The strongest evidence that two languages descend from one ancestor is not a word they share. It is a verb that is irregular in both of them, in the same way.
Indoor weather, and late. Lamplight on a page instead of daylight on a landscape — the climate of a desk at the hour when someone has looked at the same four words long enough for the differences between them to line up. Nothing at continental scale has happened yet; nothing is coloured and nothing branches. A short column of type, held still for as long as the alignment takes to arrive unassisted. Close, patient, textual, and confident enough to let four words stand there for a while before anything explains them.
The family tree of languages is not a picture of resemblance. It is the residue of a proof that refuses resemblance as evidence — and the same refusal that makes descent demonstrable is what leaves the field's largest questions permanently, publicly open.
Language is the oldest evidence we have for where people came from — older than most archaeology and all writing — because most human movement happened before anyone wrote anything down. What makes that evidence worth an argument is the standard behind it. Comparative linguistics throws out resemblance and reasons instead from differences that repeat, a discipline strict enough that it once described consonants no attested language showed, on the sole ground that the system required them, and left the claim sitting where an archive coming out of the ground could have destroyed it. It did not. That strictness is expensive, and the corpus does not hide the bill: a horizon at roughly six to ten thousand years, nothing to compare when a relative dies undescribed, and no agreed answer to where its own founding family began. A field able to state precisely where its evidence runs out is more useful to think with than a field that is right.
The families are not the subject, and neither is the tree they get drawn as. Underneath both sits the standard of proof that has to hold before either may be drawn at all: one refusal — that two things looking alike establishes nothing — carried into every place where it leaves the record silent.
This subject runs on more than one time axis. The chapters are not one sequence.
The reader loses the instrument they arrived with. Word resemblance stops being weak evidence and becomes no evidence, and the question turns from what languages have in common into what could possibly count.
The comparative method · Contact, borrowing and the limits of resemblance · Sir William Jones · Sanskrit · Indo-European · Dravidian
Regularity replaces similarity as the standard of proof. The reader can see why a rule reaches words nobody has looked at yet, and why a borrowed word arrives too late to show the shift.
The regularity of sound change · Grimm's Law — the first Germanic sound shift · Jacob Grimm · The comparative method
A residue of disobedient words becomes a second law, and reconstruction stops looking like convenient notation. The reader watches a property of a language nobody has ever heard do real explanatory work.
Verner's Law — the exception that rescued the rule · Grimm's Law — the first Germanic sound shift · Sanskrit · The regularity of sound change
The discipline acquires the property readers rarely grant the humanities: a claim about the unobservable that later evidence could have destroyed, did not, and reshaped the family tree instead.
Laryngeal theory · Ferdinand de Saussure · Hittite · Hattusa (Boğazkale) · Anatolia · Indo-European
The tree becomes a map. The reader learns to find a homeland by counting divisions rather than speakers, and then watches that inference agree with a settlement chronology assembled from entirely other material.
Austronesian · The Austronesian expansion · Taiwan · Malagasy · Trans-New Guinea
Runs alongside another chapter, not after
The first chapter's refusal comes back at continental scale, and the reader sees it applied unevenly to the same scholar's two proposals. The standard stops being a principle stated once and becomes a working border the discipline patrols.
Joseph Greenberg · Afroasiatic · Niger-Congo · The Bantu expansion · The comparative method
Disagreement stops reading as somebody's error. The reader can say what each kind of evidence is competent to answer — content, date, demography — and therefore why three competent answers do not add up to one place.
Proto-Indo-European · Indo-European · The Pontic–Caspian steppe · Anatolia · Bayesian phylogenetic dating of language families · Ancient DNA as a check on linguistic prehistory · Glottochronology · Sino-Tibetan · Uralic
An isolate stops being a curiosity and becomes a reading of an absence. The reader leaves knowing that the evidence base for everything above is being consumed while it is being used, and roughly where it runs out.
The language isolate · Basque (Euskara) · Sumerian · Language death · SIL International (Ethnologue) · Glottolog · The comparative method
37 records in this Journey.
Showing 8 of 20 featured records. Atlas does not choose which of the rest matter.
The rule that fixed a Germanic sound change could only be found by reading Germanic against a Vedic transmission that had no interest in Germanic whatsoever. A tradition of exact recitation preserved, as a by-product, the one piece of information that made a European law come out regular — and it had already been preserved for millennia before anyone had a use for it.
Verner's Law — the exception that rescued the rule · Sanskrit · Grimm's Law — the first Germanic sound shift
An island off the coast of Africa holds a language whose nearest known relative is spoken in southeastern Borneo, and the relationship was established from sound correspondences before either archaeology or genetics could speak to it. A migration nobody recorded was read out of the words of the people who made it.
The process that produces the subject's most striking objects is the same process that destroys its evidence. Every last relative that dies manufactures a new isolate, and every undescribed language that ceases to be spoken removes correspondences the method needs — so the field's catalogue of curiosities grows precisely as its capacity to explain them shrinks.
The language isolate · Language death · The comparative method
One family has branches documented for roughly five thousand years, which makes it the place where reconstruction can be checked against a written record instead of only against its own output. The family least discussed in popular accounts of this subject is the one that offers the method its rarest thing: an answer key.
Two large families, and the gap between how securely each is established is not a fact about the languages. It measures what two centuries of documentation and a long written attestation contribute to a demonstration — which means the most famous family in the subject is also its least representative.
Where was Proto-Indo-European spoken?
Four lines of evidence, each competently pursued, and no agreement. Reconstructed vocabulary places the speech community after certain technologies existed; two Bayesian phylogenies supported an Anatolian farming dispersal; ancient genomics found a large third-millennium movement out of the steppe; the most recent analysis proposes a hybrid that endorses neither. The Atlas graph carries two origin edges from Proto-Indo-European deliberately, so that browsing does not silently settle a question the field has not.
How old is Sino-Tibetan?
Two phylogenetic studies published in the same year agree on a northern Chinese origin and differ on the root age by roughly a thousand years. The corpus records both and adjudicates neither. The family's structure is also less settled than its size would suggest, because the material the comparative method depends on is thinner here.
How much of Saussure's prediction did Hittite actually confirm?
How many laryngeals there were, what they sounded like and how each behaved in each environment are all still argued, and the share of predicted positions where Hittite shows a reflex is a matter of detailed dispute. The corroboration is strong and it is not total; the pack quotes no figure for it, and this package does not supply one.
Do sound laws really admit no exception?
The exceptionless doctrine is a working assumption rather than an established fact. Later work on lexical diffusion and on sociolinguistic variation shows changes that genuinely do spread word by word, and the field continues to argue about how much of the doctrine survives while continuing, in practice, to assume it. Nothing in this Journey is safe from that qualification, including its strongest chapter.
How quickly did Austronesian speakers move, and when was Madagascar settled?
The rapid model the 2000 analysis supported is not the only account; slower and more tangled versions with longer pauses in island Southeast Asia remain in play. Madagascar's settlement date is contested by centuries and its route is unresolved, and the Eastern Polynesian chronology has been revised substantially in the past two decades. The branching order is far better established than the calendar.
Does Basque have any relative beyond Aquitanian — and could evidence ever settle it?
Its ancestry is traceable exactly one step, and beyond that every proposal has failed to produce correspondences at the scale the method requires. The deeper problem is structural rather than empirical: no amount of evidence can prove that a language has no relatives, so the isolate verdict is always a report on what has been found and never a finding about what exists.
How many families are there, and how much of the past is out of reach?
There is no single answer to the count: catalogues that treat each isolate as a family of one report far more families than catalogues that group aggressively, and the two major catalogues return materially different totals from the same speech communities. The pack quotes no figure. Behind the counting problem sits a harder limit — beyond roughly six to ten thousand years the signal degrades past recovery, so a large part of human linguistic history is not merely unknown but unreachable by any method now available.
What survives from before writing is not a record anybody kept. Nobody set out to leave a consonant shift or a branch structure, and the people who left them could not have read them. That is exactly why the discipline is so severe about what counts: an unintended trace has no author to vouch for it, so the only guarantee available is that chance could not have produced it and a borrower could not have carried it. The price is written into the same corpus — a horizon at roughly six to ten thousand years, and no way to compare a relative that died undescribed. On the family the method was built to explain, two homelands stay lit at once. None of that is a finished position. The two major catalogues are returning different totals from the same speech communities, and the count of families moves as they revise. Every last relative that dies is manufacturing a new isolate, and every undescribed language that stops being spoken is taking correspondences the method runs on out of reach: the evidence is being consumed while it is being read. The disagreement stays on the graph, in public, and the material it argues over keeps getting smaller.
This is unfinished in the world, not only in the telling.
Ask the world another question. Dynamic keeps this Journey as its context and never changes what is written above.
Open Dynamic Atlas →