Historical Linguistics & Language Families

A sentence from around the year 1000

Here is the opening of the Lord's Prayer as it stands in the West Saxon Gospels, an Old English translation copied in England around the year 1000: Faeder ure thu the eart on heofonum, si thin nama gehalgod. In the manuscript the th sounds are written with the letters thorn and eth, which English later abandoned. Read it aloud and you can catch a few words. Faeder is father. Ure is our. Nama is name. But eart on heofonum for art in heaven is already strange, si is a subjunctive verb form English no longer has, and the word order of thu the eart, literally thou who art, is not modern word order. This is English. It is not a foreign language, and no one alive could read a page of it without training.

Four hundred years later, Chaucer began the Canterbury Tales with Whan that Aprille with his shoures soote, the droghte of March hath perced to the roote. Now the vocabulary is nearly all recognisable, though soote means sweet and his means its. Two hundred years after that, Shakespeare is difficult only in patches. Today you are reading this. Nobody along that chain ever decided to change anything, no committee met, and no generation noticed the shift happening to it. And yet at every point the language was in motion, and the motion added up.

Key idea: Language change is total rather than lexical. Sounds, endings, word order, and meanings all move, and over a thousand years the accumulated movement makes an earlier stage of the same language unreadable without study.

Three kinds of evidence, and what each is good for

A course in historical linguistics is not asking you to take change on faith. There are three independent evidence streams, and they check each other.

The first is the philological record: texts written at known dates. This is the strongest evidence where it exists, because it is dated and physical, and it is the reason Indo-European is the best reconstructed family in the world. It also has hard limits. Writing is conservative, so spelling lags pronunciation by centuries. Scribes copy the forms they were taught rather than the forms they use. Written texts over-represent the literate, the formal, and the powerful, and for most of the world's roughly seven thousand languages there is no historical record at all.

The second is real-time observation: recording the same community, or even the same person, at intervals. Jane Harrington and colleagues did this with an unusually clean sample, the annual Christmas broadcasts of Queen Elizabeth II, and measured her vowels across four decades. Her speech had shifted measurably toward the more general southern British pattern of the later twentieth century. If the vowels of a monarch whose accent was a national reference point moved while she was using them, nobody's are fixed.

The third is apparent time: comparing older and younger speakers in one community at one moment and reading the age difference as a picture of change in progress. This is the workhorse method of variationist sociolinguistics, and if you took LING 320 you have met it already. It rests on an assumption that has to be argued rather than assumed, namely that a speaker's vernacular is broadly set in adolescence and does not drift much afterward. That assumption is a good approximation and not a law, which is exactly why real-time studies matter: they are the check.

Why this matters: No single line of evidence is sufficient. The written record is dated but conservative and patchy; real-time studies are direct but slow and rare; apparent time is fast but rests on an assumption. Historical linguistics is built on their convergence.

The Great Vowel Shift, and why English spelling looks broken

The most useful single example of a large, documented change is the Great Vowel Shift, which reorganised the long vowels of English roughly between 1400 and 1700. Every long vowel moved up in the mouth, and the two that were already at the top broke into diphthongs. Nothing was lost: the system stayed the same size and simply rotated.

Word	Middle English vowel	Modern English vowel	What happened
bite	ee as in machine	eye	Already highest; broke into a diphthong
house	oo as in boot	ow	Already highest; broke into a diphthong
meet	ay as in Spanish e	ee	Raised one step into the slot bite vacated
boot	oh as in Spanish o	oo	Raised one step into the slot house vacated
name	ah	ay	Raised one step
boat	aw	oh	Raised one step

Now the payoff. William Caxton set up the first English printing press at Westminster in 1476, near the start of the shift. Print fixed spellings while the vowels underneath them kept moving for another two centuries. That is why the letter i in machine and the letter i in mine represent different sounds, why the spelling of great and meat and threat no longer agree, and why English learners are told the vowels have no rules. The spelling is not broken. It is an accurate photograph of a pronunciation that has since walked out of the frame.

Notice also what the shift was not. It was not a single event, it did not happen everywhere in the English-speaking world at the same pace, and Scottish and northern varieties came through it differently, which is one reason a Scots speaker says hoose where a southern English speaker says house. Change spreads through communities the way anything social does, unevenly.

The point: A single systematic sound change, dated and traceable, explains a whole class of apparent irregularities in English spelling, because writing froze while speech kept moving.

Change is not decay, and here is the receipt

Every generation for as long as we have records has believed that the language is being ruined by the young. In 1712 Jonathan Swift published a Proposal for Correcting, Improving and Ascertaining the English Tongue, arguing that English had been in decline since the Restoration and asking for an academy to fix it permanently. Three centuries later, English is used by more people for more purposes than any language in history. Swift's specific complaints, including his loathing of shortened words like mob, are now invisible.

The best evidence against the decay story is older and better. Somewhere in the late Roman period a teacher compiled the document known as the Appendix Probi, a list of corrections in the form say this, not that: auris non oricla, calida non calda, vetulus non veclus, speculum non speclum. Each left-hand form is the correct classical Latin word. Each right-hand form is the vulgar error being stamped out. Now look at where the errors went. Latin oricla is the ancestor of Spanish oreja and Italian orecchia, both meaning ear. Calda gives Italian caldo, hot. Veclus gives Italian vecchio, old. The forms the schoolmaster condemned are the forms that survived and became the Romance languages. The forms he defended are the ones that died with the schoolroom.

Nor do languages simply lose things. English did shed most of its case endings between roughly 1000 and 1400: Old English had four cases and three genders on nouns, and modern English has a possessive marker and nothing else. But look at what English built while it was dismantling. It developed a rigid subject verb object order that carries the grammatical information the endings used to carry, an elaborate auxiliary system (do, have, be, will, would, may, might, must) that expresses tense, aspect, mood, and negation with a precision Old English could not match, and a set of phrasal verbs that no Anglo-Saxon would recognise. Complexity was not lost. It moved.

Remember: The forms that prescriptivists condemn are frequently the forms that go on to become the standard language of the next era, and languages that simplify in one part of the grammar reliably elaborate in another.

Why change happens at all

Four causes are usually distinguished, and real cases involve several at once.

• Articulation. Speech is a physical act, and adjacent sounds influence each other. The n of in becomes an m before a p in the ordinary pronunciation of input, because the lips are already closing for the p. Repeated often enough by enough speakers, an ease-driven adjustment becomes the standard form.
• Perception and acquisition. Children do not receive a grammar. They infer one from a finite sample of speech, and any inference can land slightly off the target. Multiply small mislandings across a generation and the grammar shifts. This is why the transmission bottleneck between generations is where most structural change is now thought to originate.
• Social evaluation. Variants carry meaning about who you are. Speakers adopt forms associated with people they want to resemble and avoid forms associated with people they do not. William Labov's 1963 study on Martha's Vineyard showed islanders unconsciously exaggerating a local vowel pronunciation in proportion to how strongly they identified with the island against summer visitors.
• Contact. Speakers of other languages supply words, sounds, and occasionally grammar. English took skirt from Norse alongside its own inherited shirt, so the language now has two words from one ancestral root with split meanings.

The upshot: Change is not one process. Physical, cognitive, social, and contact pressures all act on language simultaneously, which is why no single explanation covers every case.

The uniformitarian principle

One assumption underwrites the whole enterprise, and it is worth stating plainly because everything later in the course leans on it. The uniformitarian principle holds that the forces acting on languages today are the same forces that acted on them in the past. Ancient speakers were not cognitively different from you. Proto-Indo-European was a language of the ordinary kind, with dialects, slang, style-shifting, and speakers who found the young sloppy.

The principle is doing real work. It is why a reconstruction that requires an ancestor with no vowels, or with a sound change that has never been observed anywhere, is rejected as implausible. It is why we can use the typology of living languages, what is common, what is rare, what never occurs, as a constraint on what we reconstruct. And it is why the direction of change matters: some changes are common and effectively one-way, such as k becoming ch before a front vowel, while the reverse is vanishingly rare. That asymmetry is a tool, and Lesson 2 will put it to work.

Key idea: Reconstructions are constrained by what languages actually do, because the past is assumed to have obeyed the same processes we can observe operating now.

Common misconceptions

• Language change means language decline. The Appendix Probi condemned the exact forms that became Spanish and Italian. Judgments of decline track unfamiliarity, not any measurable loss of expressive capacity.
• Writing preserves a language. Writing preserves spellings, which is a different thing. English spelling froze around 1476 while the vowels kept moving, which is the source of most of its irregularity.
• Isolated languages do not change. Icelandic is often cited as frozen, and its written form is unusually conservative, but Icelandic pronunciation has changed substantially since the sagas were composed. No spoken language stands still.
• Change happens because people are lazy. Ease of articulation is one pressure among several, and it does not explain changes that make speech harder, such as the breaking of a long vowel into a diphthong, which the Great Vowel Shift did twice.
• We can watch a language change if we listen carefully enough. Individuals rarely perceive change in progress. It is recovered by measurement across decades or by comparing age groups, not by ear.

What to carry forward

• Change touches every level of a language, and a thousand years of it makes an earlier stage of English unreadable to its own descendants.
• Three evidence streams support the study of change: dated texts, real-time recordings of the same speakers, and apparent-time comparison of age groups.
• The Great Vowel Shift rotated the English long vowels between about 1400 and 1700 and, because print froze spelling in the middle of it, produced most of the mismatch between English orthography and speech.
• The Appendix Probi shows condemned forms becoming the Romance languages, and English shows inflectional loss compensated by word order and auxiliaries. Neither supports a decay model.
• Articulation, acquisition, social evaluation, and contact all drive change, usually together.
• The uniformitarian principle licenses using observed processes in living languages as constraints on what may be reconstructed for the past.

Sources

1. Linguistic Society of America. (n.d.). Is English changing? linguisticsociety.org
2. Britannica. (n.d.). English language. Encyclopaedia Britannica. britannica.com
3. Wikipedia contributors. (n.d.). Great Vowel Shift. en.wikipedia.org
4. Labov, W. (1963). The social motivation of a sound change. Word, 19(3), 273-309. doi.org
5. Harrington, J., Palethorpe, S., & Watson, C. I. (2000). Does the Queen speak the Queen's English? Nature, 408(6815), 927-928.
6. Campbell, L. (2020). Historical linguistics: An introduction (4th ed.). MIT Press.

Six words in four Polynesian languages

Start with data rather than theory. Here are six ordinary words in Tongan, Samoan, Maori, and Hawaiian. The apostrophe stands for a glottal stop, the catch in the middle of English uh-oh. The letter g in Samoan spelling represents the ng sound of English singer.

Meaning	Tongan	Samoan	Maori	Hawaiian
eye	mata	mata	mata	maka
three	tolu	tolu	toru	kolu
forbidden, sacred	tapu	tapu	tapu	kapu
house	fale	fale	whare	hale
canoe	vaka	va'a	waka	wa'a
person	tangata	tagata	tangata	kanaka

Nobody wrote Proto-Polynesian down. There is no Polynesian equivalent of Latin sitting in a library waiting to settle the question. And yet by the end of this lesson you will have reconstructed a good deal of its sound system from this table alone, and you will be able to say what each of these four languages did to get from there to here. That procedure is the comparative method, and it is the single most powerful technique historical linguistics has.

Key idea: The comparative method recovers an unattested ancestor from its surviving descendants by treating systematic correspondences between them as evidence of shared inheritance.

Step one: assemble candidate cognates

A cognate is a word in two or more languages that descends from a single word in their common ancestor. You cannot know in advance which words are cognate, so you begin with candidates: words that are similar in form and similar in meaning across the languages you are comparing.

Two practical rules shape the search. First, look in basic vocabulary: body parts, low numerals, kinship terms, common natural objects, basic verbs, pronouns. These are learned early, used constantly, and borrowed less often than words for trade goods, technology, or religion. Second, do not insist on identical meaning. Semantic drift is normal, so a word meaning house in one language may correspond to one meaning roof or household in another. Latin hostis meant stranger and then enemy, while its English cognate guest went the other way. Both descend from one root.

A warning about looks. Resemblance is where you start, not what you conclude. Modern Greek mati and Malay mata both mean eye and are not related in any way: Greek and Malay belong to different families and the Greek word is a reduced form of an older Greek word. English much and Spanish mucho mean the same thing, look the same, and are not cognate at all: much descends from Old English mycel, mucho from Latin multus. Similarity is cheap. Systematicity is what costs something to fake.

The point: Cognate candidates come from basic vocabulary and need not match in meaning exactly, but visual resemblance on its own proves nothing.

Step two: extract the correspondence sets

Now stop looking at words and start looking at positions. Line the words up sound by sound and ask what corresponds to what.

In eye, three, forbidden, and person, wherever Tongan, Samoan, and Maori have t, Hawaiian has k. That is not one word behaving oddly. It is a rule holding across the whole vocabulary, and it is our first correspondence set: t : t : t : k.

Keep going and the table fills out.

Correspondence (Tongan : Samoan : Maori : Hawaiian)	Seen in
t : t : t : k	eye, three, forbidden, person
k : ' : k : '	canoe
l : l : r : l	three, house
f : f : wh : h	house
v : v : w : w	canoe
ng : g : ng : n	person
m : m : m : m; a : a : a : a	everywhere

Notice the crucial structural fact. The set t : t : t : k and the set k : ' : k : ' are different sets. Hawaiian k appears in one and Hawaiian glottal stop in the other, and the two never trade places. Two distinct correspondence sets require two distinct proto-sounds, even when one language has merged something. This is the discipline that separates the comparative method from eyeballing.

Why this matters: The unit of evidence is the correspondence set, not the word pair. One set can be an accident; a system of sets, each holding across the vocabulary, cannot be.

Step three: rule out the other three explanations

Systematic resemblance has exactly four possible causes, and common descent is only one of them. Before reconstructing anything, you have to eliminate the other three.

Chance. Any two languages share some accidental look-alikes. What chance cannot produce is dozens of matches that all obey the same sound rules in the same positions. Don Ringe showed in 1992 how to compute the probability of a given number of accidental matches, and the answer for a real correspondence system is vanishingly small. Our Polynesian table is not four coincidences; it is one rule applied to every word containing the sound.

Borrowing. Words travel. Finnish kuningas, king, is not evidence that Finnish is Germanic; it is a loan from an early Germanic form, and it is famous precisely because Finnish preserved the old shape better than any Germanic language did. Borrowing is detected by several signals: loans cluster in culture words rather than basic vocabulary, they often fail to show the sound changes the native words underwent, and they can appear in one branch of a family and not others. If Hawaiian had borrowed its words for eye and person from a neighbour, they would not have obeyed the same t to k rule as everything else.

Universals. Some resemblances arise everywhere for reasons that have nothing to do with history. Onomatopoeic words converge because they imitate the same sounds. Nursery words converge harder still: forms like mama, papa, nana, dada turn up worldwide because they are built from the earliest sounds infants produce, which is why the Chinese word for mother and the English word for mother resemble each other without any relationship between the languages. Exclude these categories from your evidence.

What is left, once chance, borrowing, and universals are excluded, is common descent. That is the inference the comparative method licenses, and nothing weaker.

Remember: Resemblance has four possible sources. A relationship claim is only as good as the argument that eliminated the other three.

Step four: reconstruct, using direction rather than majority

Now the interesting part. Take the set t : t : t : k. Three languages have t and one has k, so a vote would give you t. But majority rule is a bad principle, because it counts languages rather than weighing evidence, and closely related daughters can inherit the same innovation from a shared intermediate ancestor and outvote a conservative outlier.

The right criterion is directionality: reconstruct the form from which all the attested forms can be derived by changes that languages are actually known to undergo. Consider the alternative. If the proto-sound were k, then Tongan, Samoan, and Maori all independently changed k to t, while Hawaiian alone kept it. That is a rare change. If the proto-sound were t, Hawaiian alone changed t to k, and everyone else kept it. Fronting and backing of stops both occur, but there is a second consideration that settles it, and it is the elegant part.

Look at the other set, k : ' : k : '. Reconstruct that as proto k, since weakening of k to a glottal stop is extremely common cross-linguistically and the strengthening of a glottal stop into k is rare. So Samoan and Hawaiian turned proto k into a glottal stop. That freed the k slot in Hawaiian, and proto t moved into it. Hawaiian did not perform two unrelated changes. It performed a chain shift: k became a glottal stop, and t moved up into the space k had left. Reconstructing proto t and proto k makes Hawaiian a single coherent event. Reconstructing anything else makes it a coincidence.

Apply the same reasoning across the table and Proto-Polynesian comes back with the consonants p, t, k, m, n, ng, f, s, w, l, h and the five vowels a, e, i, o, u, which is very close to what specialists reconstruct from a far larger dataset. Then state what each daughter did.

Language	Changes from the reconstructed ancestor
Tongan	w becomes v; otherwise conservative in these sets
Samoan	k becomes glottal stop; w becomes v
Maori	l becomes r; f becomes wh
Hawaiian	k becomes glottal stop; t becomes k; ng becomes n; f becomes h

By convention, reconstructed forms carry an asterisk: the ancestor of the first row is written as an asterisk followed by mata. The asterisk is not decoration. It marks the difference between something attested and something inferred, and this course uses it strictly.

The core of it: Reconstruct the form from which the others follow by changes languages are known to make, and prefer the reconstruction that turns a set of separate oddities into one coherent event.

Grading the method where the answer is known

All of that is only as good as its track record, and Romance is where the method gets marked. Spanish, Italian, French, Portuguese, and Romanian descend from Latin, and Latin is sitting in the library.

Take the word for to sing: Spanish cantar, Italian cantare, Portuguese cantar, Romanian canta, French chanter. The correspondence before the vowel a is k : k : k : k : sh. Directionality says reconstruct k, because k regularly becomes a palatal or postalveolar sound before certain vowels while sh rarely hardens to k. Latin has cantare. The method wins.

Now change one input and watch the result flip. Take the word for hundred: Italian cento with an initial ch sound as in church, Spanish ciento with th or s, French cent with s, Portuguese cento with s. Here the correspondence is not k at all. But the environment has changed: the following vowel is e rather than a. Reconstruct k once, and add a conditioned change, palatalisation of k before front vowels, which happened across most of the Romance area and not in the environment before a. Latin centum, pronounced with an initial k, confirms it exactly. One proto-sound, two environments, two sets of reflexes.

Romance also shows you the method's honest failures. The Romanian word for father is tata, which is not descended from Latin pater at all but from the nursery word tata. Fill in that cell mechanically and you will reconstruct nonsense. And Spanish perro, dog, has no accepted Latin source, while Italian cane, French chien, and Portuguese cao all continue Latin canis. A meaning slot is not guaranteed to hold a cognate, and a good analyst leaves cells empty rather than forcing them.

Bottom line: Run on a family whose ancestor survives, the comparative method reconstructs Latin closely, including conditioned changes it was never told about. That is the evidence that its results elsewhere are worth trusting.

What the method cannot do

Four limits matter, and honest practitioners state them up front.

It cannot recover what every daughter lost. If Proto-Polynesian had a sound that disappeared in all four languages without a trace, nothing in the data can reveal it. Reconstructions are therefore always a floor on the ancestor's complexity, never a ceiling.

It reconstructs a language without variation. Real Proto-Polynesian had dialects, registers, and change in progress, like every language. The comparative method flattens all of that into a single idealised system, and the flatness is an artefact of the method, not a fact about the past.

It depends on regularity, and regularity has exceptions. Borrowing, analogy, and irregular reduction in very frequent words all break the neat correspondences. A residue of exceptions is normal; a majority of exceptions means you have the wrong analysis.

It has a time depth limit. As shared vocabulary erodes, the number of correspondence sets available thins out until chance resemblances become indistinguishable from inherited ones. Where exactly that boundary lies is contested, and Module 6 takes up the argument properly. Note that a related technique, internal reconstruction, works on a single language by treating irregular alternations as fossils of an earlier regular pattern: English sing, sang, sung and foot, feet are visible residue of processes that were once general.

In short: Reconstruction gives a minimum, idealised, exception-tolerating picture of an ancestor, reliable within a limited time depth and silent about everything the daughters all discarded.

Common misconceptions

• Cognates are words that look alike. Cognates are words linked by regular correspondence. English much and Spanish mucho look alike and are unrelated; English cow and Sanskrit gaus look unalike and are cognate.
• The proto-form is the most common form among the daughters. Majority counting is a heuristic, not a principle. Directionality and system coherence decide, which is how Hawaiian's chain shift is recovered despite Hawaiian being outvoted three to one.
• Reconstructed languages are guesses. They are inferences from systematic evidence, tested by the fact that the same procedure reconstructs Latin correctly. They are also incomplete, which is a different criticism and a fair one.
• A proto-language is primitive or simple. Proto-languages are ordinary languages spoken by ordinary people. Where reconstructions look simple, it is usually because the evidence has thinned, not because the language was.
• If two languages share a word, they are related. Shared words are evidence of contact at minimum. Japanese has thousands of words from Chinese and belongs to neither Chinese family nor any relationship with it that has been demonstrated.

Putting it together

• The comparative method proceeds from candidate cognates in basic vocabulary to correspondence sets that hold across the whole lexicon.
• Systematic resemblance has four causes; a descent claim requires eliminating chance, borrowing, and universal tendencies first.
• Distinct correspondence sets require distinct proto-sounds, even when some daughter has merged them.
• Reconstruct by directionality and economy, not majority vote; the Hawaiian chain shift of k to glottal stop and t to k is recovered exactly this way.
• Tested on Romance, where Latin is attested, the method reconstructs the ancestor closely and even recovers conditioned changes such as palatalisation before front vowels.
• The method gives a floor on the ancestor's complexity, cannot see what all daughters lost, flattens variation, and thins out at great time depth.

Sources

1. Wikipedia contributors. (n.d.). Comparative method (linguistics). en.wikipedia.org
2. Britannica. (n.d.). Romance languages. Encyclopaedia Britannica. britannica.com
3. Hammarstrom, H., Forkel, R., Haspelmath, M., & Bank, S. (n.d.). Glottolog. Max Planck Institute for Evolutionary Anthropology. glottolog.org
4. Ringe, D. (1992). On calculating the factor of chance in language comparison. Transactions of the American Philosophical Society, 82(1), 1-110.
5. Campbell, L. (2020). Historical linguistics: An introduction (4th ed.). MIT Press.

Two words that should have matched and did not

Latin has pater and frater. Sanskrit has pitar and bhratar. Greek has pater and phrater. Gothic, the oldest substantially attested Germanic language, written down in the fourth century, has fadar and brothar. Look at the middle consonant in the Gothic pair. Brothar has the th sound of English thin, exactly as expected. Fadar has a d. By the rule that was supposed to govern all of this, both should have had th. One word obeyed and one did not, and in the 1870s that single stubborn d was the most important unsolved problem in linguistics.

This lesson is about the rule those words were supposed to obey, and about what happened when it failed. The failure turned out to be more instructive than the rule.

Key idea: Grimm's Law describes a systematic shift of consonants in Germanic, and the handful of words that disobeyed it drove the discovery that sound change is regular without exception once the conditioning environment is correctly identified.

Grimm's Law, in three series

Rasmus Rask noticed the pattern in 1818 and Jacob Grimm set it out fully in the second edition of his Deutsche Grammatik in 1822. It is a rotation, and like the Great Vowel Shift it comes in a set of coordinated steps rather than one change. Reconstructed forms carry an asterisk, so *p means a Proto-Indo-European sound inferred, not attested.

Series one: voiceless stops became voiceless fricatives. PIE *p, *t, *k became Germanic f, th, h.

PIE	Latin or Greek or Sanskrit	English	Sound
*p	Latin pes, pedis (foot)	foot	p becomes f
*p	Latin piscis (fish)	fish	p becomes f
*t	Latin tres (three)	three	t becomes th
*k	Latin centum (hundred)	hundred	k becomes h
*k	Latin cornu (horn)	horn	k becomes h
*k	Latin canis (dog)	hound	k becomes h

Series two: voiced stops became voiceless stops. PIE *b, *d, *g became Germanic p, t, k.

PIE	Latin or Greek or Sanskrit	English	Sound
*d	Latin decem (ten)	ten	d becomes t
*d	Latin duo (two)	two	d becomes t
*d	Latin dens, dentis (tooth)	tooth	d becomes t
*g	Latin genus (birth, kind)	kin	g becomes k
*g	Latin granum (grain)	corn	g becomes k
*b	Lithuanian dubus (deep)	deep	b becomes p

The last row is thin, and honestly so. Proto-Indo-European *b was extremely rare, possibly absent, so there are few examples to give. That gap is itself a datum, and Lesson 7 returns to what it might mean for the reconstructed sound system.

Series three: voiced aspirated stops became voiced stops. PIE *bh, *dh, *gh became Germanic b, d, g. Latin treated the same sounds differently, turning them into f or h at the start of a word, which is why the Latin cognates look so unlike the English ones.

PIE	Sanskrit	Latin	English
*bh (brother)	bhratar	frater	brother
*bh (carry)	bharami	fero	bear
*dh (door)	dvar	fores	door
*gh (stranger, guest)		hostis	guest

Sanskrit is doing important work in that last table. It kept the voiced aspirates as a distinct series, which is a large part of why its discovery by European scholars in the late eighteenth century reorganised the field: it showed that a three-way distinction the western languages had collapsed was original.

Why this matters: Grimm's Law is not a list of curiosities. It is a coordinated rotation of an entire consonant system, and it applies to the whole inherited vocabulary rather than to selected words.

The payoff you can use today

Here is the practical consequence that makes Grimm's Law worth memorising. English has two vocabularies stacked on top of each other. One is inherited directly from Proto-Germanic and has been through Grimm's Law. The other was borrowed from Latin, French, and Greek long after Grimm's Law had finished operating, so it was never subject to it. The two layers are full of pairs from the same PIE root, and the systematic difference between them is exactly the shift.

Inherited English (shifted)	Borrowed from Latin or Greek (unshifted)	The correspondence
father	paternal	f to p
foot	pedal, pedestrian	f to p
fish	piscine	f to p
three	triple, trinity	th to t
tooth	dental	t to d
ten	decimal, decade	t to d
heart	cardiac	h to c
horn	cornucopia, unicorn	h to c
kin	genus, generic	k to g

Once you see it you cannot unsee it. When an English word beginning with f has a learned relative beginning with p, or a word with h has one with c, you are looking at the same PIE root that took two routes into modern English, one through Germanic mouths and one through a scholar's Latin dictionary.

The point: The systematic mismatch between plain English words and their formal Latinate synonyms is Grimm's Law made visible in the modern vocabulary.

Debugging the exceptions

Now back to fadar. By the 1870s a group of young scholars at Leipzig, later called the Neogrammarians, had staked out a strong position: sound laws operate without exception, applying mechanically to every word that meets their conditions. Hermann Osthoff and Karl Brugmann put it in print in 1878. It was a bold empirical bet, and the scattered violations of Grimm's Law were the standing objection to it.

The violations had a shape. In a set of words, exactly where Grimm's Law predicted the voiceless fricatives f, th, h, or the inherited s, Germanic showed voiced sounds instead: b, d, g, z. Compare the Gothic pair we started with. Brothar has th; fadar has d. Both should have th.

In 1877 Karl Verner, a Danish scholar, published the solution in a paper whose title translates as An exception to the first sound shift. He looked past Germanic to Sanskrit, which preserves the position of the inherited Proto-Indo-European pitch accent. Sanskrit bhratar is accented on the first syllable. Sanskrit pitar is accented on the second. Line the accent up against the Germanic outcome and the exception evaporates.

Word	Sanskrit accent	Preceding syllable accented?	Gothic	Result
brother	on the root, bhratar	yes	brothar	th, as Grimm predicts
father	on the ending, pitar	no	fadar	d, voiced by Verner's Law

Verner's Law states it generally: the Germanic voiceless fricatives produced by Grimm's Law, and inherited *s, became voiced when the immediately preceding syllable did not carry the Proto-Indo-European accent, unless the sound was word-initial. Germanic then shifted its accent to the first syllable of the word, which destroyed the evidence for the conditioning environment inside Germanic itself. That is why the law was invisible for fifty years: you cannot see it without looking at Sanskrit or Greek, which kept the old accent placement.

The upshot: The exceptions to Grimm's Law were not exceptions. They were a second regular law whose conditioning factor, accent position, had been erased by a later change.

The fossils in your own speech

Verner's Law left visible residue. Because the accent in Proto-Indo-European moved around within a verb's paradigm, some forms of a verb met the condition and others did not, producing a consonant alternation within a single word family. Germanists call this grammatical alternation, and English still carries fragments of it.

The clearest survivor is was and were. Old English had wesan, to be, with past forms waes and waeron. The s in the singular met the accent condition one way and the plural the other, giving s against z; the z then became r in West Germanic and Norse by a further regular change called rhotacism. That is the entire explanation for why the past tense of English be changes its consonant between singular and plural. It is a two-stage fossil of a Proto-Indo-European accent that no Germanic speaker has produced in three thousand years.

Two more survive as separate words: seethe and sodden, which was once its past participle, and lose beside forlorn. Where the modern language has an unexplained consonant alternation between related words, Verner's Law is usually worth checking first.

Worth holding on to: The irregular verb was against were is not an irregularity. It is the last visible trace of two regular sound laws stacked on top of each other.

What the episode settled, and what it did not

The Verner result changed the field's self-understanding. If the most famous counterexample to the regularity claim dissolved on close inspection into a second regular law, then apparent exceptions were a reason to look harder, not a reason to abandon the principle. That conviction is why the comparative method works at all. Without regularity, correspondence sets would not exist and there would be nothing to reconstruct from.

Regularity is a working principle, though, not a metaphysical law, and three genuine sources of irregularity remain. Borrowing introduces words that never underwent the change: English father was inherited and paternal was not, so the pair does not violate Grimm's Law, it sits outside it. Analogy remakes forms on the model of other forms, and Lesson 4 is devoted to it. And some very high frequency words reduce irregularly, which is why English going to becomes gonna while flowing to does not become flonna.

There is also a live technical dispute worth knowing about. The Neogrammarian picture has sound change applying to all eligible words at once, gradually in phonetic detail. An alternative, associated with William Wang and the term lexical diffusion, has change spreading word by word through the vocabulary, abruptly for each word. William Labov examined both in 1981 and concluded that both happen: some changes look Neogrammarian, others look diffusionist, and the difference correlates with the kind of change involved. The regularity principle survives as the default expectation rather than as a claim that nothing else ever occurs.

In short: Regularity earned its status by surviving its hardest test, and it remains the default assumption, with borrowing, analogy, and lexical diffusion as recognised and bounded departures.

Common misconceptions

• Grimm's Law means English words come from Latin. The opposite. It describes how English and Latin words both descend from a common ancestor along different routes, which is why they differ so systematically.
• Grimm's Law is still operating. It ran to completion in Proto-Germanic, before the earliest attested Germanic texts. Words borrowed afterward, such as pedal and cardiac, were never eligible.
• Verner's Law was a patch invented to save Grimm. It made a testable prediction about accent position in languages Verner was not studying, and Sanskrit and Greek confirmed it. That is a discovery, not a repair.
• Sound laws admit no exceptions in an absolute sense. The claim is that a change applies to every word meeting its phonetic conditions. Borrowed words, analogically reshaped words, and some very frequent items sit outside those conditions.
• The shifts happened because Germanic speakers were separated from Latin speakers. Neither Latin nor Germanic is the ancestor of the other. Both are sisters, and each shifted away from the shared ancestor in its own way, Latin also having changed *bh to f.

What you now know

• Grimm's Law rotated the Proto-Indo-European stops in Germanic in three coordinated series: voiceless stops to fricatives, voiced stops to voiceless, voiced aspirates to voiced stops.
• Because Latin and Greek loanwords entered English after the shift, English carries paired vocabularies whose differences, father and paternal, ten and decimal, are the law itself.
• A residue of words disobeyed the law, showing voiced sounds where voiceless ones were predicted.
• Verner showed in 1877 that the voicing depended on the position of the Proto-Indo-European accent, preserved in Sanskrit and Greek but erased in Germanic by a later accent shift.
• English was against were is a surviving fossil of Verner's Law plus rhotacism.
• The episode established regularity as the field's default expectation, with borrowing, analogy, and lexical diffusion as recognised departures.

Sources

1. Wikipedia contributors. (n.d.). Grimm's law. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Verner's law. en.wikipedia.org
3. Britannica. (n.d.). Indo-European languages. Encyclopaedia Britannica. britannica.com
4. Verner, K. (1877). Eine Ausnahme der ersten Lautverschiebung. Zeitschrift fuer vergleichende Sprachforschung, 23(2), 97-130.
5. Ringe, D. (2017). From Proto-Indo-European to Proto-Germanic (2nd ed.). Oxford University Press.
6. Labov, W. (1981). Resolving the Neogrammarian controversy. Language, 57(2), 267-308.

The plural of book used to be bec

In Old English the plural of boc, book, was bec. It was formed exactly the way fot gave fet and toth gave teth: an ancient vowel change triggered by a plural ending that later dropped off. The pattern was regular. Today we say books, and nobody has said bec in eight hundred years. Meanwhile feet and teeth are still standing. Cu, cow, had the plural cy, which survived long enough to give the poetic word kine before it too gave way to cows. Nothing happened to any of these sounds. No articulatory pressure removed the vowel change from book and left it in foot. Something else did the work.

That something is analogy: the remaking of a form on the model of other forms, so that an item that stood outside a pattern is pulled into it. Analogy is the second great engine of language change, and it is the one that operates on the grammar rather than on the sounds.

Key idea: Analogy changes forms without any sound change, by reshaping irregular items to match a productive pattern that already exists elsewhere in the language.

The four-part proportion

Analogical changes can usually be written as a proportion, the way you would set up a ratio in arithmetic. Take a child who says brang.

sing : sang :: bring : X, and the child solves for X as brang.

The child has not made a random error. She has extracted a genuine pattern from the input, applied it to a new item, and produced exactly what the pattern predicts. The adult form brought is the anomaly here, and she has not yet learned to exempt it. The same proportion produces goed, foots, mouses, and hitted, and every one of them shows a learner whose grammar is working correctly on incomplete information.

The reason this matters historically is that sometimes the innovation sticks. Consider dive. Its past tense was dived for the entire recorded history of English. Then, in nineteenth century American English, dove appears, built on the proportion drive : drove :: dive : X. Dove was an error that won. Sneak beside snuck is a second case, also American and also nineteenth century, and it is still in competition with sneaked today. You are living inside an analogical change in progress, and you probably have an opinion about it, which is how these things go.

The point: A child's error and a completed historical change are the same operation. The difference is only whether the community adopts the result.

Two directions: levelling and extension

Analogy comes in two flavours, and the distinction is worth keeping straight.

Levelling removes an alternation inside a paradigm, making a word's forms more alike. Old English helpan was a strong verb with the past forms healp and hulpon: help, holp, holpen. Today it is help, helped, helped. The vowel alternation was levelled out. English has lost hundreds of strong verbs this way. Old English had roughly three hundred; modern English has fewer than two hundred, most of them uncommon, and the class has taken no new members from borrowing in centuries.

Extension spreads a pattern to items that did not have it, making a word's forms less like their own history and more like the neighbours. Dove and snuck are extensions. So, in the other direction, is the American use of dived being replaced. Extension is rarer than levelling, and it happens most often when a small irregular class is phonetically coherent enough to look like a rule: the verbs that take the vowel of drove, wrote, rode, and spoke share enough shape to invite recruitment.

Why this matters: Levelling makes paradigms internally consistent; extension recruits new members into an existing irregular class. Both increase regularity in the sense of pattern conformity, which is why they push in the same overall direction.

Why feet survived and bec did not

If analogy regularises, why do any irregulars survive? The answer is frequency, and it has been measured.

An irregular form has to be stored and retrieved as a whole item, because no rule generates it. Storage is maintained by use. A word you hear thousands of times a year, such as feet or went or children, is reinforced constantly and stays available. A word you hear rarely, such as the plural of a book in an eleventh century monastery, is not, and when the speaker cannot retrieve the stored form the productive rule fills the gap. Say it once and you have created a new form; say it enough and the community has a new word.

Erez Lieberman and colleagues quantified this in 2007 by tracking English irregular verbs across roughly twelve hundred years of texts. They found that the rate at which an irregular verb regularises varies inversely with the square root of its frequency of use: a verb used one hundred times less often regularises about ten times faster. The verbs that have survived as irregulars are, overwhelmingly, the most common verbs in the language. Be, have, go, do, say, make, take, come, see, get. That is not a coincidence and it is not a mystery.

Remember: Irregularity is maintained by frequency. The most common words in a language are its most irregular ones because they are the only ones used often enough to keep an unpredictable form alive.

Reanalysis: the listener draws the line somewhere new

The other force in this lesson is reanalysis, and it works differently. Analogy changes a form. Reanalysis changes the structure assigned to a form that has not changed at all.

Start with the simplest kind, at a word boundary. Middle English had a napron, borrowed from French. Say it aloud and there is nothing in the sound to tell a listener whether the n belongs to the article or to the noun. Enough listeners assigned it to the article, and English acquired an apron. The same thing happened to a nadder, which became an adder, and to an ewt, which became a newt, with the n travelling the other way. The word orange came into English through a chain from Arabic naranj, and it lost its initial n somewhere along the route for the same reason. In each case the sequence of sounds stayed put and only the boundary moved.

The same happens inside words. Hamburger is named after Hamburg, so its parts are Hamburg and the suffix er. English speakers reanalysed it as ham plus burger, and burger became a productive element that has since attached to cheese, veggie, and everything else. Helicopter is Greek helico, spiral, plus pter, wing; it was rebracketed as heli plus copter. Alcoholic yielded aholic. The Watergate scandal of 1972 gave English a suffix gate that is still attaching to new scandals fifty years later. And the noun pea has an odd history in the same family: the older word was pease, a mass noun, which sounded like a plural and was reanalysed as one, generating a new singular that had never existed.

Worth holding on to: Reanalysis needs no phonetic change at all. It happens when a listener assigns a different structure to the same string, and the new structure then generates new forms.

Folk etymology and structural reanalysis

Folk etymology is reanalysis plus repair: an opaque word is reshaped to look like something familiar. Old English brydguma was bride plus guma, an old word for man. When guma died out, the word was reshaped using groom, which had nothing to do with it, giving bridegroom. Old French crevice became crayfish, though it is not a fish. Old English angnaegl, a painful nail, became hangnail. Asparagus has been sparrowgrass in English dialects for centuries. In each case a speaker met an unanalysable word and analysed it anyway.

The most consequential kind of reanalysis is syntactic. Consider the English sentence I am going to London, which in the sixteenth century meant precisely what it says, a journey. In a sentence like I am going to marry her, the same string can be parsed two ways: as going, with a purpose, in order to marry, or as a single future marker going to, followed by marry. Once enough listeners took the second parse, the construction was free of any movement at all, which is why I am going to stay right here is not a contradiction, and why going to reduced to gonna while the motion verb in I am going to the store cannot. The sounds did not change first; the structure did, and the sound reduction followed. Lesson 6 develops this into a general account.

So what?: Reanalysis is the main route by which the syntax of a language changes, because it lets a new structure enter without any new form being introduced.

Sturtevant's paradox

Put the last two lessons together and you get a useful formulation, credited to Edgar Sturtevant. Sound change is regular and produces irregularity. Analogy is irregular and produces regularity.

The first half is easy to see. A regular sound change applies to every eligible word, but words in a paradigm do not all present the same environment, so the change splits the paradigm. The vowel alternation in foot and feet was created by a perfectly regular process operating on a plural ending that no longer exists. Regular process, irregular result.

The second half is the mirror image. Analogy cannot be stated as a law, because you cannot predict which of the eligible irregulars will be levelled or when. It hit book and cow and left foot and tooth. Unpredictable process, regularising result. The two forces run continuously and in opposite directions, which is why every language is a mixture of tidy patterns and stubborn exceptions, and why the exceptions cluster in the vocabulary you use most.

The core of it: Sound change creates the irregularities that analogy then cleans up, and neither ever finishes, which is why no language is either fully regular or fully chaotic.

Common misconceptions

• Children's errors are just mistakes. Forms like goed and brang are the correct output of a rule the child has genuinely extracted. Historically, some such forms have won.
• Irregular verbs are relics that will all disappear. The high-frequency ones are actively maintained by use. Be and go have been irregular for the whole recorded history of English and show no sign of yielding.
• Folk etymology means an incorrect story about a word's origin. In linguistics it names an actual change to the word's form, as in bridegroom and crayfish. The false-story sense is a different, popular use of the phrase.
• Analogy is a kind of sound change. They are different mechanisms. Analogy operates on grammatical patterns and can move a form in a direction no sound change would produce.
• Reanalysis requires a mispronunciation. The string can be phonetically identical. What changes is the structure the listener assigns to it, which is why an apron and a napron sound the same.

Pulling it together

• Analogy remakes forms on the model of existing patterns and can be written as a four-part proportion, the same operation behind a child's brang and the historical rise of dove.
• Levelling removes alternations within a paradigm; extension recruits new members into an existing irregular class.
• Irregular forms survive in proportion to their frequency, a relation measured across twelve hundred years of English texts.
• Reanalysis assigns new structure to an unchanged string, moving word boundaries (an apron), morpheme boundaries (burger, gate), and syntactic structure (going to).
• Folk etymology reshapes opaque words into familiar-looking ones: bridegroom, crayfish, hangnail.
• Sturtevant's paradox: regular sound change generates irregularity, and irregular analogy generates regularity, and both run without stopping.

Sources

1. Britannica. (n.d.). Inflection. Encyclopaedia Britannica. britannica.com
2. Britannica. (n.d.). English language. Encyclopaedia Britannica. britannica.com
3. Wikipedia contributors. (n.d.). Folk etymology. en.wikipedia.org
4. Lieberman, E., Michel, J.-B., Jackson, J., Tang, T., & Nowak, M. A. (2007). Quantifying the evolutionary dynamics of language. Nature, 449(7163), 713-716.
5. Hock, H. H., & Joseph, B. D. (2019). Language history, language change, and language relationship (3rd ed.). De Gruyter Mouton.

Three sentences from 1611 that no longer say what they said

The King James Bible tells its readers at Philippians 4:6 to be careful for nothing. It is not counselling recklessness. Careful in 1611 meant full of care in the sense of anxiety, so the instruction is do not be anxious about anything. At 1 Thessalonians 4:15 the living shall not prevent them which are asleep, which does not mean obstruct the dead; prevent came from Latin praevenire, to come before, and the verse means the living will not precede them. And at Mark 10:14, suffer the little children to come unto me is not about pain. Suffer meant allow.

Three words in one book, all still in the language, all still spelled the same, all now carrying meanings their 1611 users would not recognise. Nothing about the sounds has changed. This lesson is about the third great engine of language change, the one that moves meanings while leaving forms alone, and about why it is both the most obvious kind of change and the hardest to use as evidence.

Key idea: Meanings change independently of forms, which is why a text can become misleading rather than unreadable, and why the trap is worse: you feel you have understood it.

The standard taxonomy, with real cases

Semantic changes have been sorted into recurring types since the nineteenth century. The categories are descriptive labels rather than explanations, but they are genuinely useful, because they tell you what to look for.

Type	Definition	Attested English case
Narrowing	The word comes to name a subset of what it named before	deer, from Old English deor, any animal
Narrowing		meat, from Old English mete, food of any kind
Broadening	The word comes to name more than it named before	barn, from bere-aern, a building for barley
Broadening		holiday, from haligdaeg, a holy day
Amelioration	The word rises in evaluation	knight, from cniht, boy or servant
Pejoration	The word falls in evaluation	silly, from Old English saelig, blessed and happy
Metaphor	Transfer by resemblance	grasp an idea; the mouse on a desk; a computer virus
Metonymy	Transfer by association or contiguity	the crown for the monarchy; the White House said
Bleaching	Loss of specific content, often leaving an intensifier	very, from Old French verai, true

Notice how many of these you can still see happening. The computer mouse was named by metaphor within living memory and its resemblance to the animal is already fading from awareness; ask a ten year old to draw the shape it was named for.

The point: The types are not exotic. Each one is a process you can watch operating in vocabulary that entered English in your own lifetime.

Narrowing and broadening, worked

Old English deor named animals in general, the way its German cousin Tier still does. Modern English deer names one family of ruminants. That is a drastic narrowing, and it happened because a competitor arrived: after the Norman conquest, animal came in from French and took the general sense, leaving deor to specialise. The same competitive pressure narrowed hund, which meant dog, into hound, a hunting dog, when dog itself spread; and mete, food, into meat, flesh, a specialisation that has left fossils behind. Sweetmeats are not made of flesh. When something is meat and drink to you, it is food and drink.

Starve is the sharpest case. Old English steorfan simply meant to die, as German sterben still does. English narrowed it first to dying of hunger and, in some regions, dying of cold. The general sense went to die, a Norse borrowing.

Broadening runs the other way. A barn was a barley house, ancestor of the general farm building. A holiday was a holy day, now any day off. Arrive comes through French from a Latin phrase meaning to come to the shore, and it kept the nautical restriction for a while before broadening to any destination at all, which is why you can now arrive by aeroplane without contradiction. Place goes back to Greek and Latin words for a broad street, and has broadened to cover nearly any location.

Why this matters: Narrowing and broadening are often driven by competition. When a new word arrives with a general meaning, the incumbent tends to specialise, and when a competitor dies, the survivor expands.

The words that went up, and the words that went down

Evaluative change is where the social history is visible. Old English cniht meant a boy or a household servant, and the word climbed steadily until it named a member of the mounted aristocracy. Nice arrived from French around 1300 meaning foolish or simple, from Latin nescius, ignorant, and made a long journey through fastidious and precise before settling into agreeable, which is why nice distinction still means a fine one. Pretty descends from Old English praettig, cunning or crafty.

The downward cases usually track social contempt. Villain was a French word for a farm worker attached to a villa, an estate. Its slide into scoundrel is a fossil of what the writing classes thought of agricultural labourers. Vulgar is Latin vulgaris, belonging to the common people, and it took the same route. Silly is the most striking reversal in English: Old English saelig meant blessed and lucky, then innocent, then harmless, then feeble, then foolish.

Two related processes deserve names. Words for the frightening tend to weaken through overuse: awful once meant inspiring awe, terrific meant inspiring terror, and both have been drained by service as intensifiers, one downward and one upward. And words for the taboo run on what Steven Pinker named the euphemism treadmill: a polite substitute absorbs the connotations of the thing it names, becomes impolite in turn, and is replaced. English has cycled through a long line of terms for the room with the toilet in it, each one a genteel evasion that eventually needed evading. The pattern tells you that the negative charge sits on the referent, not on the word, which is why replacing the word buys only a decade or two.

Remember: Evaluative change is not random drift. It follows the social standing of the thing referred to, which is why the euphemism treadmill keeps turning.

Metaphor, metonymy, and bleaching

Two mechanisms account for a large share of all semantic change, and they work on different principles.

Metaphor moves a word across a resemblance, usually from the concrete to the abstract. You grasp an idea, follow an argument, see what someone means, and are moved by a story. Every one of those is a physical verb doing intellectual work. The direction is overwhelmingly one way: concrete words come to have abstract senses far more often than abstract words acquire concrete ones. That asymmetry is one of the few near-regularities semantics offers.

Metonymy moves a word across an association. The crown stands for the monarchy because monarchs wear one. The White House said stands for an administration housed in a building. A dish can be the food on it, and Hollywood can be an industry that is no longer concentrated in that district. Metonymy is why so many institutional names are addresses.

Bleaching is what happens when a word is used so often for emphasis that it stops meaning anything specific. Very began as the French and Latin word for true. Really and truly are heading the same way; so are awfully and terribly. The most contested current case is literally, whose use as an intensifier is treated as a modern collapse of standards but is attested in respected nineteenth century prose. It is doing exactly what very did, and the outrage it provokes is itself a familiar historical pattern.

The upshot: Metaphor transfers by likeness, metonymy by association, and bleaching hollows a word out through repeated emphatic use. Together they cover most of what happens to meanings.

Why there is no law of semantic change

Now the methodological point that this whole lesson exists to deliver, and it matters for everything in Module 3.

Sound change is regular. Given the environment, you can predict the outcome, which is what makes correspondence sets and reconstruction possible. Semantic change is not regular in that sense. There is no rule saying that words for animals narrow, or that words for servants ameliorate. Deer narrowed and bird broadened. Knight rose and villain fell. Each history is intelligible after the fact and none was predictable before it.

Some tendencies do hold. Change runs more often from concrete to abstract than the reverse. Elizabeth Traugott and Richard Dasher have argued at length that meanings tend to become more speaker-oriented over time, moving from describing the world to expressing the speaker's stance, a process called subjectification. These are statistical tendencies, useful for judging plausibility, and they are not laws.

The consequence for reconstruction is concrete and constraining. When you reconstruct a proto-form, you reconstruct the sound shape with reasonable confidence and the meaning with much less. A root whose descendants mean fear in one branch and flee in another and tremble in a third supports a reconstructed meaning somewhere in that region, stated as a range rather than a gloss. Any argument that leans hard on the precise meaning of a proto-word, and there are many such arguments in the debate over the Indo-European homeland, is leaning on the weakest joint in the method. Keep that in mind for Lesson 8.

Bottom line: Reconstructed sounds are far more secure than reconstructed meanings, and any inference about a proto-culture inherits the weakness of the semantics it depends on.

Common misconceptions

• A word's true meaning is its oldest meaning. This is the etymological fallacy. Decimate once meant to kill one in ten, and nice once meant ignorant. Current meaning is determined by current use, not by ancestry.
• Semantic change is a modern decline. The King James Bible was already using careful, prevent, and suffer in senses that had changed from earlier English and have changed again since.
• Intensifier literally is a new abuse. It is bleaching, the same process that produced very, really, and truly, and it is attested well back into the nineteenth century.
• Euphemisms make an unpleasant topic more polite permanently. The connotation attaches to the referent, so replacement terms acquire the old charge and have to be replaced in turn.
• Meanings can be reconstructed as precisely as sounds. They cannot. Semantic reconstruction yields a plausible range, which is why cultural inferences from reconstructed vocabulary need to be handled carefully.

The short version

• Meanings change while forms stay put, which makes old texts misleading rather than obviously foreign: careful, prevent, and suffer in 1611 all meant something else.
• The standard types are narrowing, broadening, amelioration, pejoration, metaphor, metonymy, and bleaching, each attested in ordinary English vocabulary.
• Narrowing and broadening are frequently driven by competition with a newly borrowed word, as with deer beside animal and hound beside dog.
• Evaluative change tracks the social standing of the referent, which is what keeps the euphemism treadmill turning.
• Metaphor moves meaning by resemblance and runs mostly from concrete to abstract; metonymy moves it by association.
• Semantic change has tendencies but no laws, so reconstructed meanings are ranges rather than glosses, and cultural arguments built on them are correspondingly weaker than phonological ones.

Sources

1. Wikipedia contributors. (n.d.). Semantic change. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Euphemism. en.wikipedia.org
3. Linguistic Society of America. (n.d.). Is English changing? linguisticsociety.org
4. Traugott, E. C., & Dasher, R. B. (2002). Regularity in semantic change. Cambridge University Press.
5. Oxford English Dictionary. (n.d.). Entries for careful, prevent, suffer, nice, silly, deer, and starve. Oxford University Press.

A verb form that should not exist

The seventh century Latin chronicle attributed to Fredegar contains the word daras, meaning you will give. Classical Latin does not have that word. Classical Latin says dabis. Daras is what happens when the two-word phrase dare habes, literally to give you have, is spoken fast enough for long enough that it stops being two words. It is one of the earliest snapshots historians of Romance have of a new future tense being born out of a sentence about obligation.

That single form contains an entire process. Latin's inherited future disappeared, and every Romance language replaced it with a construction built from the verb to have. French je chanterai, Spanish cantare, Italian cantero, Portuguese cantarei: in each of them the ending is a squashed present tense of habere fused onto the infinitive. This lesson is about the process that turns a phrase like that into an ending, which is called grammaticalisation, and it is the main way that languages acquire new grammar.

Key idea: New grammatical machinery is not invented. It is recruited from ordinary content words and phrases that gradually lose their independence, their meaning, and their phonetic bulk.

Following the Romance future all the way down

Take the stages in order, because each one is attested.

Stage one, a full lexical phrase. Latin cantare habeo means I have something to sing, or I have a song to sing. Habeo is a full verb with its own meaning, possession, and cantare is its complement. The phrase is compositional: it means what its parts mean.

Stage two, obligation. Having something to do slides easily into having to do it. The same slide happened independently in English, where I have to leave is obligation rather than possession, and in dozens of unrelated languages. This is the invited inference doing the work: if I have a letter to write, I probably must write it, and hearers who draw that inference often enough make it part of the meaning.

Stage three, futurity. Obligation implies futurity, since anything you must do is not yet done. Once the future reading is available, the obligation content becomes optional, then absent. Now the construction means only that the event lies ahead.

Stage four, fusion. Word order fixes: habeo can no longer wander around the clause. It reduces phonetically: habeo to ao to a. And it attaches: cantare habeo becomes cantarai becomes chanterai. What was a verb is now a suffix, and no French speaker can hear it.

Here is the receipt that the fusion is historical rather than an accident of similar shapes. In older Spanish and in Portuguese still, an object pronoun could be placed between the infinitive and the ending, which is only possible if the ending was once a separate word. Portuguese dar-lhe-ei, I will give to him, wraps the pronoun lhe inside the future form. You cannot insert something into the middle of a suffix. You can insert it between two words that are on their way to becoming one.

Worth holding on to: The Portuguese construction with a pronoun inside the future form is direct morphological evidence that the Romance future ending began life as a separate verb.

The parameters, and how to recognise the process

Grammaticalisation is recognised by a bundle of properties that tend to occur together. They are worth learning as a checklist.

Parameter	What it means	In the Romance future
Semantic bleaching	Specific content is lost	Possession, then obligation, then nothing but futurity
Decategorialisation	The item loses the properties of its word class	Habeo stops behaving like a verb; it takes no subject of its own
Phonetic reduction	The form shortens	habeo to ai in French
Fixation	Free position becomes fixed position	The ending can appear only after the stem
Obligatoriness	An optional choice becomes a required category	A French verb must be marked for tense
Coalescence	Separate words fuse into one	Two words become one inflected form

Paul Hopper and Elizabeth Traugott summarise the trajectory as a cline: content word, then grammatical word, then clitic, then affix, then nothing at all. English illustrates every stage at once, because it has items sitting at each point. The noun body gave the suffix in slowly and quickly through Old English lic, an item meaning form or body. Old English had, condition, gave hood in childhood. Dom, judgement, gave dom in freedom and kingdom. The full noun while, a period of time, is still a noun in a long while and a subordinating conjunction in while you were out.

English modals tell the same story. Will was willan, to want. Shall was sculan, to owe. Can was cunnan, to know how. May, must, and might all descend from full verbs with concrete meanings, and all of them have lost the ability to take the endings that ordinary verbs take, which is decategorialisation you can test: nobody says he cans.

The point: Grammaticalisation is diagnosable. Look for bleaching, loss of word-class behaviour, phonetic reduction, fixed position, and a new obligatory category, arriving together.

A second run, where the cycle restarts

The Romance future runs in one direction and stops. The second worked case is more interesting, because the same process feeds back on itself. This is Jespersen's cycle, named after the Danish linguist Otto Jespersen, and French is the classic demonstration.

Stage	Form	What is happening
Latin	non dico	A single strong negator before the verb
Old French	je ne dis	The negator has worn down to ne and is now weak
Middle French	je ne dis pas	A noun, pas, meaning step, is added for emphasis
Standard modern French	je ne dis pas	Both parts are obligatory; pas has lost the meaning step
Colloquial modern French	je dis pas	Ne drops in speech; pas alone now carries the negation

Watch what pas did. It began as an ordinary noun. It entered the negation as a minimiser, the way English says not a step, not a bit, not a drop. Other minimisers competed and lost: French once used point, a dot, and mie, a crumb, in the same slot, and point survives only as a literary variant. Pas won, bleached until it no longer meant step, became obligatory, and has now taken over the whole job from the original negator, which is disappearing. The system has returned to stage one with a different word in the slot, and it is available to start again.

English ran the same cycle. Old English ic ne secge became Middle English I ne seye not, where not descends from na wiht, no thing or no creature, another minimiser. Then ne dropped, leaving I say not. Then English did something Romance did not, and recruited the verb do to carry the tense, giving I do not say and its reduced form don't. The cycle is the same; the outcome differs because English had do available and French did not.

So what?: Grammaticalisation is not a one-time event in a language's history. Categories wear out and are rebuilt from fresh lexical material, sometimes repeatedly, and languages take different routes through the same cycle depending on what material is at hand.

Unidirectionality, and the honest objections

The strongest claim in this literature is unidirectionality: grammaticalisation runs from lexical toward grammatical and not the other way. Content words become function words, function words become affixes, affixes erode to nothing. Nouns do not spontaneously grow out of case endings.

The claim has a great deal of support, and Bernd Heine and Tania Kuteva's survey of grammaticalisation paths in hundreds of languages found the same routes recurring across unrelated families: verbs meaning go and come becoming future markers, verbs meaning finish becoming perfect markers, nouns meaning back and head becoming spatial adpositions. The convergence is impressive and it is what makes the concept predictive rather than merely descriptive.

But the claim is not absolute, and it would be dishonest to teach it as if it were. Counterexamples, collected under the name degrammaticalisation, are real if uncommon. English used to have a genitive case suffix on the noun; the modern possessive is a clitic that attaches to the end of a whole phrase, as in the king of Spain's daughter, where it marks the phrase rather than the word king. That is movement in the wrong direction along the cline. The suffix ish has escaped into an independent word, as in the answer to a question about whether you are ready: ish. The preposition up has become a verb in to up the ante. The fair statement is that grammaticalisation is a strong and cross-linguistically robust tendency with a small set of documented exceptions, and that a proposal violating it needs more evidence than one following it, not that it is impossible.

In short: Unidirectionality is a well supported statistical generalisation rather than a law, and treating it as either an absolute or a mere hypothesis misstates the evidence.

Common misconceptions

• Grammar was designed or invented at some point. Every grammatical marker whose history can be traced turns out to come from ordinary vocabulary. Nobody designed the French future; it was a phrase about having things to do.
• Phonetic reduction causes grammaticalisation. The order usually runs the other way. Bleaching and structural reanalysis come first, and reduction follows because the item is now predictable and unstressed.
• The stages replace one another cleanly. They overlap for centuries. English still has full lexical have alongside obligation have to alongside perfect auxiliary have, all coexisting.
• Once a language loses a category it is impoverished. Latin lost its future and Romance built a new one within a few centuries. Categories are replaced far more often than they are simply lost.
• Unidirectionality means degrammaticalisation never happens. A small number of documented cases exist, including the English possessive clitic. The generalisation is strong, not exceptionless.

Where this leaves us

• Grammaticalisation recruits ordinary words into grammatical roles, and it is the principal source of new grammatical machinery.
• The Romance future runs from Latin cantare habeo, a phrase about obligation, to a fused inflectional ending, with Portuguese dar-lhe-ei preserving the seam.
• The process is diagnosable by bleaching, decategorialisation, phonetic reduction, fixation, obligatoriness, and coalescence occurring together.
• English suffixes such as ly, hood, and dom, and the modal verbs will, shall, and can, all began as independent content words.
• Jespersen's cycle shows negation being worn down, reinforced by a minimiser such as French pas, and rebuilt, so the process can repeat.
• Unidirectionality is strongly supported across unrelated families but admits documented exceptions such as the English phrasal possessive.

Sources

1. Wikipedia contributors. (n.d.). Grammaticalization. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Jespersen's cycle. en.wikipedia.org
3. Britannica. (n.d.). Romance languages. Encyclopaedia Britannica. britannica.com
4. Hopper, P. J., & Traugott, E. C. (2003). Grammaticalization (2nd ed.). Cambridge University Press.
5. Heine, B., & Kuteva, T. (2002). World lexicon of grammaticalization. Cambridge University Press.

Calcutta, 2 February 1786

On 2 February 1786, William Jones, a judge on the Supreme Court of Bengal, delivered the third anniversary discourse to the Asiatick Society he had helped found. He had been studying Sanskrit, and in that lecture he said that Sanskrit's resemblance to Greek and Latin in verb roots and grammatical forms was too strong to have arisen by accident, and that all three probably descended from a common source which no longer existed.

Jones was not the first European to notice. A Florentine merchant, Filippo Sassetti, had remarked on Sanskrit and Italian similarities in the 1580s, and a French Jesuit in India, Gaston Coeurdoux, had made a fuller case in a paper submitted in 1767 that sat unread for decades. What made Jones's statement the starting gun was its audience and its precision: he pointed at the verb morphology, not at scattered vocabulary, and he named a lost ancestor rather than deriving one language from another. Within forty years Rasmus Rask, Franz Bopp, and Jacob Grimm had turned that suggestion into a working comparative programme. Proto-Indo-European is the result, and it remains the most thoroughly reconstructed unattested language in the world.

Key idea: Indo-European is not a hypothesis about vocabulary resemblance. It rests on systematic correspondence in inflectional morphology, which is far harder to borrow and far harder to match by chance.

The branches

Ten branches are recognised. Two of them are extinct and are known only from texts recovered by archaeologists in the twentieth century, which is worth pausing on: the family's shape was worked out before those two branches were known, and they fitted.

Branch	Earliest substantial attestation	Modern representatives
Anatolian	Hittite, from around 1650 BCE	None; entirely extinct
Indo-Iranian	Sanskrit and Avestan, second millennium BCE	Hindi, Urdu, Bengali, Persian, Pashto, Kurdish
Greek	Mycenaean in Linear B, around 1400 to 1200 BCE	Greek
Italic	Latin and its neighbours, from the seventh century BCE	Spanish, Portuguese, French, Italian, Romanian
Celtic	Old Irish and Gaulish inscriptions	Irish, Scottish Gaelic, Welsh, Breton
Germanic	Gothic, fourth century CE	English, German, Dutch, the Scandinavian languages
Balto-Slavic	Old Church Slavonic, ninth century CE	Russian, Polish, Czech, Lithuanian, Latvian
Armenian	Fifth century CE	Armenian
Albanian	Fifteenth century CE	Albanian
Tocharian	Manuscripts from the Tarim Basin, sixth to eighth centuries CE	None; entirely extinct

The Tocharian case is worth a sentence on its own. Manuscripts recovered from western China around 1900 turned out to record two closely related Indo-European languages spoken thousands of kilometres east of any other branch. The family had predicted nothing about them, and yet their inflections slotted into the reconstruction without special pleading.

Why this matters: Two whole branches were discovered after the reconstruction was built, and both confirmed rather than broke it. That is the kind of test a reconstruction can fail and did not.

What the evidence actually looks like

Vocabulary first, because it is the easiest to read.

Meaning	Sanskrit	Greek	Latin	Gothic or English	Lithuanian
two	dva	duo	duo	twai, two	du
three	trayas	treis	tres	threis, three	trys
mother	matar	meter	mater	modor, mother	motina
is	asti	esti	est	ist, is	esti

Now the stronger evidence, which is grammatical rather than lexical. The verb to be is irregular in every one of these languages, and it is irregular in the same way, with a form beginning with a vowel plus s in the third person singular and a different stem elsewhere. Shared irregularity is the gold standard of comparative evidence, because regular patterns can be borrowed or reinvented and shared arbitrary quirks cannot. Add the case systems, which line up in category and often in ending, and the three-way gender system, and the athematic verb endings, and the case for a single ancestor becomes overwhelming.

The point: Shared irregularities in inflection carry far more evidential weight than shared words, because there is no plausible route to them other than common inheritance.

The reconstructed sound system, and its oddities

Proto-Indo-European is usually reconstructed with a large stop inventory arranged in three series and, unusually, three separate positions at the back of the mouth.

Series	Labial	Dental	Palatal	Plain velar	Labiovelar
Voiceless	*p	*t	palatal k	*k	*kw
Voiced	*b (very rare)	*d	palatal g	*g	*gw
Voiced aspirate	*bh	*dh	palatal gh	*gh	*gwh

Two features of this table have generated a century of argument. The first is the near-absence of *b. A system with p, t, k but no b, while having d and g, is typologically strange: languages that lack one voiced stop usually lack g, not b. That oddity is the main motivation for the glottalic theory, a minority proposal that reinterprets the voiced series as ejectives, which would make the gap unremarkable. It has never won majority support, but it is a serious argument rather than a crank one, and it is a good illustration of the uniformitarian principle from Lesson 1 being used as a constraint on reconstruction.

The second is the three dorsal series, which produced the family's most famous isogloss. The palatal series stayed a stop in one group of languages and became a sibilant in another. Latin centum, hundred, begins with k. The Avestan word for hundred is satem. Those two words gave their names to the centum and satem groups. Early scholars took the split as the family's primary division, west against east. It is not. Tocharian is centum and sits far to the east, and the modern view treats the satem development as an areal innovation that spread across neighbouring branches rather than a defining break in the family tree.

Proto-Indo-European also had ablaut, a systematic vowel alternation within roots: an e grade, an o grade, and a zero grade with no vowel at all. Greek shows it cleanly in leipo, I leave; leloipa, I have left; elipon, I left. English carries inherited residue of the same alternation in sing, sang, sung and in drive, drove, driven. Those are not English irregularities. They are five thousand year old morphology.

Remember: The strong verbs of English are not an English quirk. They are the last working remnant of the Proto-Indo-European ablaut system.

The prediction that came true fifty years later

This is the episode that makes the case for the method better than any argument could.

In 1879 a twenty-one year old student in Leipzig named Ferdinand de Saussure published a study of the Indo-European vowel system. Certain roots behaved irregularly: they had long vowels or unexpected vowel colours where the ablaut system predicted something else. Saussure proposed that these roots had originally contained additional consonants, which he called coefficients sonantiques, that had disappeared everywhere but left their fingerprints on the neighbouring vowels. He did not know what they sounded like. He posited them purely because the system required something in those slots.

No attested Indo-European language had them. For decades the proposal was regarded by many as an elegant piece of algebra with no phonetic reality.

Then in 1915 Bedrich Hrozny showed that Hittite, recovered from cuneiform tablets at Bogazkoy in Anatolia, was Indo-European. In 1927 Jerzy Kurylowicz pointed out that Hittite had a consonant, written with a cuneiform sign transliterated as h, and that it turned up precisely in the positions where Saussure had predicted a missing consonant. A sound posited from internal arithmetic in 1879, in a language nobody had ever heard, was found in a language nobody had known existed when the prediction was made.

These consonants are now called laryngeals and are written *h1, *h2, and *h3. Modern reconstructions use them constantly: the word for father is written *ph2ter, which is why the a of Latin pater is there at all. Not every detail of laryngeal theory is settled, and how many there were is still argued. What is settled is that Saussure was right that something was there.

The upshot: Comparative reconstruction made a risky prediction about an unattested sound and a later discovery confirmed it. That is why the method's results are treated as knowledge rather than as speculation.

Trees, waves, and dating

August Schleicher drew the family as a branching tree in the 1860s, and the tree diagram has been standard ever since. Johannes Schmidt objected in 1872 that the shared features do not nest neatly: innovations overlap across branches in ways a tree cannot represent, and he proposed a wave model in which changes spread outward from centres like ripples. Both are right about something. Branching captures the splits; waves capture the continued contact between neighbouring dialects after the splits. Current practice uses trees while acknowledging that the earliest divisions were probably a dialect continuum rather than clean breaks.

One branching claim does command wide agreement: Anatolian split off first, and it is different enough that some scholars prefer to call the common ancestor of Anatolian and everything else Indo-Hittite. Tocharian is usually placed as the next to separate.

Dating is a separate problem with two anchors. The lower anchor is attestation: Hittite texts from about 1650 BCE, Mycenaean Greek tablets from about 1400 BCE, so the family had already fractured well before then. The upper anchor is cultural vocabulary. Words for wheel, axle, yoke, and conveying by vehicle reconstruct securely across widely separated branches: the wheel word gives English wheel, Greek kuklos, and Sanskrit cakra, and the yoke word gives Latin iugum, Sanskrit yugam, and English yoke. Wheeled vehicles do not appear in the archaeological record before roughly 3500 BCE. If the reconstructed vocabulary is inherited rather than separately borrowed, the community that spoke it had not yet dispersed before then. That argument is central to the homeland debate, and Lesson 8 takes it apart properly, including the objections to it.

Bottom line: The family tree is a useful simplification of what was probably a dialect continuum, and the date of the proto-language is bracketed between the archaeology of wheeled vehicles above and the earliest attested daughters below.

Common misconceptions

• Sanskrit is the ancestor of the Indo-European languages. Sanskrit is a daughter, not a parent. It is archaic and important as evidence, especially for accent and for the voiced aspirates, but Greek and Hittite preserve things Sanskrit lost.
• Proto-Indo-European is a known language you could learn. It is a set of reconstructed forms with an asterisk in front of them. Reconstructions of connected texts, such as Schleicher's fable, are teaching exercises, not recovered documents.
• Proto-Indo-European was the first or original human language. It is one proto-language among many, spoken perhaps six thousand years ago, tens of thousands of years after language itself began.
• The centum and satem groups are the two halves of the family. That was a nineteenth century view. Tocharian, in the far east, is centum, and the satem change is now read as an areal innovation.
• Reconstruction is guesswork dressed up in asterisks. Laryngeal theory predicted an unattested consonant that Hittite then supplied, and two unknown branches later slotted into the framework without adjustment.

Summing up

• Jones's 1786 discourse pointed at shared verb morphology and a lost common ancestor, and that framing launched comparative linguistics.
• Ten branches are recognised, two of them extinct and recovered only in the twentieth century, both of which fitted the existing reconstruction.
• The strongest evidence is shared irregular inflection, especially the verb to be, rather than shared vocabulary.
• The reconstructed sound system has three stop series and three dorsal positions, a near-missing *b that motivates the glottalic proposal, and the ablaut alternation that survives in English strong verbs.
• Saussure's 1879 postulation of missing consonants was confirmed when Hittite, discovered later, showed a consonant in exactly those positions.
• Trees and waves both capture part of the truth, Anatolian split first, and the proto-language is dated between the archaeology of wheeled vehicles and the earliest attested daughters.

Sources

1. Britannica. (n.d.). Indo-European languages. Encyclopaedia Britannica. britannica.com
2. Wikipedia contributors. (n.d.). Proto-Indo-European language. en.wikipedia.org
3. Wikipedia contributors. (n.d.). Laryngeal theory. en.wikipedia.org
4. Fortson, B. W. (2010). Indo-European language and culture: An introduction (2nd ed.). Wiley-Blackwell.
5. Mallory, J. P., & Adams, D. Q. (2006). The Oxford introduction to Proto-Indo-European and the Proto-Indo-European world. Oxford University Press.

A pot with a wagon scratched on it

At Bronocice in southern Poland, excavators recovered a clay vessel from the late fourth millennium BCE with an image incised into its surface that is generally read as a four-wheeled wagon, complete with a yoked draught pole. It is among the oldest depictions of a wheeled vehicle anywhere. Wheeled transport, on the current archaeological evidence, does not exist before roughly 3500 BCE and then appears across a broad zone of Europe and the Near East within a few centuries.

Now set that beside a linguistic fact. The Proto-Indo-European word for wheel reconstructs across widely separated branches: it gives English wheel, Greek kuklos, and Sanskrit cakra. Words for axle, yoke, draught pole, and conveying by vehicle reconstruct in the same way. If those words are inherited from a single ancestral community rather than borrowed later by each branch separately, then that community was still together after wheels existed. That single inference is the hinge of one of the longest running arguments in the humanities, and this lesson is about how much weight it can carry.

Key idea: Reconstructed vocabulary is evidence about the world its speakers lived in, but only as strong as the argument that the words were inherited rather than borrowed and that their meanings did not drift.

Linguistic palaeontology, and its three ways of failing

Linguistic palaeontology is the practice of inferring facts about a prehistoric culture from the vocabulary reconstructed for its language. The logic is simple: if a word for a thing reconstructs to the proto-language, the speakers presumably knew the thing. The practice is old, occasionally powerful, and has produced some of the most confident nonsense in the field. It fails in three ways, and you should know all three before trusting any conclusion drawn from it.

Failure one: meanings drift. This is where Lesson 5 earns its place. A reconstructed root gives Latin fagus, beech, and English beech, so nineteenth century scholars reconstructed a word for beech and drew a line on the map where beech trees grow, placing the homeland to the west of it. But the Greek reflex of the same root is phegos, and it means oak. The reconstruction supports a tree of some kind, and no more than that. Every argument that ran through the beech line collapsed.

Failure two: parallel borrowing. A word can spread across neighbouring languages after they separate and still end up looking inherited, particularly if the daughters were still in contact and the sound changes had not yet finished. The wheel vocabulary is exactly where this objection bites, and defenders of the inheritance reading answer that the wheel words show the branch-specific sound changes we would expect of inherited material rather than the shape of late loans. That is a real argument on both sides, not a slogan.

Failure three: absence proves nothing. If no common word for a thing reconstructs, it may mean the speakers did not know the thing. It may also mean the word existed and every branch happened to replace it, which is entirely normal. Arguments of the form there is no Proto-Indo-European word for X are the weakest kind in this literature.

Why this matters: A reconstruction supports a cultural claim only when the meaning is stable across branches, the form shows inherited sound development, and the argument does not rest on an absence.

What the vocabulary supports reasonably well

With those cautions in place, a fair amount does survive.

Domain	Reconstructed item and reflexes	What it supports
Cattle	Sanskrit gauh, Greek bous, Latin bos, English cow	Cattle keeping, and cattle as a measure of wealth
Sheep and wool	Latin ovis and lana, Sanskrit avih and urna, English ewe and wool	Sheep herding and textile production
Horse	Latin equus, Greek hippos, Sanskrit asvah, Old English eoh	Familiarity with horses, though not necessarily riding
Vehicles	Wheel, axle, yoke, draught pole across several branches	Wheeled transport before dispersal, if inherited
Honey and mead	Sanskrit madhu, Greek methu, English mead	Beekeeping or honey gathering, and a fermented drink
Sky deity	Greek Zeus pater, Latin Iuppiter, Vedic Dyaus Pitar	A sky father god invoked by an inherited two-word formula

The last row is the single most convincing item in the whole inventory, because it is not one word but a phrase, sky father, preserved as a unit in three branches that had no contact with each other for millennia. That is very hard to explain by borrowing.

The kinship terminology is the most interesting case, because the argument comes from the shape of the gap rather than from a single word. Proto-Indo-European reconstructs a full set of terms for the husband's relatives: husband's father, husband's brother, husband's sister, all with solid reflexes in Latin, Greek, and Sanskrit. There is no comparable set for the wife's relatives. The natural reading is that a married woman moved into her husband's household and needed names for the people she now lived among, while a married man did not acquire a new set of daily relations to name. The society was, in the usual terms, patrilocal and patrilineal. This is a genuinely strong inference, and note why: it rests on a structured pattern across many words rather than on the meaning of any single one.

The point: The most reliable cultural inferences come from patterns across many reconstructed items, or from inherited multi-word formulas, not from the gloss of a single root.

Two homelands, argued at their strongest

Where and when did these people live? Two families of answer have dominated for forty years, and each has serious evidence.

The steppe hypothesis. Associated with Marija Gimbutas, who developed it from the 1950s, and given its fullest modern statement by David Anthony in 2007, this places the homeland on the Pontic-Caspian steppe north of the Black and Caspian seas, in roughly the fourth millennium BCE, with the Yamnaya archaeological culture as the best candidate for the late proto-language community. Its case: the reconstructed vocabulary of wheeled vehicles, horses, wool, and mobile pastoralism matches steppe archaeology of that period and does not match a farming society two thousand years earlier. Horse domestication and wheeled wagons appear in that zone at roughly the right time. The reconstructed social vocabulary, with its patrilocal kinship and its cattle-based wealth and its formulas of heroic reputation, fits a mobile herding society. And the linguistic dating anchors, if you take the wheel vocabulary as inherited, exclude a date much before 3500 BCE.

The Anatolian hypothesis. Proposed by Colin Renfrew in 1987, this places the homeland in Anatolia around 7000 BCE and has the languages spread with farming, carried by the demographic expansion of agriculturalists into a thinly populated Europe. Its case: agriculture is the one process in European prehistory large enough to have replaced the languages of an entire continent, and there is no need to invent an invasion to explain it. Renfrew objected that Gimbutas's model relied on waves of conquest for which the archaeology was thin. Bayesian phylogenetic analyses of vocabulary by Russell Gray and Quentin Atkinson in 2003, and by Remco Bouckaert and colleagues in 2012, produced age estimates of roughly eight to nine thousand years, which is far too old for the steppe model and close to the Anatolian date.

Those computational results were then contested with the same tools. Will Chang and colleagues published an analysis in 2015 that constrained the phylogeny with attested ancient languages, and recovered dates consistent with the steppe hypothesis instead. Two teams, similar methods, opposite answers, depending on the assumptions built in. That is worth noticing on its own, and Lesson 13 returns to why vocabulary-based dating behaves this way.

Remember: The steppe case rests on the fit between reconstructed vocabulary and fourth millennium archaeology; the Anatolian case rests on the demographic scale of the farming expansion and on computational dates that later work has disputed.

What ancient DNA changed, and what it did not

From 2015 the argument acquired a new evidence stream. Wolfgang Haak and colleagues, publishing in Nature, and a parallel team led by Morten Allentoft, sequenced ancient genomes from across Europe and found a striking result: around 2900 to 2500 BCE, populations in central Europe associated with the Corded Ware culture derive a large fraction of their ancestry, in some samples a majority, from a steppe population closely related to the Yamnaya. A very large movement of people from the steppe into Europe had happened, at roughly the time and along roughly the route the steppe hypothesis required.

Be precise about what that shows. It shows that people moved. It does not by itself show that a language moved with them, because genes and languages are not the same thing and dissociate constantly: the population of Ireland is genetically continuous across a period in which the community language changed to English. What the genetics did was remove the objection that no adequate demographic event existed on the steppe side, which had been Renfrew's strongest point. That is a serious change in the state of the argument, not a proof.

And the genetics has produced complications for the steppe account too. A 2018 study of ancient genomes from Anatolia led by Peter de Barros Damgaard found little or no steppe-derived ancestry in Bronze Age Anatolian individuals, which is awkward if Hittite arrived from the steppe. More recent genomic work has continued to refine the picture, including proposals locating the source population for both Anatolian and the later steppe groups in the region north of the Caucasus. Meanwhile Paul Heggarty and colleagues published a large phylogenetic study in 2023 arguing for a hybrid solution: a deeper origin south of the Caucasus for the family as a whole, with the steppe serving as the secondary centre from which most European branches spread.

So what?: Ancient DNA established that a major steppe migration into Europe occurred, which strengthened one side, but it also complicated the Anatolian branch, and it cannot on its own demonstrate that any particular language accompanied any particular set of genes.

Where the argument actually stands, and what would move it

A majority of Indo-Europeanists currently favour some version of a steppe origin, and it would be misleading to pretend the two positions are evenly balanced. It would be equally misleading to call the matter closed. Specific problems remain unsolved: the position of Anatolian, the near absence of steppe ancestry in Bronze Age Anatolia, and the persistent gap between vocabulary-based dating methods and archaeologically anchored ones.

Three kinds of new evidence would genuinely move the argument. An Indo-European text older than the Hittite tablets, from anywhere, would constrain the chronology directly. Dense ancient genomic sampling from Anatolia and the southern Caucasus across the fifth and fourth millennia would test the competing routes. And a dating method whose assumptions both camps accept in advance would remove the current situation, in which the answer tracks the assumptions.

One thing this course will state without hedging. The homeland question is a question about a speech community, and it has nothing to do with race. In the nineteenth and twentieth centuries the term Aryan, which is simply the self-designation of the Indo-Iranian speakers, was taken up by European racial theorists and then by the Nazi state to underwrite a fantasy of a master people. That usage has no basis in the linguistics. A language family is a set of languages related by descent, and its speakers at any period were of whatever ancestry the region held. Some of Gimbutas's own later cultural claims, particularly her reconstruction of a peaceful matriarchal Old Europe overrun by patriarchal invaders, are likewise not accepted by most archaeologists and should not be mistaken for the linguistic evidence.

Bottom line: The steppe model currently has the better fit to the evidence, the Anatolian problem is real and unresolved, and the entire question is about language transmission rather than ancestry or race.

Common misconceptions

• Ancient DNA settled the homeland question in 2015. It showed a large steppe migration into Europe. Languages and genes travel independently often enough that the inference still requires argument.
• Proto-Indo-European speakers were a race, or the Aryans. They were a speech community. The racial reading is a nineteenth century invention that was later put to murderous political use and has no linguistic content.
• If a word does not reconstruct, the speakers did not have the thing. Words are replaced constantly. Arguments from absence are the weakest form in linguistic palaeontology.
• Reconstructed meanings are as secure as reconstructed sounds. The beech line argument failed precisely here, since the same root means oak in Greek.
• Computational dating has resolved the chronology. Two well constructed Bayesian analyses have produced answers two to three thousand years apart depending on their assumptions, which is the problem rather than the solution.

Where this leaves us

• Linguistic palaeontology infers culture from reconstructed vocabulary and fails through semantic drift, parallel borrowing, and arguments from absence.
• The vocabulary supports cattle and sheep keeping, wool, honey, horses, probably wheeled vehicles, and a sky father deity preserved as an inherited two-word formula.
• The asymmetry in kinship terms, with the husband's relatives named and the wife's not, is a strong structural argument for patrilocal residence.
• The steppe hypothesis places the homeland north of the Black Sea in the fourth millennium BCE; the Anatolian hypothesis places it in Anatolia around 7000 BCE with the spread of farming.
• Ancient DNA from 2015 showed a large Yamnaya-related migration into central Europe, strengthening the steppe case without proving language transmission, while later work has complicated the position of Anatolian.
• The question concerns a speech community, not a race, and the Aryan racial reading is a political fabrication with no linguistic basis.

Sources

1. Wikipedia contributors. (n.d.). Proto-Indo-European homeland. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Kurgan hypothesis. en.wikipedia.org
3. Britannica. (n.d.). Indo-European languages. Encyclopaedia Britannica. britannica.com
4. Anthony, D. W. (2007). The horse, the wheel, and language: How Bronze-Age riders from the Eurasian steppes shaped the modern world. Princeton University Press.
5. Renfrew, C. (1987). Archaeology and language: The puzzle of Indo-European origins. Cambridge University Press.
6. Haak, W., Lazaridis, I., Patterson, N., et al. (2015). Massive migration from the steppe was a source for Indo-European languages in Europe. Nature, 522(7555), 207-211.

Eight hundred languages in one country

Papua New Guinea has a population of roughly ten million people and more than eight hundred languages, which is somewhere near one in eight of all the languages on earth. Iceland has a population of about 380,000 and one language. Neither number is an error, and the contrast is the first thing to understand about the world's linguistic map: diversity is not distributed anything like evenly, and the reasons are historical rather than geographical accidents.

Counting the world's languages is harder than it sounds, because the line between a language and a dialect is partly political. Ethnologue's recent editions list somewhat over seven thousand living languages, and Glottolog, which is stricter about what counts as a distinct variety, arrives at a comparable figure. Both catalogues sort those languages into roughly four hundred independent families and isolates. That number is the surprising one. The world is not a handful of large families. It is a few very large ones and a very long tail.

Key idea: Just over seven thousand languages fall into about four hundred demonstrated families and isolates, with a small number of families accounting for most speakers and a very long tail accounting for most of the diversity.

The large families

Family	Roughly where	Scale	Some members
Indo-European	Europe, Iran, South Asia, and the colonial diaspora	Most first-language speakers of any family	English, Spanish, Hindi, Russian, Persian, Bengali
Sino-Tibetan	China, the Himalaya, mainland Southeast Asia	Second by speakers, several hundred languages	Mandarin, Cantonese, Burmese, Tibetan
Niger-Congo, or Atlantic-Congo	Sub-Saharan Africa	The most languages of any family, well over a thousand	Swahili, Yoruba, Igbo, Zulu, Shona
Afro-Asiatic	North Africa, the Horn, Southwest Asia	Around three hundred languages	Arabic, Hebrew, Amharic, Hausa, Somali, ancient Egyptian
Austronesian	Taiwan to Madagascar to Polynesia	Over a thousand languages, the widest premodern spread	Malay and Indonesian, Tagalog, Malagasy, Maori, Hawaiian
Dravidian	Southern India and pockets further north	Around eighty languages	Tamil, Telugu, Kannada, Malayalam
Turkic	Turkey through Central Asia to Siberia	Around forty languages	Turkish, Azerbaijani, Uzbek, Kazakh, Uyghur
Uralic	Northern Europe and northwestern Siberia	Around forty languages	Finnish, Estonian, Hungarian, the Sami languages
Austroasiatic	Mainland Southeast Asia and eastern India	Around 150 languages	Vietnamese, Khmer, the Munda languages
Trans-New Guinea	The New Guinea highlands	Several hundred languages; its exact extent is debated	Enga, Dani, Kuman
Pama-Nyungan	Roughly seven eighths of the Australian continent	Around 250 languages historically	Warlpiri, Pitjantjatjara, Guugu Yimithirr
Mayan, Uto-Aztecan, Quechuan, Na-Dene, Algic	The Americas	Tens of languages each	Kaqchikel, Nahuatl, Quechua, Navajo, Cree

Two of these deserve a second look. Austronesian achieved the greatest geographical spread of any family before European colonial expansion: from a homeland in Taiwan, its speakers reached Madagascar off the coast of Africa and Rapa Nui in the eastern Pacific, an arc of more than half the circumference of the planet, using boats. And the Bantu subgroup of Niger-Congo, some five hundred languages, spread across most of central and southern Africa from a homeland near the modern Nigeria and Cameroon border, beginning in roughly the second millennium BCE.

Why this matters: The biggest families are the ones whose speakers expanded fastest, usually with agriculture, herding, or seafaring. Family size measures historical expansion, not linguistic merit.

Isolates

A language isolate is a language with no demonstrated relatives: a family with exactly one member. Around a hundred are recognised, and the list includes Basque in the western Pyrenees, Ainu in northern Japan, Burushaski in the Karakoram, Zuni and Seri in North America, and, among extinct languages, Sumerian and Elamite.

Basque is the one everyone knows, and it is worth being precise about what its status means. Basque is not older than other languages, not more primitive, and not a survival from before language families existed. Every language is exactly as old as every other, since each is a continuous chain of transmission back to whenever human language began. What isolate status means is narrower and stranger: Basque's relatives, if it had any, are gone, and no proposed connection to any other family has met the standard of evidence.

Isolate status can also be provisional. Korean is often called an isolate, though it is now usually grouped with Jeju as a small Koreanic family. Sometimes evidence arrives: Edward Vajda's 2010 proposal linking Na-Dene in North America with Yeniseian in central Siberia has been taken more seriously by specialists than most long-range proposals, though it is not settled. That is what a serious hypothesis at the frontier looks like, and Lesson 14 examines the frontier properly.

Remember: An isolate is a language whose relatives cannot be demonstrated, not a language without ancestry. Basque has exactly as much history as Latin; the difference is what survived alongside it.

Why diversity clumps where it does

Look again at the Papua New Guinea number. Why there? Johanna Nichols proposed a distinction that has become standard, between spread zones and residual zones.

A spread zone is territory across which one language repeatedly expands to replace others: open, traversable, and suited to mobile populations. The Eurasian steppe is the model case. Indo-Iranian, then Turkic, then Mongolic swept across it in turn, each replacing what came before, so a vast area holds few families. Australia, most of which is covered by a single family, is another.

A residual zone accumulates diversity instead of losing it: mountainous, fragmented, or otherwise hard to cross, so populations persist rather than being absorbed. The Caucasus holds three indigenous families in an area smaller than many countries. The New Guinea highlands, precolonial California, and the Amazon basin are all residual zones. The pattern is not about the people. It is about whether the terrain permits one group to spread over everyone else.

The point: Language diversity is high where geography prevents replacement and low where it permits expansion, which is why a single mountain range can hold more families than a continent-sized plain.

What is demonstrated and what is proposed

Several familiar groupings are not established families, and this distinction is one you should be able to make on sight.

Altaic once grouped Turkic, Mongolic, and Tungusic, sometimes with Korean and Japanese attached. The languages do share a great deal: vowel harmony, agglutinative morphology, subject object verb order. The problem is that the shared material is overwhelmingly typological and lexical rather than the kind of shared irregular morphology that demonstrates descent, and the vocabulary that does match is concentrated in exactly the semantic fields where borrowing is expected. Most specialists now treat the resemblances as the product of long contact across a spread zone, and Altaic is not accepted as a genetic unit. A related term, Transeurasian, is used by researchers who continue to argue the case.

Nilo-Saharan is a proposed African grouping whose internal evidence many Africanists consider too thin to support, and it is often described as a residual category for languages not placed elsewhere. Khoisan, once treated as a family on the basis of click consonants, is now generally analysed as at least three separate families plus isolates. Clicks are a shared areal feature, not proof of relationship, and some Bantu languages including Zulu and Xhosa have borrowed them.

The general lesson is that typological similarity, shared word order, shared morphological style, shared sound inventories, is not evidence of common descent. It is evidence of contact, of shared ancestry, or of nothing at all, and only the comparative method can tell which.

The upshot: A grouping is a family when systematic correspondences and shared morphological irregularities demonstrate it. Typological resemblance across a contact zone is not the same thing and does not become so through repetition.

Sign languages have genealogies too

One more correction to the map, and it surprises most people. Sign languages are full natural languages with their own histories, and they form families that have nothing to do with the spoken languages around them.

American Sign Language descends principally from French Sign Language, brought to the United States in 1817 when Laurent Clerc, a deaf French teacher, came to Hartford with Thomas Hopkins Gallaudet to found a school, where it mixed with signing already in use among deaf Americans, including the community on Martha's Vineyard. British Sign Language has an entirely separate history. The consequence is that ASL and BSL are not mutually intelligible, while ASL and French Sign Language remain substantially related, even though the surrounding spoken languages line up the other way.

Sign languages also give historical linguists something they almost never get: a documented birth. When a school for deaf children opened in Nicaragua in the late 1970s and brought together children who had had no shared language, a new sign language emerged among them within a few cohorts, and it became more grammatically systematic in each successive group of younger signers. Researchers were present while it happened. Nothing in the spoken record offers a comparable view of a language coming into existence.

Worth holding on to: Sign language families cut across spoken language families entirely, because they were transmitted through schools and deaf communities rather than through the surrounding hearing population.

Common misconceptions

• Some languages are older than others. Every living language is the endpoint of an unbroken chain of transmission of the same length. Basque is not older than Spanish; its relatives are simply gone.
• Big families are more successful languages. Family size records the expansion of populations, usually through agriculture, pastoralism, or seafaring. It says nothing about the languages themselves.
• Languages that look alike are related. Turkic, Mongolic, and Tungusic look alike in structure and are not accepted as a family. Shared typology can come from contact.
• Sign languages are visual codes for the surrounding spoken language. They are independent languages with independent histories, which is why ASL is related to French Sign Language and not to British Sign Language.
• Language diversity is highest where populations are largest. It is highest where terrain prevents any one group from spreading over the others, which is why New Guinea and the Caucasus beat entire continents.

What to carry forward

• Somewhat over seven thousand living languages fall into roughly four hundred demonstrated families and isolates.
• Indo-European leads by speakers, Niger-Congo and Austronesian by number of languages, and Austronesian achieved the greatest premodern geographical spread.
• An isolate such as Basque is a one-member family, not an ancient or primitive language.
• Spread zones such as the Eurasian steppe lose diversity to successive expansions; residual zones such as the Caucasus and New Guinea accumulate it.
• Altaic, Nilo-Saharan, and Khoisan are proposals or areal groupings rather than demonstrated families, because typological similarity is not evidence of descent.
• Sign languages form their own families, and the emergence of Nicaraguan Sign Language gave researchers a documented view of a language being born.

Sources

1. Eberhard, D. M., Simons, G. F., & Fennig, C. D. (Eds.). (n.d.). How many languages are there in the world? Ethnologue. ethnologue.com
2. Hammarstrom, H., Forkel, R., Haspelmath, M., & Bank, S. (n.d.). Glottolog. Max Planck Institute for Evolutionary Anthropology. glottolog.org
3. Linguistic Society of America. (n.d.). How many languages are there in the world? linguisticsociety.org
4. Linguistic Society of America. (n.d.). Sign language. linguisticsociety.org
5. Nichols, J. (1992). Linguistic diversity in space and time. University of Chicago Press.

Three languages that put the article in the wrong place

Romanian for man is om, and for the man is omul. Bulgarian for man is the word written мъж and transliterated mazh, and for the man it is мъжът, mazhat. Albanian for mountain is mal, and for the mountain is mali. In all three, the definite article is stuck on the end of the noun.

Now consider where those languages come from. Romanian is Romance, descended from Latin, which had no definite article at all. Bulgarian is Slavic, and no other Slavic language has a definite article, let alone a suffixed one. Albanian is its own branch of Indo-European. Three different branches, three different ancestries, one shared and unusual construction, and their nearest relatives outside the region, Italian, Polish, and so on, do not share it. They also share several other traits: the loss of the infinitive in favour of a subordinate clause, and a future tense built from a verb meaning to want.

None of this is inheritance. It is what centuries of multilingual neighbours do to each other's grammars, and it is the subject of this lesson.

Key idea: Languages in prolonged contact converge structurally, and the resulting similarities can look exactly like the signal the comparative method is designed to detect.

What gets borrowed, in what order

Borrowing is not random with respect to what kind of material is involved. There is a well established hierarchy, and it is a useful diagnostic tool.

Borrowed most readily	Example
Nouns, especially for new things	English tomato, sushi, algebra, robot
Other content words	English borrowed the verb to arrive from French
Function words	Rare; English they, them, their from Old Norse
Derivational affixes	English able, ment, and tion, all from French
Inflectional morphology	Very rare; requires intense sustained bilingualism
Syntax and phonology	Usually only in a linguistic area, over centuries

The hierarchy is why the Norse pronouns in English are so revealing. Ordinary contact gives you nouns. Old English already had perfectly good third person plural pronouns, hie and him and hira, and it replaced them with Norse forms. Nobody borrows a pronoun for want of one. That substitution is evidence of an intensity of contact, of the kind produced by two populations living intermixed and speaking closely related and partly intelligible languages, that few situations ever reach.

Scale matters too. English took roughly ten thousand words from French in the centuries after 1066, and a large majority are still in use. Japanese vocabulary of Chinese origin accounts for something close to half of dictionary entries. Spanish took several thousand words from Arabic during the centuries of Muslim rule in Iberia, many of them still carrying the fossilised Arabic definite article al: alcalde, almohada, alcohol, algebra. Borrowing on this scale does not make a language a member of the donor's family. Spanish is Romance, and English is Germanic, no matter how the dictionary looks.

Why this matters: Vocabulary borrowing is easy and can be enormous; structural borrowing is hard and requires sustained bilingualism. What was borrowed therefore tells you what kind of contact occurred.

How to spot a loanword

Four diagnostics, and you generally want more than one.

Its form breaks the native rules. English words do not normally begin with the sound sequence of ski or sk in general, because Old English changed that cluster; ski, sky, skirt, and skin are Norse or later borrowings. The sound in the middle of genre and beige does not occur in native English words at all.

It fails the sound laws. If a word ought to show a change and does not, it was not there when the change ran. Latin-derived paternal did not go through Grimm's Law because it arrived two thousand years too late.

Its distribution is patchy. An inherited word tends to occur across a family. A word found in one branch only, particularly if a neighbour has a similar one, is a candidate loan.

Its meaning sits in a borrowing-prone field. Trade, technology, religion, administration, food, and prestige goods are where loans cluster. The classic English illustration is the pairing of the animal in the field with the meat on the table: cow, pig, and sheep are inherited Germanic words, and beef, pork, and mutton come from French, which tells you something about who tended the animals and who ate them.

The point: Loanwords are identified by their form violating native patterns, their failure to undergo the relevant sound changes, their patchy distribution, and the semantic fields in which they cluster.

Substrate, superstrate, adstrate

Three terms describe the social relationship between the languages in contact, and they are worth keeping distinct.

A substrate is the language of a population that shifts to another language, leaving traces of its own in the result. The influence typically shows in pronunciation and sometimes in syntax rather than in vocabulary, because the shifting speakers carry their accents and their structural habits into the new language. Whether the Celtic languages of Britain left a substrate in English is a long-running argument: some scholars attribute the English use of do in questions and negatives to Celtic influence, since Welsh and Cornish have comparable constructions, while others point to the long delay between the shift and the appearance of the construction.

A superstrate is the language of an incoming dominant group that does not displace the local language but leaves a heavy deposit in it, usually in vocabulary of power, law, and prestige. Norman French in English is the standard example, and it is why English legal and governmental vocabulary is overwhelmingly French: judge, jury, court, parliament, government, sovereign.

An adstrate is a neighbouring language of comparable standing, with influence running both ways and neither displacing the other. Long-standing bilingual borderlands produce adstrate relationships.

One more mechanism belongs here: the calque, or loan translation, where the structure is borrowed and the material is native. French gratte-ciel and German Wolkenkratzer are word-for-word translations of skyscraper. English brainwashing is a translation of a Chinese expression. The word loanword is itself a calque of German Lehnwort, which is a small joke the discipline enjoys.

Remember: Substrate influence comes from below through language shift and shows in structure and accent; superstrate influence comes from above and shows in prestige vocabulary; calques import structure while leaving no foreign material behind at all.

Linguistic areas

Now back to the Balkans. A linguistic area, often called by the German term Sprachbund, is a region in which languages from different families have converged on shared structural features through long multilingual contact.

The Balkan area is the classic case, and the shared features include the suffixed definite article we started with, the loss of the infinitive, a future built from a verb meaning to want, and the merging of the genitive and dative cases. Greek, Albanian, Romanian, Bulgarian, and Macedonian participate to varying degrees. No single language is the source of all of it; the features spread among a population that was routinely multilingual for centuries.

South Asia is the second great example, described by Murray Emeneau in a 1956 paper that gave the study of linguistic areas much of its modern shape. Indo-Aryan languages such as Hindi and Dravidian languages such as Tamil are unrelated, yet across the subcontinent they share retroflex consonants, verb-final word order, a construction using a verb of saying to mark quotation, and a set of converb constructions for chaining clauses. Sanskrit has retroflex consonants that its Indo-European relatives outside the subcontinent lack, which is one of the oldest arguments for Dravidian substrate influence.

Mainland Southeast Asia is a third: Sino-Tibetan, Austroasiatic, and Kra-Dai languages there share lexical tone, numeral classifiers, and highly isolating morphology, across family lines. And Europe itself forms a weaker area, which Benjamin Lee Whorf named Standard Average European, characterised by definite and indefinite articles, a perfect formed with a verb meaning to have, and relative clauses introduced by relative pronouns, a bundle that is far from universal worldwide.

The upshot: Areal convergence produces structural similarity across family boundaries, which is precisely the pattern that a careless comparison would misread as descent.

When contact goes all the way

At the extreme end, contact produces languages that no tree can represent. Mixed languages take major components from two sources in a way that makes the question of which family they belong to unanswerable.

Michif, spoken by Metis communities in Canada and the northern United States, takes its noun phrases, with their articles and adjectives, from French and its verb phrases, with their complex morphology, from Cree. Not vocabulary from one and grammar from the other, but two whole grammatical subsystems side by side in one language. Media Lengua, spoken in Ecuador, takes essentially all of its word roots from Spanish and essentially all of its grammatical structure from Quichua.

These languages are rare, and they matter out of proportion to their number, because they are counterexamples to the assumption that every language has exactly one parent. The comparative method assumes a tree. Most languages fit it. A few genuinely do not, and Module 4's final lesson covers the largest category of such cases.

So what?: The family tree model works for the vast majority of languages but is not a law of nature, and mixed languages show what its exceptions look like.

Common misconceptions

• Heavy borrowing changes a language's family. English took ten thousand words from French and remains Germanic, because descent is determined by the inherited core of grammar and basic vocabulary.
• Borrowing damages a language. Every documented language borrows. The languages with the largest borrowed vocabularies, including English and Japanese, are not impaired in any measurable way.
• Structural similarity across neighbouring languages proves relationship. The Balkan languages share a suffixed article and belong to three different branches. Contact explains it; descent does not.
• Substrate influence shows up mainly as vocabulary. Shifting speakers carry accents and structures rather than words, which is why substrate effects are argued from phonology and syntax and are correspondingly harder to prove.
• Every language has exactly one ancestor. Michif and Media Lengua take entire grammatical subsystems from two sources, which no single-parent tree can represent.

Pulling it together

• Material is borrowed in a rough order: nouns most readily, then other content words, then function words, affixes, and finally syntax and phonology.
• The Norse pronouns in English mark an extreme intensity of contact, because pronouns are almost never borrowed.
• Loanwords are detected by native phonotactic violations, failure to show expected sound changes, patchy distribution, and clustering in trade, prestige, and technology vocabulary.
• Substrate influence arrives from a shifting population and shows structurally; superstrate influence arrives from a dominant one and shows lexically; calques import structure using native material.
• The Balkans, South Asia, and mainland Southeast Asia are linguistic areas in which unrelated languages converged on shared features.
• Mixed languages such as Michif and Media Lengua combine whole subsystems from two sources and cannot be placed in a single-parent tree.

Sources

1. Wikipedia contributors. (n.d.). Sprachbund. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Balkan sprachbund. en.wikipedia.org
3. Britannica. (n.d.). English language. Encyclopaedia Britannica. britannica.com
4. Thomason, S. G., & Kaufman, T. (1988). Language contact, creolization, and genetic linguistics. University of California Press.
5. Emeneau, M. B. (1956). India as a linguistic area. Language, 32(1), 3-16.

A parliament debating in a language two centuries old

Papua New Guinea conducts much of its parliamentary business in Tok Pisin, which is one of its three official languages and the country's most widely used lingua franca in a nation of more than eight hundred languages. Most of its vocabulary comes from English. Gras means grass and also hair. Haus is house. Pikinini, child, comes from Portuguese pequenino by way of the Pacific trade. The word bilong, from English belong, does the work of the preposition of, so hair is gras bilong het.

Now look at the pronouns, because that is where the interesting part is. Tok Pisin distinguishes yumi, we including you, from mipela, we not including you. English has no such distinction and never has. The languages of the region, which are Austronesian, make it routinely. Tok Pisin took its words from one place and that piece of its grammar from another, and both of those facts are typical of the kind of language it is.

Key idea: Pidgins and creoles are the clearest cases in linguistics of languages with a datable origin, and they typically take their vocabulary from one source and important parts of their structure from another.

The pidgin stage

A pidgin is a contact language that arises when groups without a common language need to communicate for limited purposes, usually trade or labour. Three properties define it.

It has no native speakers. Everyone who uses a pidgin has another first language, and the pidgin is a supplement. It is restricted in domain: it does the work of the dock or the market or the plantation and is not used for storytelling, courtship, or argument. And it is structurally reduced relative to the languages that fed it, with a small vocabulary, little or no inflection, and heavy reliance on context.

None of that means a pidgin is broken. It is a functioning system, with its own conventions, that has been shaped for a narrow purpose. But it is not anyone's language, and that is the crucial fact, because it means the ordinary process of transmission to children is missing.

Why this matters: A pidgin is defined by having no native speakers and a restricted range of use, not by being defective. Its limits are the limits of the job it was built for.

What happens when children arrive

A creole is what a pidgin becomes when a generation of children acquires it as a first language. That change of status changes the language, because a first language has to do everything: tell a story, express doubt, mark time, embed one clause inside another, insult someone precisely.

The expansion is systematic and rapid, often within a generation or two. Vocabulary grows. Word order fixes. Subordinate clauses develop. And, most strikingly, a set of preverbal particles arises to mark tense, mood, and aspect, in patterns that recur across creoles with entirely different vocabulary sources.

Haitian Creole, which has around twelve million speakers and has been co-official with French since the constitution of 1987, shows the system cleanly.

Haitian Creole	Meaning	Marker
Mwen manje	I eat, or I ate	No marker; interpretation from context and verb type
Mwen te manje	I had eaten	te, anterior
Mwen ap manje	I am eating	ap, ongoing action
Mwen pral manje	I am going to eat	pral, future

Every one of those markers is a French word: te from ete, ap from apres, pral from pour aller. Not one of them functions as it does in French, and French does not mark tense with preverbal particles at all. The material is French; the system is not. That combination is the signature of creole formation, and it is why simply counting French-derived words tells you almost nothing about what Haitian Creole is.

Remember: Creoles standardly build a preverbal tense, mood, and aspect system out of lexifier material that did not have that function in the lexifier.

Where the Atlantic creoles came from

The Caribbean and West African creoles have a specific and brutal history. On plantations built on enslaved labour, colonial powers deliberately mixed people from different language backgrounds to impede communication and organisation. The socially dominant language, the lexifier, supplied most of the vocabulary because access to it was the route to survival. The languages of the enslaved population, the substrate, contributed structure.

Creole	Lexifier	Where
Haitian Creole	French	Haiti
Jamaican Creole, or Patwa	English	Jamaica
Sranan Tongo	English	Suriname
Papiamentu	Portuguese and Spanish	Aruba, Bonaire, Curacao
Mauritian Creole	French	Mauritius
Cape Verdean Creole	Portuguese	Cape Verde

The demographic detail matters for the theories that follow. Where the enslaved population vastly outnumbered the speakers of the lexifier, and where new arrivals kept the proportion of people acquiring the language imperfectly high, creolisation was fastest and the result furthest from the lexifier.

The point: The Atlantic creoles are not accidents of drift. They were produced by a specific labour system that separated the source of vocabulary from the source of grammar.

Three explanations, argued

Why do creoles from different continents, with different vocabulary sources, resemble each other structurally? Three answers compete, and none has won outright.

The bioprogram hypothesis. Derek Bickerton argued in the early 1980s that when children are given impoverished and inconsistent input, they fall back on innate structural defaults, and that creoles reveal those defaults directly. His main evidence came from Hawaiian Creole English, where he interviewed elderly speakers who had grown up as the language nativised and found that their generation used structures absent from their parents' pidgin. The recurring tense and aspect systems across unrelated creoles were, on this view, the human language faculty showing through.

The substrate account. Claire Lefebvre and others argue that the structural similarities come from the substrate languages, which in the Atlantic case were largely from a limited region of West Africa and shared many features. On this account Haitian Creole's grammar is substantially that of Gbe languages such as Fon, with French words substituted in. The similarities across creoles then reflect the similarities among the substrate languages and the recurrence of the same social situation, not an innate template.

The gradualist account. Robert Chaudenson and Salikoko Mufwene argue that the process was slower and more continuous than either of the above, that the input was not a stripped-down pidgin but the ordinary non-standard colonial varieties of French or English, and that creoles developed through the normal mechanisms of second language acquisition and language change over several generations, with no abrupt break.

Where does this stand? The strong bioprogram claim has few adherents now, partly because the historical record has turned out to show more gradual development than a single-generation account predicts. Most current work is multi-causal: substrate structures, universal tendencies in second language acquisition, and the specifics of who was present in what proportions all contribute, and the mix differs by colony. The honest summary is that the general shape of the process is agreed and the weighting of the causes is not.

The upshot: Creole similarities are now generally attributed to a combination of substrate influence, acquisition universals, and demographic history rather than to any single mechanism.

Creole exceptionalism

A second dispute is sharper, because it is partly about the discipline itself.

Some linguists, notably John McWhorter, have argued that creoles form a recognisable structural class, characterised by relatively little inflectional morphology, little lexical tone, and little semantically opaque derivation, and that this profile follows from their recent origin: complexity of that kind takes time to accumulate.

Michel DeGraff has argued against what he calls creole exceptionalism, on both empirical and historical grounds. Empirically, he holds that the proposed diagnostics do not hold up under examination, that plenty of non-creole languages show the same profile, and that creoles vary among themselves as much as other languages do. Historically, he points out that the idea of creoles as a simplified class emerged from nineteenth and twentieth century writing that also described their speakers in explicitly racist terms, and argues that the burden of proof on any claim of special simplicity should be correspondingly heavy.

Both positions are held by serious linguists and the dispute is live. What is not in dispute is the point that matters most outside the seminar room: Haitian Creole is a complete language with a full grammar, a literature, and constitutional status, and calling it broken French is false in the same way and for the same reasons discussed in LING 320.

Bottom line: Whether creoles constitute a distinct structural class is genuinely contested; whether they are full languages is not.

Common misconceptions

• A creole is a broken or simplified version of its lexifier. Haitian Creole marks tense, mood, and aspect with a preverbal system French does not have. It is a different language, not a damaged one.
• Pidgin and creole are interchangeable terms. A pidgin has no native speakers and a restricted range; a creole is a first language doing everything a first language must do.
• Creoles have no grammar. They have rigid and describable grammars. Speakers reject ungrammatical sentences as reliably as speakers of any other language.
• Creoles arise wherever languages meet. They arise under specific and largely brutal demographic conditions, above all plantation slavery, in which access to the lexifier was restricted and speakers of many substrate languages were mixed deliberately.
• The bioprogram hypothesis is the accepted explanation. It was influential and is now a minority position; current accounts combine substrate influence, acquisition universals, and colony-specific demography.

The takeaway

• A pidgin has no native speakers, serves restricted purposes, and is structurally reduced; a creole is what it becomes when children acquire it as a first language.
• Creolisation expands vocabulary, fixes word order, develops subordination, and typically builds a preverbal tense, mood, and aspect system.
• Haitian Creole marks anterior, ongoing, and future with te, ap, and pral, all from French words that do not carry those functions in French.
• Atlantic creoles arose from plantation slavery, with the lexifier supplying vocabulary and the substrate languages contributing structure.
• Bioprogram, substrate, and gradualist accounts compete, and current work combines them rather than choosing one.
• Whether creoles form a distinct structural class is contested; that they are complete languages is not, and Haitian Creole has been co-official in Haiti since 1987.

Sources

1. Britannica. (n.d.). Creole languages. Encyclopaedia Britannica. britannica.com
2. Britannica. (n.d.). Pidgin. Encyclopaedia Britannica. britannica.com
3. Wikipedia contributors. (n.d.). Tok Pisin. en.wikipedia.org
4. DeGraff, M. (2003). Against creole exceptionalism. Language, 79(2), 391-410.
5. Bickerton, D. (1984). The language bioprogram hypothesis. Behavioral and Brain Sciences, 7(2), 173-188.

A slab of granodiorite found in July 1799

French soldiers rebuilding a fort near the town of Rashid in the Nile delta, which Europeans called Rosetta, turned up a broken slab of dark stone carrying three blocks of writing. The top block was Egyptian hieroglyphs, the middle was the cursive Egyptian script called demotic, and the bottom was Greek, which European scholars could read at once. The Greek recorded a priestly decree issued at Memphis in 196 BCE, and it stated that the decree had been inscribed in all three scripts. The stone passed to Britain under the terms of the surrender in Egypt and has been in the British Museum since 1802.

The last person who could read hieroglyphs had died more than a thousand years earlier. Twenty-three years after the stone surfaced, someone read them again. This lesson is about how that is done, why it usually is not, and what written evidence is worth to a historical linguist once it exists.

Key idea: A script is a code for a language, and reading an unknown script is fundamentally the problem of working out which language it encodes and how.

What a writing system encodes

Writing systems differ in the size of the linguistic unit each symbol stands for. There are six broad types, and most real systems mix them.

Type	A symbol represents	Examples
Logographic	A word or morpheme	Chinese characters, most of which also carry a phonetic component
Syllabic	A syllable or mora	Japanese kana, Linear B, Cherokee
Alphabetic	A consonant or vowel	Greek, Latin, Cyrillic
Abjad	A consonant, with vowels optional	Arabic, Hebrew
Abugida	A consonant with an inherent vowel, modified by marks	Devanagari, Ethiopic, Thai
Featural	Articulatory features, assembled into syllable blocks	Korean hangul

Two points follow that people routinely get wrong. First, no writing system in use for a natural language is purely pictographic. Pictures cannot express grammatical endings, proper names, or abstractions, and every system that began with pictures acquired phonetic signs quickly, usually by the rebus principle of using a picture for its sound rather than its sense. Chinese characters are not pictures of ideas; the large majority contain a component indicating pronunciation.

Second, writing is a technology, not language itself. It was invented independently only a small number of times, probably in Mesopotamia around 3200 BCE, in China by around 1200 BCE, and in Mesoamerica in the first millennium BCE, with Egyptian either independent or stimulated by contact with Mesopotamia. Every other script in the world descends from or was inspired by one of those. Compare that with speech, which every human community has. For a historical linguist this matters because writing is evidence about language rather than language itself, and it is evidence with a systematic lag.

Why this matters: Writing systems encode sound at various grain sizes, no system is purely pictographic, and writing is a rare invented technology rather than a universal property of language.

What a decipherment needs

Decipherments come in three difficulty grades, and knowing which one you face tells you your chances.

Known language, unknown script. This is the tractable case. If you already know what the language is, you can work out how the signs encode it. Linear B was solved this way, in the end.

Known script, unknown language. Also possible, though what you get is a pronunciation rather than a meaning. Etruscan can be read aloud, since it uses an alphabet derived from Greek, and it is still only partly understood, because the language is an isolate with limited texts.

Both unknown. Close to hopeless without exceptional luck. This is where the undeciphered scripts sit.

Beyond that, four resources make the difference: enough text to work with, a bilingual inscription if you are fortunate, proper names, which usually keep their sound across languages and provide a phonetic foothold, and a known context that lets you guess what a text is likely to say. Almost every successful decipherment has used proper names as its opening.

The point: The decisive question is not how strange the signs look but whether the underlying language is known or recoverable.

Three decipherments

Egyptian. Thomas Young made real progress between 1814 and 1819, correctly identifying the cartouches, the oval rings enclosing royal names, and reading some phonetic values in the name Ptolemy. Jean-Francois Champollion completed the work and announced it in a letter published in September 1822. Two things made the difference. He recognised that hieroglyphs were not one thing but a mixture of phonetic signs, logograms, and determinatives that mark semantic category. And he knew Coptic, the last stage of the Egyptian language, written in a Greek-derived alphabet and preserved in Christian liturgy. Coptic gave him the language behind the script. Comparing the cartouches of Ptolemy and Cleopatra, which share several sounds, let him check individual sign values against each other.

Linear B. Arthur Evans, excavating at Knossos on Crete from 1900, recovered clay tablets in two related scripts he named Linear A and Linear B. He was certain that the language was not Greek, and he sat on much of the material for decades. Alice Kober, working in the 1940s at Brooklyn College without any assumption about the language, identified sets of words that appeared in three related forms, and showed that the script must be recording an inflected language and that certain signs shared a consonant or a vowel. Her grids gave the structure of the syllabary without giving its sound values. Michael Ventris, an architect who had been obsessed with the script since seeing Evans lecture as a schoolboy, built on that work and announced in 1952 that the language was Greek, several centuries older than Homer, and that Evans had been wrong. John Chadwick, a Greek philologist, joined him and they published in 1953.

Then came the confirmation, which is the part worth remembering. Carl Blegen, excavating at Pylos, had a newly found tablet that Ventris had never seen. Applying the proposed values, it read out as a list of vessels, including a word transcribed ti-ri-po-de, which is the Greek for two tripods, beside a drawing of a three-legged pot. A decipherment that produces the right answer on a text it was not built from is a decipherment that is correct.

Maya. The Maya case took longer and shows how a field can obstruct itself. In the 1560s the Spanish bishop Diego de Landa burned Maya books and also, in the same period, recorded from an informant what he took to be a Maya alphabet, which was actually a set of syllable signs. Eric Thompson, the dominant Mayanist of the mid twentieth century, held that the script was symbolic rather than phonetic and that the inscriptions concerned astronomy and ritual rather than history. Yuri Knorozov, working in the Soviet Union in 1952 with limited access to materials, argued from Landa's list that the signs were largely syllabic. Heinrich Berlin identified emblem glyphs standing for particular cities in 1958, and in 1960 Tatiana Proskouriakoff showed that the dates on the monuments at Piedras Negras clustered into human lifespans, which meant the inscriptions were dynastic history. Once those pieces were in place the script opened, and Maya texts are now read as the historical records of named rulers.

Remember: Each decipherment turned on identifying the underlying language and finding a phonetic foothold, and each was confirmed by producing sense on material the decipherer had not used.

What is still unread, and why

Several scripts have resisted, and the reasons differ.

Linear A can be given approximate sound values, because it shares many signs with Linear B, but the language behind it, usually called Minoan, is not known to be related to anything. You can pronounce it and not understand it.

The Indus script, from the Bronze Age cities of the Indus valley, has thousands of inscriptions, but almost all are extremely short, averaging around five signs, with no bilingual and no securely identified language. There is even a serious argument, advanced by Steve Farmer, Richard Sproat, and Michael Witzel in 2004, that it does not encode language at all but is a system of non-linguistic symbols. That claim is contested, and the honest position is that we do not know.

Rongorongo, from Rapa Nui, survives in about two dozen objects, and the community that could read it was destroyed by slave raiding and epidemic disease in the nineteenth century before anyone recorded how it worked.

The common factor is not difficulty of the signs. It is the absence of enough text, an unknown language, or both.

The upshot: Scripts stay unread when the corpus is too small, the language is unidentified, or the tradition of reading them was broken before it could be recorded.

The traps in written evidence

Finally, the cautions, because written evidence is the best kind a historical linguist can have and it still misleads in predictable ways.

Spelling is conservative, so a text records how a word was written rather than how it was said, and the gap widens over time; this is the Great Vowel Shift problem from Lesson 1. Scribal convention can hide distinctions the language made or invent ones it did not. Logographic writing can conceal pronunciation entirely: Hittite scribes often wrote Sumerian and Akkadian logograms in the middle of Hittite sentences, so a word can appear in text hundreds of times without its Hittite pronunciation ever being written. And a syllabary can be systematically ambiguous. Linear B does not distinguish l from r, does not mark the difference between voiced, voiceless, and aspirated stops, and cannot write a consonant at the end of a syllable, so a Greek word can be spelled in a way that supports several readings. That is one reason Evans could look at the tablets for thirty years without seeing Greek in them.

Written records are also socially skewed. They over-represent the literate, the administrative, and the formal, which is why we know a great deal about Mycenaean palace inventories and almost nothing about how Mycenaeans spoke to their children.

Bottom line: Writing gives dated evidence, which is irreplaceable, but it lags pronunciation, can hide it entirely under logograms, is systematically ambiguous in some scripts, and records a narrow slice of a society.

Common misconceptions

• Hieroglyphs are picture-writing. They are a mixed system of phonetic signs, logograms, and semantic determinatives, which is exactly what Champollion had to recognise before he could read them.
• Chinese characters represent ideas rather than sounds. Most characters contain a phonetic component, and the system encodes a specific language, not thought in general.
• Ventris cracked Linear B alone from nothing. Alice Kober's structural analysis of inflectional sets built the grid he worked from, and Chadwick's philology was needed to interpret the Greek.
• An undeciphered script must use very difficult signs. Linear A signs are largely readable; the obstacle is that the language behind them is unknown.
• A language without writing has no history. The comparative method reconstructs histories for unwritten languages; writing supplies dates, not history itself.

Looking back

• Writing systems encode language at the level of morpheme, syllable, consonant, or feature, and no system used for a natural language is purely pictographic.
• Writing was independently invented only a few times and is a technology rather than a universal property of language.
• Decipherment depends on whether the underlying language is known, on corpus size, on bilinguals, and above all on proper names as a phonetic foothold.
• Champollion succeeded because he knew Coptic and recognised that hieroglyphs mixed phonetic and semantic signs; Ventris succeeded on Kober's structural grids and was confirmed by the Pylos tripod tablet.
• The Maya script opened once Knorozov argued for syllabic values, Berlin identified emblem glyphs, and Proskouriakoff showed the inscriptions were dynastic history.
• Linear A, the Indus script, and rongorongo remain unread because of unknown languages, tiny corpora, or a broken reading tradition.
• Written evidence lags pronunciation, hides it under logograms, is ambiguous in syllabaries such as Linear B, and over-represents formal and administrative language.

Sources

1. British Museum. (n.d.). Everything you ever wanted to know about the Rosetta Stone. britishmuseum.org
2. Britannica. (n.d.). Rosetta Stone. Encyclopaedia Britannica. britannica.com
3. Wikipedia contributors. (n.d.). Linear B. en.wikipedia.org
4. Chadwick, J. (1958). The decipherment of Linear B. Cambridge University Press.
5. Coe, M. D. (1992). Breaking the Maya code. Thames and Hudson.

A formula that turns a word list into a date

In 1952, in the Proceedings of the American Philosophical Society, Morris Swadesh proposed that the core vocabulary of a language is replaced at a roughly constant rate: about 14 words in every hundred per thousand years. If that were true, you could count how many basic words two languages still share, plug the number into a formula, and read off how long ago they separated. Radiocarbon dating had been developed a few years earlier and was transforming archaeology. Swadesh was offering linguistics the same gift.

The method is called glottochronology, and the broader practice of comparing languages by counting shared vocabulary is lexicostatistics. Glottochronology is not much used today, and the reasons it failed are worth working through carefully, because they teach more about historical method than the method itself ever did.

Key idea: Glottochronology assumed that basic vocabulary decays at a constant rate like a radioactive isotope, and almost every step of that analogy turns out to be wrong.

The machinery

The procedure has three steps.

Step one: the list. Swadesh compiled a list of meanings chosen to be culture-neutral and universal, and therefore resistant to borrowing: pronouns, low numerals, body parts, natural phenomena, basic verbs. Two versions are standard, one of 200 items and a tighter one of 100.

Step two: count cognates. For each meaning, decide whether the two languages have cognate words. The proportion of cognate pairs is the retention figure, written C.

Step three: apply the formula. Time since separation is calculated as the natural logarithm of C, divided by twice the natural logarithm of the assumed retention rate r. The doubling is because both languages have been losing vocabulary independently since the split. With the 100-item list, r is taken as 0.86.

Work an example. Suppose two languages share 70 of 100 basic meanings as cognates. The natural log of 0.70 is about negative 0.357. The natural log of 0.86 is about negative 0.151, and twice that is negative 0.302. Dividing gives about 1.18, which the method reads as 1,180 years since separation. Run it again with 50 percent shared: the log of 0.50 is about negative 0.693, and dividing gives about 2.3, so roughly 2,300 years.

Notice what just happened. Two numbers went in and a confident date came out, with no error bars, no indication of how the cognacy judgments were made, and no way to tell a good case from a bad one. That tidiness is the problem in miniature.

Why this matters: The formula converts a single proportion into a single date with no expression of uncertainty, which is a warning sign in any method that claims to date the past.

The test that broke it

The decisive move was obvious and it took ten years for anyone to make it: run the method on languages whose histories are documented and see whether it gets the right answer.

Knut Bergsland and Hans Vogt did exactly that in a 1962 paper in Current Anthropology, using cases where the elapsed time was known independently from written records. The results were not close.

Icelandic, compared with the Old Norse of a thousand years earlier, had retained very nearly all of its basic vocabulary, far more than the assumed rate permits. Norwegian, separated from the same ancestor over the same interval, had retained noticeably less. Same starting point, same elapsed time, different rates. An East Greenlandic Inuit variety had lost basic vocabulary several times faster than the model allowed. Georgian and Armenian, both with long written records, likewise failed to match.

One number cannot be constant and also take different values for two languages over the same thousand years. The constant rate assumption is not approximately right with noise around it. It is wrong in a way that varies systematically with the social circumstances of the community, which is precisely the information the method claimed to have abstracted away from.

The point: Tested against known histories, the replacement rate varied enough between communities to make any date derived from it uninterpretable without knowing the history you were trying to recover.

Why the rate varies

The reasons are not mysterious once you look.

Word taboo. In a number of Australian and Amazonian societies, words resembling the name of a recently deceased person are avoided, sometimes for years, and substitutes are adopted. A society with that practice replaces basic vocabulary far faster than one without. This is a cultural variable, not a linguistic constant.

Literacy and standardisation. A written standard, a literary canon, and schooling all slow lexical replacement by keeping older forms in circulation. Iceland's saga tradition is exactly this effect, and it explains the Bergsland and Vogt result directly.

Contact. Intense contact accelerates replacement, and the assumption that basic vocabulary is borrowing-proof is simply false. English is the standard counterexample: the pronoun they, on the Swadesh list, is a Norse loan. So are the basic items give and egg in their modern forms. Any list-based method that assumes such items are safe will read borrowing as inheritance and produce a date that is too recent.

Cognacy judgment is not neutral. To decide whether two words are cognate you need the comparative method, which means glottochronology depends on the very analysis it was supposed to replace. Done carelessly, it counts look-alikes; done carefully, it presupposes the answer.

Remember: Lexical replacement rate is a social variable, driven by taboo practices, literacy, and contact, which is why no single constant can stand in for it.

What replaced it

Vocabulary-based dating did not disappear. It was rebuilt on better statistical foundations, borrowing methods from evolutionary biology, where the analogous problem of dating divergences from character data had been worked on for decades.

Modern Bayesian phylogenetic analyses differ from glottochronology in several concrete ways.

Glottochronology	Bayesian phylogenetics
One fixed replacement rate for all languages	Rates allowed to vary across branches and across meanings
One date, no uncertainty stated	A posterior distribution, reported as a credible interval
Pairwise comparison	Whole trees inferred and compared simultaneously
No external calibration	Calibration on attested ancient languages with known dates
Borrowing ignored	Borrowing modelled or coded out, imperfectly

These are real improvements, and the resulting work is serious science. It is also, as Lesson 8 showed, not yet decisive. Russell Gray and Quentin Atkinson's 2003 analysis produced Indo-European dates supporting the Anatolian hypothesis; Will Chang and colleagues, in 2015, constrained the analysis by placing attested ancient languages as ancestors rather than as sister lineages, and recovered dates supporting the steppe hypothesis instead. The data were largely the same. The assumptions differed, and the assumptions determined the answer by two to three thousand years.

That is not a reason to dismiss the methods. It is a reason to read their outputs as conditional statements: given this cognate coding, this calibration, and this model of borrowing, the date is such and such. Any presentation that drops those conditions is overselling.

The upshot: Bayesian methods fixed the fixed-rate assumption and the missing uncertainty, and did not fix the dependence of the result on cognacy coding, calibration choices, and the treatment of borrowing.

What historical linguists actually date with

So how do dates get assigned at all? Four sources, in descending order of confidence.

Attested texts. A dated inscription in a language is the gold standard: Hittite tablets place Anatolian in the second millennium BCE and nothing overturns that.

Loanword strata. If language A borrowed a word from language B in a form that reflects B's pronunciation at a known period, the contact is dated. Finnish borrowings from Germanic, which preserve very archaic Germanic shapes, are the classic case, and they date the contact relative to Germanic sound changes.

Terminus arguments from vocabulary. The wheel argument from Lesson 8 is one: if the reconstructed word is inherited, the community postdates the invention. These are only as strong as the inheritance claim.

Correlation with archaeology and genetics. The weakest, because it requires an assumption linking a material culture or a population to a language, and that link is exactly what is usually in question.

Which leaves an honest summary worth memorising. Relative chronology, which change happened before which, is well established and follows directly from the comparative method. Absolute chronology, how many years ago, is much weaker, and the confidence with which a date is stated in popular accounts of prehistory rarely reflects the confidence the evidence supports.

Bottom line: The order of events in a language's history can be established firmly; the calendar dates attached to them are inferences of a much softer kind, and should be read as such.

Common misconceptions

• Glottochronology is the linguistic equivalent of radiocarbon dating. Radioactive decay is a physical constant. Lexical replacement is a social behaviour that varies with taboo practice, literacy, and contact.
• Basic vocabulary cannot be borrowed. English they, give, and egg are Norse borrowings sitting on the Swadesh list itself.
• Bayesian phylogenetics has settled the dating question. Two careful analyses of the same family produced dates thousands of years apart because their calibration assumptions differed.
• A date with a credible interval is more reliable than one without. It is more honest, which is different. The interval reflects uncertainty within the model, not uncertainty about whether the model is right.
• Because glottochronology failed, vocabulary comparison is worthless. Lexicostatistics is still useful for grouping languages and generating hypotheses. It is the conversion of similarity into calendar years that failed.

What to remember

• Swadesh proposed in 1952 that basic vocabulary is replaced at a constant rate, allowing a separation date to be computed from a cognate percentage.
• The formula divides the log of the retained proportion by twice the log of the assumed rate, and produces a single date with no stated uncertainty.
• Bergsland and Vogt tested it in 1962 on documented histories and found rates varying widely, with Icelandic retaining far more than the model allowed.
• Replacement rate varies with word taboo, literacy and standardisation, and contact intensity, and basic vocabulary is demonstrably borrowable.
• Bayesian phylogenetic methods allow rates to vary, quantify uncertainty, and calibrate on attested languages, but remain sensitive to cognacy coding, calibration, and borrowing.
• Relative chronology is secure; absolute chronology rests on attested texts, loanword strata, terminus arguments, and correlations, in that order of confidence.

Sources

1. Wikipedia contributors. (n.d.). Glottochronology. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Swadesh list. en.wikipedia.org
3. Hammarstrom, H., Forkel, R., Haspelmath, M., & Bank, S. (n.d.). Glottolog. Max Planck Institute for Evolutionary Anthropology. glottolog.org
4. Swadesh, M. (1952). Lexico-statistic dating of prehistoric ethnic contacts. Proceedings of the American Philosophical Society, 96(4), 452-463.
5. Bergsland, K., & Vogt, H. (1962). On the validity of glottochronology. Current Anthropology, 3(2), 115-153.
6. Gray, R. D., & Atkinson, Q. D. (2003). Language-tree divergence times support the Anatolian theory of Indo-European origin. Nature, 426(6965), 435-439.

One hundred and fifty families reduced to three

In 1987 Joseph Greenberg published Language in the Americas. Specialists at that time recognised somewhere around 150 independent families and isolates in North, Central, and South America, the accumulated result of a century of detailed comparative work. Greenberg's book reduced them to three: Eskimo-Aleut, Na-Dene, and a single vast grouping he called Amerind containing everything else, from Cree to Quechua to Guarani.

The reaction from Americanists was close to unanimous rejection, and it was not a matter of temperament. Lyle Campbell's review in the journal Language in 1988 catalogued errors in the data itself: forms misattributed to the wrong language, morpheme boundaries drawn to make items match, glosses stretched, and material taken from unreliable older sources without checking. More fundamentally, the objection was about method, and that is what this lesson is about: what the comparative method can establish, where it stops, and what happens to claims made beyond that line.

Key idea: The disagreement about deep relationships is not about whether distant relationships exist. It is about what counts as evidence for one, and about whether a method that cannot fail is a method at all.

Why the signal decays

The comparative method depends on correspondence sets, and correspondence sets depend on cognates surviving in enough daughters to be seen. Two processes work against that as time passes.

Vocabulary turns over. Words are replaced by borrowing, by taboo, by coinage, and by ordinary competition, so the pool of inherited items shrinks steadily. Meanwhile sound change accumulates. Each change is regular, but ten or twenty of them stacked in each of two lineages can leave cognates with no surviving segment in common, which is why English cow and Sanskrit gaus are related and English much and Spanish mucho are not.

The chance resemblance rate, however, does not decay. Any two languages will always share some accidental look-alikes, because phoneme inventories are small and words are short. So the signal falls while the noise stays flat, and eventually they cross. Where that crossing happens is disputed, and estimates commonly fall somewhere between six and ten thousand years for well documented families. Beyond it, you cannot distinguish a real relationship from an accidental one, using vocabulary comparison alone.

Why this matters: The limit is not a lack of effort or imagination. It is the point at which inherited similarity has decayed below the level of similarity that chance produces for free.

How much chance actually gives you

It is worth making the noise concrete, because intuitions here are badly calibrated. Take two unrelated languages with roughly twenty consonants and five vowels each. Compare a list of a few hundred meanings, allow a match if the first consonant and the vowel agree, and allow meanings to count as similar rather than identical. The expected number of matches by chance is not small. It is comfortably into double figures, which is more than enough to fill a persuasive-looking table.

Don Ringe worked out the arithmetic formally in 1992, and the result is the reason that a list of look-alikes, however long, is not evidence. What defeats chance is not quantity of resemblance but structure: a correspondence that holds in a specified position across the vocabulary, and, best of all, matching irregular morphology. Recall the Indo-European verb to be from Lesson 7. The chance that two unrelated languages independently make their commonest verb irregular in the same way is negligible, and that is the kind of argument that survives.

The point: Long lists of similar-looking words are what chance produces anyway. Only recurrent correspondence in fixed positions, and shared arbitrary morphology, count as evidence.

Multilateral comparison, and the objection to it

Greenberg's method, which he called multilateral or mass comparison, is to inspect large numbers of languages at once, note resemblances in vocabulary and grammatical elements, and group languages by the impression the whole array creates. He argued that comparing many languages simultaneously makes chance resemblances stand out as random and real ones as patterned, and that requiring correspondence sets first is unnecessary bureaucracy.

The objections are specific.

It does not establish correspondences, so it cannot distinguish inheritance from chance in any individual case. It does not exclude borrowing, so shared items in a contact zone count as evidence of descent. It does not exclude nursery words or onomatopoeia. And its results are not reproducible in the way that a claim about a correspondence is: two analysts inspecting the same array can reach different impressions with no procedure for adjudicating between them.

The deepest objection is that the method has no criterion of failure. There is no observation that would show a proposed grouping to be wrong, because any list of languages will yield some resemblances. A procedure that cannot fail cannot confirm either.

In fairness, Greenberg's classification of African languages, published in 1963 and using the same approach, has held up substantially better, and its broad outlines, including Niger-Congo and Afro-Asiatic, are widely accepted today. Defenders take that as vindication of the method; critics reply that Africa's groupings have since been supported by conventional comparative work, that Nilo-Saharan, the weakest of his African families, remains the least accepted, and that a method producing one good result and one bad one is exactly what a method with no failure criterion would produce.

Remember: Multilateral comparison is rejected not because its conclusions are unwelcome but because it lacks any procedure for ruling a proposed grouping out.

Three proposals, graded

Nostratic is the most serious of the deep proposals. The name was coined by the Danish linguist Holger Pedersen in 1903, and the modern version was developed in the 1960s by Vladislav Illich-Svitych and Aharon Dolgopolsky in Moscow. It proposes a superfamily linking Indo-European, Uralic, Kartvelian, Dravidian, Afro-Asiatic, and the Altaic groupings, at a depth of perhaps twelve to fifteen thousand years.

What makes it methodologically respectable is that it does the right kind of work: it compares reconstructed proto-forms rather than modern words, and it states sound correspondences. The objections are correspondingly technical. Comparing reconstructions compounds the uncertainty in each one, so an error in Proto-Uralic propagates into every Nostratic etymology using it. The correspondences proposed are loose enough to admit a great many candidates. The semantic latitude allowed is wide. And different Nostraticists reconstruct materially different systems, which is not what you expect if the signal is real. Most historical linguists regard it as unproven; a minority regard it as promising; almost nobody treats it as established.

Amerind is the weakest, for the reasons above: a method with no failure criterion applied to data with documented errors, producing a grouping that would have to be far older than the comparative method can reach.

Dene-Yeniseian is the interesting case, because it shows what a good long-range proposal looks like. Edward Vajda's 2010 work argues that the Na-Dene languages of North America, including Navajo and Tlingit, are related to Yeniseian in central Siberia, of which Ket is the last surviving member. His argument does not rest on vocabulary lists. It rests on the verb: both groups build verbs from a template of prefix positions in the same order, with functionally matching elements in matching slots. That is shared arbitrary morphological structure, the strongest evidence type there is. The proposal is not universally accepted, and specialists have raised real objections, but it is taken seriously in a way that Amerind is not, and the reason is entirely about the kind of evidence offered.

The upshot: Long-range proposals are graded by evidence type, not by ambition. Shared morphological paradigms buy respect; lists of similar words do not.

What a serious proposal has to do, and what nobody claims

Five requirements, and they are the same ones that apply at shallow time depths.

• State recurrent sound correspondences, not resemblances, and show them holding across a body of vocabulary.
• Exclude borrowing, chance, and universal tendencies explicitly, as in Lesson 2.
• Prefer shared irregular morphology to lexical matches wherever it can be found.
• Publish the data in full, from reliable sources, so others can check the forms.
• Specify what evidence would count against the proposal.

At the far end of the scale sits the idea of Proto-World, a single ancestor of all human languages, and the global etymologies proposed by Merritt Ruhlen, of which the best known is a set of forms resembling tik and meaning finger or one. Almost no historical linguist accepts these. The point is not that a single ancestor is impossible; if language arose once, there was one, and that is a perfectly reasonable thing to believe. The point is that no evidence could survive that far, so the claim is untestable, and an untestable claim about the past is not a finding.

Which leads to the closing distinction, and it is the one the whole lesson exists to draw. Failing to demonstrate a relationship is not the same as demonstrating that none exists. Basque may well have relatives; the evidence is simply gone. Amerind may even be true. What the field is entitled to say is not that these languages are unrelated but that no one has shown them to be related, which is a different and more modest claim. Keeping those two apart is the whole discipline in one sentence.

Bottom line: The correct statement about a failed long-range proposal is that the relationship has not been demonstrated, not that it has been disproved.

Common misconceptions

• Rejecting Amerind means claiming Native American languages are unrelated to each other. It means the specific three-way grouping has not been demonstrated. Around 150 families are recognised, and further groupings among them are actively researched.
• Long-range proposals are rejected out of academic conservatism. Dene-Yeniseian is a long-range proposal taken seriously because of the kind of evidence it offers. The standard is about method, not distance.
• More data always helps a comparison. Adding languages and meanings raises the number of chance matches too. Without correspondence sets, a bigger dataset produces a more persuasive illusion.
• Nostratic and Amerind are equally weak. Nostratic compares reconstructions and states correspondences, which is the right kind of argument even if the execution is disputed. Amerind does neither.
• If all languages descend from one ancestor, deep reconstruction must be possible. Common descent and recoverable evidence are separate questions; the evidence decays whether or not the relationship existed.

Where this leaves us

• Greenberg's 1987 reduction of the Americas to three groupings was rejected on both data and method, most influentially in Campbell's 1988 review.
• The comparative method has a practical ceiling, often placed between six and ten thousand years, because inherited signal decays while chance resemblance does not.
• Chance produces a substantial number of look-alikes between any two languages, so lists of resemblances are not evidence; recurrent correspondences and shared irregular morphology are.
• Multilateral comparison is rejected chiefly because it specifies no observation that could show a grouping to be wrong.
• Nostratic argues in the right form and is unproven; Amerind is weak; Dene-Yeniseian is taken seriously because it rests on matching verb-prefix templates.
• A failed demonstration of relationship is not a demonstration of unrelatedness, and Proto-World is untestable rather than false.

Sources

1. Wikipedia contributors. (n.d.). Nostratic languages. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Amerind languages. en.wikipedia.org
3. Linguistic Society of America. (n.d.). What is linguistics? linguisticsociety.org
4. Campbell, L. (1988). Review of Language in the Americas by J. H. Greenberg. Language, 64(3), 591-615.
5. Ringe, D. (1992). On calculating the factor of chance in language comparison. Transactions of the American Philosophical Society, 82(1), 1-110.
6. Vajda, E. J. (2010). A Siberian link with Na-Dene languages. In J. Kari & B. A. Potter (Eds.), The Dene-Yeniseian connection (pp. 33-99). University of Alaska.

Anchorage, 21 January 2008

Marie Smith Jones died in Anchorage on 21 January 2008, at the age of eighty-nine. She was the last fluent speaker of Eyak, a language of the Copper River delta in southern Alaska. For the last decades of her life there was no one she could have an ordinary conversation with in her first language.

Eyak did not vanish without trace, because the linguist Michael Krauss had worked with her and with the handful of earlier speakers for decades, and what exists now is a dictionary, a grammar, and recordings. That is not nothing. It is also not a language, in the sense in which a language is something people use to argue and joke and raise children. Krauss had published a paper in the journal Language in 1992 that put the general problem in front of the field: on his estimate, the great majority of the world's languages were on a trajectory that would end them within a century.

This final lesson is about that trajectory, why it exists, what documentation can preserve, and what a community can do with what has been preserved.

Key idea: A language is lost when it stops being transmitted to children, and documentation preserves a record of it without preserving the thing itself.

The scale, in standard terms

UNESCO's classification, used in its atlas of endangered languages, has five levels, and they are worth knowing because they are the vocabulary everyone in this area uses.

Category	Definition
Vulnerable	Most children speak it, but use is restricted to certain domains such as the home
Definitely endangered	Children no longer learn it as a first language in the home
Severely endangered	Spoken by grandparents and older generations; parents may understand but not speak it to children
Critically endangered	The youngest speakers are grandparents or older, and they use it partially and infrequently
Extinct	No speakers remain

The single line that matters most in that table is the second. Once children stop acquiring a language at home, the outcome is essentially settled unless something is deliberately done, because the remaining speakers are a fixed and ageing population. Everything above that line is a warning; everything below it is a countdown.

The distribution of speakers explains why so many languages sit near that line. A small number of languages have enormous speaker populations, and roughly half of all languages have fewer than ten thousand speakers each. Small speaker numbers do not by themselves endanger a language, since a community that transmits its language to children can be tiny and stable, but small populations have little margin when economic and educational pressures arrive.

Why this matters: Endangerment is measured by intergenerational transmission, not by speaker numbers. A language with fifty thousand adult speakers and no child speakers is worse off than one with five hundred speakers of all ages.

How languages are actually lost

Language death is rarely an event. The usual pattern is a slow narrowing that takes three generations.

First comes domain loss: the language stops being used at work, then in school, then in public, retreating to the home. Then comes a generation of bilinguals who speak the ancestral language to their parents and the dominant language to each other. Then a generation that understands but does not speak, sometimes called passive bilinguals or semi-speakers. Then no one.

The pressures driving that sequence are mostly not mysterious. Economic opportunity attaches to the dominant language. Schooling is delivered in it. Media and now the internet operate in it. Migration and urbanisation break up the dense local networks that maintain a small language. Parents make individually reasonable decisions about their children's prospects, and the aggregate of those decisions is a shift no one chose.

But a large share of loss was not gradual and was not chosen at all. In the United States, Canada, and Australia, government and church-run boarding and residential schools removed Indigenous children from their families and punished them for speaking their languages, over roughly a century. Canada's Truth and Reconciliation Commission documented the residential school system and its effects in its 2015 report; Australia's Bringing Them Home report of 1997 documented the removal of Aboriginal and Torres Strait Islander children. In Wales, schoolchildren caught speaking Welsh were made to wear a token of shame. These policies were designed to end the transmission of languages, and in many cases they succeeded. Any account of endangerment that presents it purely as the natural outcome of modernisation is leaving out the part that was policy.

Remember: Shift proceeds through domain loss, then a bilingual generation, then passive understanding, then nothing, and a substantial share of the world's language loss was produced deliberately by schooling policies aimed at exactly that result.

Three arguments for doing something

The scientific argument. Every claim about what human language can be is constrained by the languages we have observed, and the sample is heavily skewed toward large, well-described, mostly European ones. Small languages routinely turn out to do things the sample said were unusual or impossible: elaborate obligatory evidential systems that require a speaker to mark how they know what they are asserting, case marking that stacks several layers on one noun, spatial systems that use fixed compass directions rather than left and right. Each language lost narrows what any theory of language can be tested against.

The knowledge argument. Languages carry detailed knowledge about local ecologies, plants, weather, navigation, and medicine, encoded in vocabulary and in the categories the grammar makes obligatory, and much of it has never been written in any other language. When the language goes, that knowledge usually goes with it.

The rights argument. This is the one the field increasingly treats as primary. Communities have a claim to their own languages, recognised in instruments including the United Nations Declaration on the Rights of Indigenous Peoples, adopted in 2007, which addresses the right to use, develop, and transmit languages to future generations. Loss of a language is frequently experienced by its community as a bereavement and as one consequence among many of a history of coercion.

One honest note about the first argument. A linguist's interest in an unusual case system is not a reason for anyone to keep speaking a language, and framing endangerment primarily as a loss to science gets the priorities backwards. The decisions belong to the community. What outside linguists can offer is technical help, archiving, training, and materials, on terms the community sets.

The point: The scientific and knowledge arguments are real, and the decision about a community's language belongs to that community, which is why current practice is built around consent and collaboration rather than collection.

Documentation is not description

The distinction, drawn sharply by Nikolaus Himmelmann in 1998, has reshaped fieldwork.

Description produces an analysis: a grammar, a dictionary, an account of the phonology. It is the linguist's interpretation, organised by linguistic categories, and it answers the questions the linguist thought to ask.

Documentation produces a record: a large, durable, annotated corpus of actual language use, in many genres and by many speakers, with recordings, transcriptions, translations, and metadata, deposited in an archive so that people who ask different questions later can still get answers. Narratives, conversations, procedural texts about how something is made, songs, arguments, children speaking.

The reason the shift matters is that a grammar written in 1950 answers 1950's questions. A recorded and annotated corpus can be reanalysed indefinitely, and, crucially, it can be used by the community, which a technical grammar generally cannot. Modern practice therefore emphasises archiving in institutions built for the purpose, clear agreements about who may access what, and materials produced in forms the community can actually use: storybooks, dictionaries for learners, curriculum.

The upshot: A description answers the questions the analyst asked; a documentary corpus preserves the material for questions nobody has asked yet, including the community's own.

What revitalisation looks like when it works

Revitalisation is difficult and its results are uneven, but the record is not empty, and the approaches that have worked share a common feature: they create situations in which the language is used, not taught about.

Immersion for the very young. The Maori kohanga reo, or language nests, began in New Zealand in 1982, putting preschool children with fluent elders in an all-Maori environment, with a school system following as those children aged. Hawaiian followed with Punana Leo from 1983, and Hawaiian-medium education now runs through to university level. Both languages had been in serious decline; both now have thousands of young speakers who acquired them in childhood.

Master and apprentice pairing. Developed for California's languages, where many had only a few elderly speakers and no realistic prospect of a school programme, this pairs one fluent elder with one committed adult learner for many hours a week of shared activity conducted entirely in the language, with no translation. It has produced new speakers of languages that had none under fifty.

Reclamation from the archive. The Wampanoag language of southeastern Massachusetts had no speakers for more than a century. It did have an unusually rich written record, because seventeenth century missionary work had produced a large body of text, including a complete Bible translation, and because Wampanoag people themselves had written letters, deeds, and petitions in their own language. Beginning in 1993, Jessie Little Doe Baird and the Wopanaak Language Reclamation Project used those documents, together with comparative evidence from related Algonquian languages, to reconstruct the language and bring it back into use. Baird's daughter was raised speaking it, the first first-language speaker in generations.

The largest case, with its caveat. Hebrew is the only language usually described as having gone from no native speakers to a national language with millions, a transformation associated with Eliezer Ben-Yehuda and the settlement of Palestine from the 1880s. It is a genuine achievement and a poor template, because the conditions were exceptional: continuous liturgical and literary use for two millennia so the language was never truly dormant, a large body of text, a population with many members already literate in it, and eventually a state with the power to make it the language of schooling and government. Very few endangered languages have any of that.

Worth holding on to: What revitalisation programmes have in common is the creation of contexts where the language is the medium rather than the subject, whether that context is a preschool, a pair of people cooking together, or a household.

Common misconceptions

• Languages die naturally, like species. Much of the loss of the last two centuries followed from specific policies, including schooling systems explicitly designed to stop transmission.
• A documented language is a saved language. An archive preserves a record. Only children acquiring the language preserve the language, which is why the transmission line in the UNESCO table is the one that matters.
• Small languages are simple or limited. Small languages routinely have grammatical machinery, such as obligatory evidential marking, that the largest languages lack entirely.
• If Hebrew was revived, any language can be. Hebrew had continuous textual and liturgical use, mass literacy in it, and eventually state backing. Its conditions are close to unique.
• Bilingualism causes language loss. Stable bilingualism is normal worldwide and can persist for centuries. Loss occurs when one language stops being transmitted, not when a second is added.

What you now know

• Eyak ended as a spoken language with the death of Marie Smith Jones in 2008, leaving a dictionary, grammar, and recordings compiled over decades.
• UNESCO's categories run from vulnerable to extinct, and the decisive threshold is whether children still acquire the language at home.
• Loss typically proceeds through domain narrowing, a bilingual generation, and passive understanding, and a large share of it followed from residential and boarding school policies designed to end transmission.
• The scientific, knowledge, and rights arguments all support action, and the rights of the community rather than the interests of researchers govern the decisions.
• Documentation builds a durable annotated corpus for questions not yet asked; description produces the analyst's grammar and dictionary. Both are needed and they are not the same.
• Immersion nests, master-apprentice pairing, and archival reclamation have all produced new speakers, and Hebrew's revival relied on conditions almost no other language has.

Sources

1. Linguistic Society of America. (n.d.). Why do languages die? linguisticsociety.org
2. UNESCO. (n.d.). Languages. unesco.org
3. Endangered Languages Project. (n.d.). Catalogue of endangered languages. endangeredlanguages.com
4. Krauss, M. (1992). The world's languages in crisis. Language, 68(1), 4-10.
5. Hinton, L., Huss, L., & Roche, G. (Eds.). (2018). The Routledge handbook of language revitalization. Routledge.