Lots of places have a language like this: the older generation uses it with each other and switches to something else with the young. Nobody announced the switch and nobody remembers when it started, and by the time someone notices and starts worrying, thirty years have gone.
Language loss usually gets told as a number (how many are left) and a moment (the last speaker). Neither is where it actually happens.
That 1992 paper estimated that only about 10% of languages were safe long-term and that up to half might already have no children learning them. The author never said "one every two weeks" — later writers took "90% in a hundred years" and turned it into a rate, compressing an estimate about a distribution into a beat that sounds precise.
The other route is to count. The Catalogue of Endangered Languages tallied the languages documented as having actually fallen out of use in recent decades and got roughly one every three months. The same catalogue lists 3,116 still-spoken endangered languages, about 45% of those remaining, and a separately compiled database using its own scale arrives at about 44% — rare agreement for this field. The denominator is itself a judgement, though: whether a way of speaking counts as its own language or as one branch under another has no clean criterion, which was the conclusion of the topic on whose speech counts as standard.
More informative is a 2022 study that modelled 6,511 languages against 51 predictor variables and projected that about 1,500 could cease to be spoken by the end of the century. Two results cut against intuition. Contact with other languages is not itself a driver — multilingual coexistence can hold steady for a very long time. And places with more years of schooling show higher endangerment, not because education is harmful but because school usually runs in one dominant language, taking a key occasion away in the children's generation. Evidence strength: large-sample correlational analysis, not a causal experiment.
In 1991 Fishman proposed a graded scale, and one level on it matters more than the rest: everyday transmission in the home, the neighbourhood, the community. His judgement was blunt — if that level fails, everything else amounts at most to biding time. So speaker counts mislead. Fifty thousand adults who speak it with no child learning it is closer to the end than five hundred speakers whose children are still learning; the first curve stays flat for forty years and then drops all at once. The standard definition of "endangered" is set on exactly this: the child-bearing generation no longer passes it on.
And it does not happen on one particular day. A language retreats occasion by occasion: out of official writing and the courts, then out of work and trade, then out of school, then out of conversation between friends, and out of the kitchen table last. The occasion held longest is the last to go and the easiest to misread — it is still there, so the whole thing looks like it is still there.
Usefully, there is an earlier signal. Dorian tracked the Gaelic of several fishing villages in East Sutherland, Scotland, from 1963 for four decades, and her 1981 book introduced the term semi-speaker: people who understand and can speak, but whose output has changed. This dialect built noun plurals several ways — changing a vowel in the stem, adding an ending, even subtracting a final sound — and the youngest group levelled almost everything to one suffix, or marked no plural at all. That morphological contraction arrives well before the population curve turns, and you can see it without any demographic data at all.
The first question belongs to the field. In 1963 Greenberg generalised from thirty languages and his first universal was that subjects almost always precede objects. In the late 1960s Derbyshire went to the Amazon to describe Hixkaryána, a Cariban language of Brazil with a few hundred speakers that nobody had written up before, and found its basic order puts the object first: "the jaguar grabbed the man" comes out as toto yahosïye kamara — first toto ("the man"), then yahosïye ("grabbed him"), and kamara ("jaguar") last. In a sample of 1,377 languages, only eleven are object-initial. So whether our "all languages …" sentences hold depends on a few dozen languages at the very edge — precisely the ones with the fewest speakers and the shortest time left.
The second question belongs to the community, and it is the one most often told badly. The usual line is that a worldview dies. But as the topic on the rise and fall of Whorf set out, the strong version did not survive, so that argument cannot carry the weight put on it — and it does not need to. What people themselves report losing is concrete: whether they can finish a thought with an elder, whether a ceremony can still be held, whether the place names still mean anything. There is also a much-cited correlation: a 2007 study across First Nations bands in British Columbia found that in communities where more than half of members reported conversational knowledge of their own language, youth suicide rates were close to zero, and about six times higher below that threshold. Evidence strength: a band-level correlation, not individual-level, and not causal — the original authors themselves compared it alongside six other measures of cultural continuity.
The third question is reversibility. Wampanoag had no speakers left after the nineteenth century, but it left a body of documents: a complete Bible translation from 1663, plus a stack of deeds, wills and petitions written in the language by Wampanoag people themselves. In 1993 jessie little doe baird started the Wôpanâak Language Reclamation Project and rebuilt the language from that material; her daughter, born in 2004, became the first native speaker in over a century, and the tribe now runs a full-time immersion school, Mukayuhsak Weekuw ("the children's house"). Documents cannot keep a language alive, but they settle the question of whether it remains possible at all.
The one case of a language going from no native speakers back to a community's everyday first language is Hebrew. The preconditions were unusual: it had never stopped being written and recited, its users were immigrants with no shared spoken language between them, and state power followed. The outcome is disputed too — Zuckermann argues that the modern Israeli language is a two-source hybrid, its vocabulary from Hebrew and its grammatical footing carrying the revivalists' native Yiddish. That the argument is possible at all tells you what "revival" means.
More reproducible is the language nest. In 1982 Māori communities opened the first kōhanga reo: preschool children spend whole days there in the target language only, looked after by elders who still speak it. Hawai'i followed the model with Pūnana Leo in 1983. A count at the time found fewer than fifty fluent speakers of Hawaiian under eighteen statewide, almost all of them on Ni'ihau; today the figure is in the thousands. For Māori, the 2023 census recorded 213,849 people reporting they could hold a conversation in it, about 15% more than in 2018 — self-reported ability, which is not the same as daily use.
And one example small enough to remember. Ned Maddrell died in 1974; he was the last person raised in a community where Manx was spoken. In 2009 UNESCO listed the language as extinct, and pupils at the Manx-medium primary school wrote in — in Manx — to ask what language they were writing in if theirs was extinct. The classification was later revised to critically endangered.
Put together, the successful projects all do the same thing: they put the language back into an occasion that happens every day without fail, and that occasion is usually the hours spent looking after small children. What fails is usually not underfunding but money spent only on class hours, textbooks and orthography — all of which sit at outer levels of the scale and never reach the one at home. That is the reverse test of the mechanism in "the dividing line is not the last speaker". As for the revived language not matching the old one: that was never a standard anyone could demand, since no language matches itself three generations on.
Latin did not disappear; it became French, Italian, Spanish. The chain never broke, it just shifted a little each generation. Disappearance means the chain breaks: no next generation takes it up and no descendant language exists. The criterion sits on transmission, not on resemblance — which also shows why "the language is deteriorating" makes no sense here. Change is what being alive looks like.
Documents give you forms, not when to say what. The people rebuilding Wampanoag had a Bible, deeds and petitions, and not one argument, one joke, one interruption. So a revived language is necessarily newly made at the pragmatic layer, mostly along the habits of the revivalists' own language — which is not a failure but the actual shape of the thing.
Speaker numbers, phylogenetic isolation, and whether usable material survives produce three quite different lists, and none of them is jointly optimal. More to the point, all three are external measures. What decides whether a language gets spoken in somebody's house tomorrow is not a ranking.