You have probably heard that English and German are relatives, or that Chinese and Tibetan are. The question is on what grounds. There are over seven thousand languages; take any two and their wordlists will throw up a few pairs that sound alike. Yet the claim is solid enough to be used for inferring where ancient people moved and when they set out. How is that step taken?
The procedure is called the comparative method: using living descendants to reconstruct an ancestor for which there is no record. Three things about it — what it accepts as evidence, how it reconstructs a word nobody has ever heard, and where it breaks down. The last one matters most: only a method that knows its own boundary is entitled to say any of the above.
First, throw out resemblance. Besides the Persian bad, there is another pair every course covers: Greek theos and Latin deus both mean "god" and look alike, yet they are not relatives, each descending from its own root. The genuine relative of Latin deus is Greek Zeus, the one that looks nothing like it. Words that look alike need not be family, and family need not look alike.
What counts is a wholesale, repeatable correspondence. Latin word-initial f turns up in Spanish as an h that is written but not pronounced: filius against hijo ("son"), facere against hacer ("to do"), ferrum against hierro ("iron"). These three are not coincidences; the f in that position went down the same road across the vocabulary, while Italian and French did not follow (figlio, fils). So the test is not whether two words resemble each other, but whether one sound in one position stands reliably opposite another.
The prettier part is that even the exceptions obey rules. Latin focus ("hearth fire") gives Spanish fuego with the f intact, yet focaria, built on the same root, gives hoguera ("bonfire") with h. The difference lies in what follows: the f stays before the diphthong ue and changes elsewhere. A rule that looks broken usually just means the conditioning has not been written out in full.
The same test also throws the loanwords out. Spanish fiesta ("festival") and febrero ("February") keep their initial f because they were not handed down continuously; they were borrowed back from written Latin after the change had already run. English shows it more starkly: the native words are father, fish, foot, while the cognates borrowed from the Latin side are paternal, piscine, pedal — f on one side of the same slot, p on the other. A word's first consonant tells you which door it came in through. Correspondence is not only evidence; it is also a sieve.
The craft itself is not mysterious. Line up cognates — sets of words descended from one ancestral form — by meaning: how each language says "son", how each says "do". Then pull out the sounds that line up within each set and check whether they line up the same way in other sets. Finally choose an ancestral form such that the fewest and most natural steps lead from it to every descendant. Direction matters too: stops between vowels tend to soften, and rarely the reverse.
Why trust this reasoning? Because one family lets you check the answer: the ancestor of the Romance branch left a great deal of writing behind. You can pretend not to know Latin, reconstruct from Spanish, Italian, French, Portuguese and Romanian, and then open the books. It matches — with one telling offset. What comes out is not the classical Latin of the textbooks but something closer to the version people actually spoke. The comparative method always gives you the latest common ancestor of the daughters, not the written standard of the same era.
Nor is the craft confined to Europe. The sounds of Middle Chinese were likewise reconstructed. The 切韻 (Qieyun, a rime dictionary of 601) spelled each syllable using two other characters and sorted the characters into classes, but never said how those classes were pronounced. Bernhard Karlgren filled in sound values by triangulating three sources: those classes, the pronunciations of modern regional varieties, and the Chinese readings borrowed wholesale into Japanese, Korean and Vietnamese. The reconstructions are still being revised; the procedure is the same one.
The hardest test came in 1879. Working on alternations of vowel length in Indo-European, Ferdinand de Saussure argued that a class of sounds found in no language then known had to be posited for those alternations to make sense. For half a century the claim was treated as an outsider's eccentricity. Early in the twentieth century, tablets in Hittite — an ancient language of Asia Minor — were deciphered, and in 1927 Kurylowicz pointed out that a Hittite consonant turns up exactly in the positions Saussure had predicted. A reconstruction is not a restatement of existing material. It is a commitment: if a new relative is ever unearthed, this is what it will have to look like.
If the method is this powerful, why has nobody joined all the world's languages into one tree? Because the signal runs out. Every thousand years another batch of cognates is replaced, further sound changes pile on top, and correspondences get harder to recognise; meanwhile chance look-alikes keep occurring at a fixed rate. Past some depth the real signal drops below the noise — the problem is not that resemblances are missing, but that they no longer show anything.
How large the noise is can be computed, and computing it is exactly what Ringe did in 1992: estimating on pure probability how many pairs of similar sound and similar meaning two wordlists should throw up, and arriving at a figure far above most people's intuition. His target was the practice of laying wordlists from a hundred languages side by side and hunting for similarities by eye. The dispute has never entirely settled, but one side of it is clear: resemblance with no systematic correspondence behind it cannot be told apart from accident.
So the ceiling sits at roughly eight thousand years — the figure usually quoted is 8000 plus or minus 2000, with an upper bound no deeper than ten thousand, partly because the oldest securely established families are about that old. Deeper proposals, such as joining Indo-European and Uralic into a larger Nostratic grouping, are not in the position of having been refuted; they cannot reach the evidential bar. At that depth, what would support them and what chance alone produces look the same.
The current frontier borrows statistical tree-building from biology, coding the presence or absence of cognates as characters and producing dated trees. Indo-European origins have been argued over on this basis for two decades: one camp roots the family in Anatolia and follows the spread of farming, the other on the East European steppe, following migration evidence from ancient DNA. In 2023 Heggarty and colleagues published a root age of about 8120 years in Science and proposed a hybrid in which each side holds a stretch. The strength of that evidence should be stated plainly: it is a model posterior, and it shifts when cognate judgements or rate assumptions change.
Shifts by how much? Look at Sino-Tibetan. In 2019 two teams applied comparable methods to the same family, reached similar conclusions about geography, and landed on about 7200 years and about 5900 years — more than a millennium apart. The method is real and so are the error bars. The ceiling is the same kind of fact: not a shortfall of craft, but information eaten by time. No algorithm, however ingenious, reads out a signal that is no longer there.
It is not circular, but it is not insurance either. Body parts, numerals and pronouns are empirically less likely to be borrowed, and that is all: English they, them and their were borrowed from Old Norse, although pronouns are usually held to be the hardest of all to borrow. The wordlist only lowers the noise. What actually carries the weight is still correspondence, since loans tend to give themselves away in the sound matches — like the Spanish f that never became h.
The comparative method assumes by default that the ancestor is a single homogeneous point, and no real language is homogeneous. If the ancestor had already split into dialects when the daughters diverged, reconstruction flattens those differences and yields a system that contradicts itself; in practice such contradictions are often written up as "several variants coexisting in the proto-language". The recent statistical models that let documented ancient languages sit on the tree as sampled points are aimed at exactly this layer.
Almost certainly the latter. Every language has ancestors, and nearly every one once had related neighbours. "Isolate" is a statement about the evidence we hold, not about the language itself — its cousins either disappeared or separated so long ago that they fell past this issue's ceiling.