WHY IT MATTERS
“She speaks well, your machine. But it is not me she is speaking to.”
Lalla Fadma, 78, a village above Tafraout, in the Anti-Atlas
This page tells four short stories. Each one shows a mechanism at work in today’s AI, and what it costs the people the machine does not hear. The scenes are drawn from ordinary situations. The people are invented; the mechanisms are not.
1 · A machine that reads only one language
Her granddaughter hands her a phone. Lalla Fadma asks, in Tachelhit mixed with Darija, whether her late husband’s land papers must be redone now that the commune has changed its registers. The phone apologizes in standard Arabic, then offers three links about the weather in Agadir.
The machine did not misunderstand her. It did not hear her at all.
AI learns the world by reading, and most of what it reads comes from a web written mostly in English. Languages that are spoken more than written barely exist in that diet. Everything else reaches the machine through translation: researchers who look inside large language models find that they often reach the meaning in an English-like form before returning to the language of the reply.
| Language | Share of web pages machines read |
|---|---|
| English | about 42% |
| French | just under 5% |
| Arabic, all written forms | under 0.7% |
| Hindi | about 0.2% |
| Swahili | about 0.01% |
| Amazigh languages | not detected |
2 · The translation trap
In a Moroccan tale, the young hero runs away to his khal. Every Moroccan listener knows at once: he is seeking refuge outside his father’s authority, with his mother’s brother. The English translation says he went to his uncle. The story is still there. Its meaning is gone.
Translation is one of the finest human gestures. Imposed on every word, it becomes a steamroller.
| What is said | What translation keeps | What is lost |
|---|---|---|
| ʿamm and khal (Darija) | “uncle” | The father’s brother belongs to the lineage; the mother’s brother is the refuge. Hindi does the same with chacha, tau and mama. |
| hshuma (Darija) | “shame” | Not an emotion but a shared norm of restraint that protects common dignity. |
| inchallah (Darija) | “God willing” | Whether the speaker promised, hoped, or politely refused. |
| Evidential suffixes (Quechua) | nothing | How the speaker knows: seen, heard said, or supposed. |
| Four languages in one sentence | one language | The choice of language itself, which was a message. |
| A voice | a line of text | The way she slowed down on her husband’s name. |
3 · A double divide
In Casablanca, a manager who speaks French and standard Arabic already uses AI for almost everything. In the Souss, a farmer who speaks Tachelhit and cannot read gets almost nothing from it. Same technology, same country.
The first divide runs between the few countries that build AI and the many that use it. The second runs inside every country, between the people machines hear and the others. As states move hospital appointments, birth registrations and social aid to voice, a machine that misunderstands rural, elderly or non-literate citizens will not refuse anyone. It will simply misunderstand some more often than others, until they stop asking.
This is not only Morocco
- India: hundreds of millions of people whose languages make up a fraction of a percent of the web.
- West Africa and the Sahel: national languages that are mostly oral, beside official languages inherited from colonization.
- The Andes: Quechua, spoken by millions, with a grammar English cannot carry.
- Quebec and First Nations: a people that defended its language against a giant neighbour, and Indigenous languages fighting to be passed on.
4 · The time left
In 1974, Ned Maddrell died on the Isle of Man, the last native speaker of Manx. In 2008, Marie Smith Jones died in Alaska, the last speaker of Eyak. She used to say that no one could know what it is like to be the last.
- One language a month. Without intervention, the rate of language loss could triple within forty years (Bromham et al., 2022).
- About 1,000 speakers of Haketia remain, against some 30,000 around 1900 (Jewish Language Project).
- Archives asleep. Thousands of hours of tales and songs sit on fading media, with no index. A recording no one can listen to is not a memory. It is a promise not kept.