, San Salvador, or Lima. The role-playing is important. It shows if a model has a single voice and a costume rack, or if it can switch dialects at will.

It’s difficult to ignore how late this work is coming in. Spanish, despite its enormous size, has frequently been treated as a single block, while English NLP has long debated dialects. The default has been to translate an English benchmark into “generic” Spanish, whatever that may be, which makes many speakers barely noticeable. Large labs and investors discuss multilingual AI; the more difficult and truthful question is whether their multilingual AI is also multidialectal.

Cautious optimism is warranted. The methodology is transparent, the dataset is available, and other researchers are already experimenting with it. It’s still unclear if the upcoming generation of models will take this to heart or continue to reduce Spanish to a single, polite standard. As you watch this happen, it seems like technology is finally being asked to listen instead of just talk. That’s a minor change. However, minor changes add up in a language this diverse.