Nobody Talks About the Language

The risk with sign language AI is not the one everybody names.

Two conversations happen every time AI and sign language end up in the same room.

The first one: interpreters are finished. The second one: soon we will fix deafness anyway.

I have heard both for ten years. Neither is the thing that keeps me up at night.

Interpreters are not first

Interpreting will change form. It will not disappear.

An interpreter is not a translation function. They hold the room. They read who is uncomfortable, who is lying, who did not understand and will not say so. In a medical appointment, a courtroom, a negotiation, a funeral, you want a human being who carries responsibility for the meaning. No model carries responsibility.

What AI can take is the routine. The repetitive loops, the short standard exchanges, the low-stakes back and forth that eats hours and gives nothing back. Remove that and interpreters spend their time where their skill actually matters. That is not a threat, that is a better job.

And if deafness gets fixed?

Maybe one day it will be, for those who want it. That is each person's decision and I will not argue with anyone's choice.

But notice what that conversation quietly assumes: that deafness is the problem and sign language is a workaround. It is not. For many of us being Deaf is not a defect to repair, it is where we come from. And even in a world where every ear could be repaired tomorrow, the thing worth keeping would still be there.

Sign language is not a substitute for speech. It is a full language with its own grammar in three dimensions, its own literature, its own humor, its own way of thinking. It should be permanent, and honestly it should be universal: hearing people gain something real by learning it, because it is another way to imagine, to reason, to say a thing.

Fix the ear, and the language still matters. That is exactly why the language is what I worry about.

The risk is not disappearance

Here is what nobody talks about.

Sign language will not vanish. It will get flattened.

Picture the pragmatic path. A model needs to serve many markets, so it needs one vocabulary. Regional variants are noise, so they get dropped. Rare signs are hard to learn, so they get replaced by common ones. Facial grammar is subtle, so it gets simplified. Multiply that by a few years of everyone training on the same convenient data, and you get one global sign language: efficient, learnable, and much poorer than any of the languages it replaced.

Make it concrete. ASL has several regional signs for BIRTHDAY: a convenient model keeps one and corrects the others. Raised eyebrows turn an ASL statement into a yes/no question: drop the face because it is hard to track, and the model cannot tell asking from telling. A signer describes how a car swerved with a classifier built on the spot, a construction that exists once and never again: a fixed vocabulary calls that noise. Each simplification is defensible alone. Together they are a different, smaller language.

The case for it is not stupid, which is what makes it dangerous. International Sign is genuinely easier to pick up than Esperanto ever was. A universal signed layer would let a hearing person communicate with almost anyone, including across spoken languages they do not share. That is a real argument and I understand its appeal.

But the way spoken language works online should tell us where that road ends. Broader reach, thinner language. Sign language would take the same path, faster, because it has no written form to anchor it.

The bias is already in the data

This is not a future problem. The bias is sitting in today's training sets.

Most sign language video available online is interpretation: someone, often hearing, translating spoken English live, under time pressure, in something close to English word order. It is the cheapest data to scrape, so that is what models get trained on. The result is a model that has learned signed English and believes it is ASL. To a hearing engineer the output looks fine. To a native signer it is off in every sentence.

Then the model becomes the judge. It powers a learning app, the app grades a native regional variant as a mistake, and the correction flows in the wrong direction: the machine trained on approximations is now correcting the people who own the language.

And remember who learns from those apps. Around 90% of Deaf children are born to hearing parents. Those families will learn to sign with whatever tool is on their phone. If the tool carries the bias, the child grows up being corrected by it, at home. One generation is enough for the bias to become the standard.

Language is thought

This is not about aesthetics or nostalgia.

In our corpus, roughly a third of the vocabulary appears exactly once: names, regional variants, one-off constructions, the things a specific person says in a specific place. That tail is not noise. That tail is the language.

And sign language is not just how we communicate, it is how we think. Space, simultaneity, visual logic. Impoverish the language and you do not simply lose vocabulary, you narrow how an entire community reasons and imagines. That is the part that cannot be recovered later.

AI will not be responsible for that. AI has no face and no conscience: it follows the direction it was given. The people building it will be responsible, and by the time the drift is visible in the data, it will be too late to reverse.

Which is why we do it country by country

This is the reason the Deaf Digital Heritage is built one language and one country at a time, instead of one convenient global dataset.

It is a harder, slower, more expensive way to build. It is also the only version that preserves what it records. Each language keeps its own corpus. Regional variants are labeled as variants, not corrected into errors. The rare signs stay in. Deaf annotators make the calls, because that is a technical requirement, not a values statement.

I want the future everyone imagines. AI that actually understands sign language. Deaf people using the same assistants as everyone else, in their own language. The communication barrier broken instead of managed.

We just have to arrive there with the language intact. We are the ones deciding that right now, in the data.

Follow @CLERC to track the build.