How Many of the World's Living Languages Have No AI Support at All

The answer depends entirely on two things people usually leave unstated: which count of living languages you start from, and what you are willing to call support. Fix both and the arithmetic is short, and the result is that the overwhelming majority of living languages have no documented support from any major provider.

The denominator

The standard reference is Ethnologue, published by SIL International. Its 27th edition, published in 2024, counts 7,164 living languages worldwide. That is the figure most people mean when they say “about seven thousand languages”, and it is the one used throughout this page. Ethnologue publishes the current count and its methodology, and the count changes with each annual edition as languages are split, merged, or reclassified as extinct.

It is worth being clear that this number is a judgement, not a measurement. The boundary between a language and a dialect is not decidable in general — the pages on Scots and on Cantonese are both, at bottom, about varieties whose status is contested. Every catalogue draws that line somewhere, and Ethnologue draws it using the ISO 639-3 code set, which SIL also maintains. Using a different catalogue would move the denominator by hundreds either way, but not by an order of magnitude, which is all the following arithmetic needs.

What counts as support

This is where most published answers quietly cheat. There are at least four distinct things “supported” can mean, and they produce wildly different numbers:

• Documented and evaluated. The provider names the language in a model card and reports evaluation results for it. This is the strictest reading and the one that matters if you are shipping.
• Documented, not evaluated. The language appears on a supported-languages list with no published quality figure attached.
• Present in training data. Almost every web-scraped corpus contains some text in hundreds of languages, because the web does. Presence is not capability; a few thousand sentences of a language buys the model almost nothing.
• Will produce output that looks like the language. The weakest reading, and the one that produces the most flattering numbers, because a model will happily produce confident, fluent, wrong text in a language it barely knows.

The numbers below use the second reading — the provider put the language on a list in public. It is the only one that can be counted from primary sources without inventing anything, and it is generous: every figure here is an upper bound on documented, evaluated support.

What providers actually document

Four figures, each with its publisher and date named, because a language count with no source attached is worthless:

• Meta, Llama 3.1 (2024) — 8 languages. The model card states support for English, German, French, Italian, Portuguese, Hindi, Spanish and Thai, and is explicit that other languages are outside the supported set. Meta publishes the model card on GitHub.
• Meta, NLLB-200 (2022) — 200 languages. A dedicated translation model, not a general chat model, and a deliberate research push at exactly this gap. Meta describes the No Language Left Behind project.
• Google Translate (June 2024) — 243 languages. Google announced the addition of 110 languages in a single release, calling it the service’s largest expansion. Google published the announcement on its blog. This is a translation product rather than a generative chat model, and the two capabilities are not interchangeable.
• Meta, Massively Multilingual Speech (2023) — over 1,100 languages for speech recognition, with language identification claimed for several thousand more. Speech, not text generation — included because it is the largest documented count anyone has published and it still does not change the conclusion.

The subtraction

With the denominator at 7,164 and the numerator taken from each provider figure above, the arithmetic is one line each. The assumption being made is that a language documented by one provider is documented at all — so these are per-system gaps, not a global gap:

Denominator: 7,164 living languages (Ethnologue, 27th ed., 2024)

Llama 3.1 text generation
  7,164 - 8     = 7,156 unsupported   ( 99.89% )

NLLB-200 translation
  7,164 - 200   = 6,964 unsupported   ( 97.21% )

Google Translate (June 2024)
  7,164 - 243   = 6,921 unsupported   ( 96.61% )

MMS speech recognition (upper bound, ~1,100)
  7,164 - 1,100 = 6,064 unsupported   ( 84.65% )

The best documented coverage of any system in that list, and it is a speech model rather than a text generator, leaves roughly five in six living languages out. For general-purpose text generation — the thing people mean when they say “AI” — the documented figure is closer to 99%, because general chat models document single or double digits of languages, not hundreds.

Note what the arithmetic does not say. It does not say 99% of people are unserved: the documented languages are the ones with the largest speaker populations, so coverage by speakers is far better than coverage by language. Both numbers are true and they answer different questions. If your question is “can I ship a product in the language my users speak”, the speaker-weighted number is the relevant one and it is much friendlier. If your question is “how much of human linguistic diversity is represented”, it is this one.

Why the real gap is wider still

Three effects push the honest number above even these figures.

First, the lists include languages where the model is documented but poor. Being on a supported-languages page carries no quality floor, and most of these languages have no benchmark at all, so there is often no public evidence either way.

Second, the tokenizer imposes a cost penalty that scales inversely with corpus share. A language that is nominally supported but is segmented at three or four tokens per character is expensive enough to be commercially unsupported for many applications, which is the argument in the tokenizer vocabulary bottleneck.

Third, many of the 7,164 have little or no digital text corpus to train on at all — not a small one, none. For those languages the gap is not a prioritisation decision that a provider could reverse next quarter; it is a data-collection problem measured in years.

What to check before quoting a number

If you are citing a figure like this in a document that matters, three checks take about ten minutes and stop you repeating a stale claim. Confirm the current Ethnologue edition and its count, because the denominator moves. Pull the supported-language list from the specific model card you are relying on rather than from the provider’s marketing page, because those routinely differ. And state which of the four definitions of “support” above you used, because a reader who assumes a different one will read your number as off by more than an order of magnitude.