The 70% MSA Trap

How We Cage the Human Tongue to Please the Machine

I walked into the voice recording session as a supervisor, not a speaker. With me was a team of Syrians preparing to record spontaneous conversations in the Levantine dialect for a foreign tech company training an AI model to understand spoken Arabic. The brief looked clear enough in the email: seventy percent Modern Standard Arabic (MSA), thirty percent local dialect — all within the context of natural, everyday conversation.

I found myself at a loss for words standing in front of the speakers, realizing that what the foreign client was asking for couldn’t be asked of a human being. It was a request fit for another machine. What person can move their tongue while running a mental counter tallying the percentage of formal versus colloquial words? How can anyone be spontaneous while simultaneously running a mathematical audit on every single word they say?

This scene isn’t just a field incident from the world of language-data freelancing, whether in Arabic or even American English. It’s a complete metaphor for something much broader and deeper: we are re-engineering the human being to become raw material ready for machine consumption, instead of developing the machine to absorb humans as they actually are — messy, contradictory, and beautiful.

When a person is forced to measure their own spontaneity in percentages, speech is no longer expression — it becomes a theatrical performance staged for an audience of algorithms.

This article isn’t an attack on artificial intelligence, nor a call to fall behind the technology curve. It’s a fundamental philosophical question that Arabic-speaking freelancers raise from inside recording booths and behind quality-check screens: Who serves whom? Are we training the machine to serve human language, or are we training ourselves to serve the machine’s language?

Part One: The Psychology of the Tongue — Why Spontaneity Dies the Moment You Count It

The human brain doesn’t operate during speech the way a calculator does. When a Syrian woman talks with her neighbor, her brain doesn’t sort words into labeled slots (formal/colloquial). It thinks about meaning, emotions, social context, and the relationship between the two speakers. A formal Arabic word emerges when the context calls for it — when the moment is official, religious, or emotionally charged — not when the speaker decides they’ve used up their colloquial quota for this part of the conversation.

What linguists call “code-switching” — the real-time shift between dialect and standard Arabic, or even between two languages — is a phenomenon driven by subtle emotional, contextual, and social triggers, most of which are unconscious. It happens automatically: we slip into dialect when describing something funny, shift to formal Arabic when quoting scripture or a proverb, and blend a foreign technical term into a local sentence structure when explaining instructions. Nobody plans any of that.

(See our article series: The Multilingual Family)

Research in cognitive linguistics has shown that imposing artificial percentage-based constraints on native speakers during voice data production generates what’s known as a “Cognitive Overload”: the speaker’s attention shifts from the level of meaning to the level of calculation. Speech slows down, natural intonation disappears, sentences stiffen, and the conversation starts sounding robotic — even when it’s coming from flesh-and-blood human beings.

In other words: the moment you ask a human recorder to be spontaneous at a fixed percentage, you’ve killed spontaneity in the same breath you requested it. What reaches the model’s training database isn’t a real human voice — it’s a human performing the role of a real human. That subtle difference will shape everything the machine learns about our language.

Part Two: The Native Speaker Problem — When a Human Fails a Test of Their Own Humanity

There’s a painful irony that many freelancers specializing in language data projects know well: the native speaker is often the first candidate to get rejected from projects officially designed to capture authentic, real-world speech.

Why does this happen? Because tech companies aren’t really looking for Syrian, Egyptian, or Moroccan dialect as it actually exists. They’re looking for what might be called a “sanitized dialect” — a Platonic ideal of a dialect: an imaginary, polished version stripped of all natural human noise.

The Levantine dialect is a living, moving thing. It differs between Damascus, Aleppo, and Homs — even between the Shaghour neighborhood and the Mezzeh neighborhood within Damascus itself. In real conversation, it’s full of: natural pauses between words, unconscious sound blending, incomplete grammatical structures, and speakers correcting themselves mid-sentence. All of that is the humanity of language at its finest — and all of it gets classified by data-cleaning algorithms as “noise” to be deleted.

We’re not looking for a real Syrian voice. We’re looking for a theatrical performance of what the algorithm imagines a Syrian voice ought to sound like.

The practical result that field supervisors on recording projects witness: the most authentic and natural-sounding speaker gets rejected because their speech is “full of noise,” while the most polished and artificially careful speaker gets accepted because their voice is “clean” by the algorithm’s standards. That’s how authenticity gets redefined: not what a person produces naturally, but what the machine accepts by its own criteria.

And here we reach the more painful point: what does the machine actually learn from all this? It learns a falsified model of language that erases real dialectal diversity, marginalizes regional accents, and entrenches a “standard dialect” that nobody has ever actually spoken in a market, a café, or a home. Then that machine gets presented to us as a tool that “understands Arabic.”

Part Three: Theater and the Machine — The Illusion of Simulation and the Authority of Reference

This isn’t the first time humans have trained one another through imitation. Theater and cinema have been doing it for centuries. Both arts are, at their core, a form of “cultural training” through which societies pass on behavioral patterns, language, and values to their next generations.

But there’s a fundamental difference between art and the algorithm. When we watch a Damascus TV series or an Aleppo stage production, we instinctively know that what we’re seeing is a performance — that the language has been reworked to serve the drama, and that what the actor says in an emotional scene isn’t the daily standard for conversation. There’s what critics call a “Suspension of Disbelief”: we participate in the illusion with full awareness that it is an illusion, and when the curtain falls we return to our real language without absorbing the actor’s performance as a linguistic benchmark.

Artificial intelligence, on the other hand, is presented to us as a reference — not a performance. And the data it trained on — engineered under duress, representing a hybrid language no people ever spoke naturally — becomes the raw material for voice assistants, educational tools, and language-correction systems. Just as younger generations might learn “English” from movies without noticing the gap between screen language and street language, future generations might learn “Arabic” from a machine that trained on humans performing their fabricated spontaneity at seventy percent MSA.

Art admits it’s a performance, and invites you to enjoy the show while keeping your real world intact. The algorithm presents the performance as the truth, and invites you to measure yourself against it.

Part Four: The Feedback Loop — When the Machine Becomes the Teacher

Let’s think through where this path leads if it continues.

Today: we clip our tongues to produce data that satisfies the algorithm — speaking in an unnatural way to train a model that’s supposedly meant to serve our natural way of speaking.

Soon: the algorithm trained on our “engineered” data starts producing written content, answering children’s questions, correcting essays, and suggesting phrasing. Millions of users treat it as a trusted authority.

Down the road: younger generations who’ve spent hours with AI assistants and learning apps will develop a linguistic sensibility shaped by this “sanitized hybrid language.” When they speak, they may unconsciously imitate the machine’s patterns — not their grandmother’s, not their neighbor’s.

This isn’t a conspiracy theory. It’s a known law of linguistic systems called the “Feedback Loop”: when artificial outputs become the inputs of the next generation, the drift from the original accumulates year by year. We’re not just training the machine on our language. Bit by bit, we’re training ourselves on the machine’s language.

We’re not moving toward the flexibility of life. We’re moving toward the rigidity of code. And we don’t yet know whether that’s evolution or collapse.

More dangerous still is what might be called “Authority Bias toward the Model”: a growing tendency among users — especially younger generations — to trust AI outputs over their own linguistic instincts, or even over what they hear from their own families. When a machine tells you “this sentence is incorrect” with deep confidence backed by a sleek interface, you’re inclined to believe it — even if that sentence is exactly what eloquent speakers have said for a hundred years.

Conclusion: The Revolution of Human Noise

We’re not here to call for stopping AI development, or to reject language data projects that provide real opportunities for Arab freelancers. We’re calling for something more modest — and far more difficult: for tech companies to redefine what “quality” means in language data.

Quality isn’t the absence of noise. It’s the ability to absorb noise as information. The pause between two words isn’t an error to be deleted — it’s silent speech carrying meaning. Unconscious sound-blending isn’t distortion; it’s the fingerprint of a living dialect. The variation between neighborhoods in the same city isn’t chaos to be standardized — it’s richness that reflects the real social fabric of a community.

When the machine does that — when it absorbs the human being in all their scattered, spontaneous, contradictory fullness — that will be genuine intelligence. What we’re doing right now is the exact opposite: forcing the human to absorb the machine and its rigid templates.

In the Syrian context specifically, and across every Arab context where the local freelancer works as a bridge between living language and the digital model, there’s an added responsibility: to recognize that when our voices enter the recording room, they don’t just train a machine — they shape, unintentionally, the image of our language in the world’s digital memory. And that image deserves to be real.

Humans have never built their language in percentages, and they never will. The human tongue doesn’t run on a (70+30) equation. It runs on life — with all its chaos, contradiction, and beauty that no algorithm can calculate.

Zy Yazan Platform © 2026