Is There Time Since Babel for the Language Families?

Summary

If each language family began at Babel, the diversity inside it grew in about 4,250 years on the Masoretic text, or 4,770 on the Septuagint.  This article measures how fast languages change and asks whether that is enough.  Under the Masoretic date it is not:  the Bronze Age Indo-European texts need language to have changed two to five times faster than anything recorded.  Under the Septuagint date, with the small founder groups the young-earth literature already posits, the gap nearly closes.  The parent article, Language Families and the Tower of Babel, states the young-earth model;  its other children ask whether language began once and how old the families are.

How fast languages change.  Documented lineages keep 70 to 85% of their basic vocabulary per millennium.  Robert Lees found 81% on thirteen dated pairs in 1953, with no pair faster than 76%.¹¹  Bergsland and Vogt showed in 1962 that the rate is not constant:  Icelandic kept 97 to 100% over about eight centuries, while East Greenlandic, spoken by about 400 people with a name taboo, had a retention estimate of 72%.²²  Words differ in rate, so a short calibration overstates the rate over a long interval.³  Small populations lose words faster;  Bromham and colleagues measured a slope worth about a factor of 2.5 over Polynesian populations from 200 to 360,000 speakers.⁴⁵  In 1,000 to 2,200 years Latin, Old Norse, Koine Greek and Old Chinese each became a group of related languages, and none became an unrelated family.  Creoles keep most of their source's basic vocabulary.⁶

The divergence-time test.  The test counts shared basic-vocabulary cognates between Indo-European languages in the IE-CoR database, calibrated on 59 ancestor-to-descendant pairs dated by texts within the last 2,450 years.⁷  The code reproduces 11 of 13 percentages that Dyen, Kruskal and Black printed in 1992.⁸  Hittite, written about 1450 BC, and Vedic Sanskrit, about 1300 BC, share 21% of their meanings, about as many as Spanish and Russian do today.⁹  Under a Babel at 2242 BC every lineage would have had to change 2.45 times as fast as the fastest lineage recorded over a millennium, and 3.5 to 4.8 times as fast once the measured variation in rate among meanings is allowed for.  Under the Septuagint's 2767 BC the factors are 1.53 and 2.2 to 3.0.¹⁰  John Oller's claim that observed rates suffice within 4,000 years fails for every Bronze Age pair that crosses branches of the family.¹¹

The escape routes.  Only two rest on young-earth writers' own proposals.  K. J. Duursma suggests faster word loss in small, isolated groups.¹²  Computed here from published young-earth founder sizes and Bromham's slope, it halves the Masoretic gap and closes the Septuagint gap at a single rate.¹³  Anne Habermehl lowers the Late Bronze Age by about 550 years, though she does not apply it to language;  with both routes, the Septuagint case fits even with rate variation.¹⁴  The Masoretic date then fits at a single rate, and with rate variation only if every favourable choice is stacked at once.  Nobody proposes Babel languages that were already partly related, or creolisation, which fails on the creole data.  Henry Morris allowed gaps in Genesis 11 of a thousand years or more;  no young-earth writer applies them to language.  A thousand years brings Hittite and Vedic within 7% of the fastest recorded rate, and only the far end of his range, five thousand years, reaches the text-dated ages with no change in rates.¹⁵¹⁶  Austronesian, with no Bronze Age texts, cannot exclude a post-Babel origin on modern word lists.¹⁷

Verdict.  Languages diversify fast, but basic vocabulary decays slowly, and the Bronze Age texts are too far apart for a Masoretic Babel at any recorded rate.  The Septuagint chronology with small founder groups comes within reach of the data.  The honest young-earth position takes the longer chronology and says plainly that Indo-European needs a burst of early change that no document records.

Part 1: How fast languages change

Everything in How Old Are the Language Families? and in Part 2 below turns on one quantity:  how much of a language's basic vocabulary survives per thousand years.  It has been measured since 1953, on documented ancestor-to-descendant pairs, and the measurements are the young-earth side's key unknown stated as a number.

Lees, 1953.  Robert Lees treated basic-vocabulary loss as first-order decay and calibrated it on thirteen pairs counted on a 215-word list, every pair dated by documents:  Old English to English, Plautine Latin to Spanish and French, Old High German to German, Middle Egyptian to Coptic, Koine to Athenian and Cypriot Greek, Classical Chinese to Mandarin, Old Norse to Swedish, Latin to Italian, Portuguese, Rumanian and Catalan.¹  The result:  "k = 0.8048 ± 0.0176 per millennium, with a confidence limit of 90%", or "on the average about 81% of the basic-root-morphemes of a language will survive as cognates after 1000 years".¹¹  The values "range from 0.760 for Coptic : Egyptian up to 0.854 for both German : Old High German and Swedish : Old Norse, with a standard-deviation of 0.0342."¹  Nothing in his data is faster than 0.760, and the fastest pair is the one spanning the longest time.  He conceded the weakness himself:  constancy over time could only be tested on "Assyro-Babylonian, which covers about 5000 years of written records", and he had not done it;¹ applied to German and English his constant dates their split to "about 716 A.D.", though "the Germanic invasions of Britain began about 449";¹ and if word classes decay at different rates "the assumption of a temporally constant k for our word list ... is clearly false."¹

Swadesh, 1955.  Morris Swadesh replaced the long list with a 100-item one, because "If the lexicostatistic test list is to be valid for different magnitudes of time depth, it has to be made up of elements having approximately the same likelihood to persist", and otherwise "there will be a tendency for the less persistent items to be lost early, thereby leaving a residual list of gradually increasing durability".¹⁸  On the new list seven of Lees's pairs keep about 86% per millennium in place of 81%.¹⁸  He dropped Coptic, the fastest pair, as showing "hidden divergence" already present in 1700 BC.¹⁸  Two things in this paper matter later.  Swadesh's reason for the revision is the rate-variation point:  the observed rate rises with depth because the fast items go first, which is what makes a constant-rate date a floor.  And Arndt's independent rescoring lowered Lees's English rate to "67.8% or 74.8% instead of 76.6%", the lower figure before removing inter-dialect borrowing, so English, flooded with French and Norse words, is among the fastest documented lineages.¹⁸

Bergsland and Vogt, 1962.  The paper that ended glottochronology's reputation tested it on four documented lineages.  "For Old Norse of the early 13th century to Modern Icelandic (rural), the retention rates are, in the 215-item test list, 97.19%; in the 200-item list 97.9%; and in the 100-item list 100%."²  So the method dates eight centuries of Icelandic history to zero:  "The ridiculously recent dates obtained by the glottochronological method cannot be due to mutual influence in later periods or to convergence ... but are simply due to the high retention rate of Modern Icelandic."²  Georgian, from the fifth-century Gospel translations, keeps 87 to 89% per millennium on the 215-item list and 95 to 97.2% on the 100-item list;  Armenian over the same span 89.7 to 94.1% and 94.6 to 97.8%.²²  At the other extreme, East Greenlandic, spoken by about 400 people at Angmagsalik with a name taboo that replaced the word for anything a dead person had been named after, "has deviated from West Greenlandic by 30% more than has the dialect of Labrador" inside a separation of a few centuries, with a retention estimate of 72%.²²  Their conclusion:  "Our findings clearly disprove the basic assumption of glottochronology 'that fundamental vocabulary changes at a constant rate.'"²  And the mechanism, which both sides should note:  "If relative isolation and the small size of the tribe accounts for the low retention rate of East Greenlandic, the same circumstances may account for the high retention rate of Modern Icelandic."²  Small and isolated goes both ways.

Rate variation among words.  Pagel, Atkinson and Meade estimated a separate replacement rate for each of 200 meanings across 87 Indo-European languages and found that how often a word is used predicts about half the variation.¹⁹  Restated with numbers in 2013:  "most words have about a 50% chance of being replaced by a new noncognate word ... roughly every 2,000–4,000 y ... However, some words, such as the numerals, pronouns, and special adverbs (e.g., I, you, here, how, not, there, what, two, five) are replaced far more slowly, with half-lives of once every 10,000, 20,000 or even more years".³  Greenhill and colleagues sorted 1,195 Austronesian cognate sets and 157 grammatical features into rate classes and found basic vocabulary the slowest part of language, with 82% of its cognate sets in the slow class and grammar changing about five times faster on the median.²⁰²⁰  So there is no faster subsystem to which a young chronology can appeal, and no slower one that keeps a deep signal.  The slow words cut against one young-earth reading in particular:  if Babel split one language into daughters, the pronouns and numerals with 10,000-year half-lives should still be cognate across every family born there.  They are not.  That fits Babel only on Wieland's and Hodge's model, in which Babel created unrelated vocabularies outright, and not on Taylor's, in which one vocabulary was re-sounded.

Bursts at splits.  Atkinson, Meade, Venditti, Greenhill and Pagel found that path length on lexical trees grows with the number of splits along the path:  "punctuational bursts of change at the time of language splitting are an important and general process in language evolution and account for 10 to 33% of the total divergence among these languages in their fundamental vocabularies", with 31% for Bantu, 21% for Indo-European and 9.5% for Austronesian.²¹²¹  Greenhill 2017 puts the Austronesian lexical figure at 29%.²⁰  This helps a young chronology in kind:  a dispersal is a burst of splitting, and the burst component is tied to the number of splits rather than to elapsed time.  It limits the help in size:  if 21% of Indo-European divergence is per split, the time-dependent part that must be compressed falls by at most a fifth.

Population size.  Bromham, Hua, Fitzpatrick and Greenhill measured, in ten pairs of Polynesian sister languages, the rate of gaining new basic words and the rate of losing inherited ones against speaker population:  "larger populations have higher rates of gain of new words whereas smaller populations have higher rates of word loss."⁴  The loss slope is -0.12 on log-log scales, with a standard error of 0.032, over populations from 200 to more than 360,000.⁴  Loss is what lowers the shared-cognate fraction, and over the study's own range that slope is a factor of about 2.5 in loss rate.⁵  The limits are the paper's own:  "overall rates of change (gain plus loss) were not significantly related to population size";  "Allowing for a period of population growth does not provide a better fit to the data, nor does accounting for founder effects";  and the reduced datasets "did not have a significantly higher rate of loss in smaller populations".⁴⁴⁴  Daniel Nettle's 1999 simulation points the same way:  "the rate is much faster with communities of less than 200-250 members. Thereafter there is no further decline in the rate".²²  It is a model, not data, and he says so:  "it is quite possible that there are other sets of computational procedures and assumptions which would produce different results."²²  His strongest datum is second-hand:  Lithgow found in Muyuw, with three thousand speakers, "a massive rate of lexical replacement (13%, which should take 685 years according to the glottochronological clock of Lees, 1953) and grammatical change over just 50 years".²²  Taylor in 1983 took the mainstream view to be the reverse, writing that rapid change "goes so very much against the popular language theory that small societies change their speech form more slowly than large ones", and so treated fast change in small groups as something the other side could not use;²³ the later literature finds small groups lose words faster, and the young-earth side has not yet used that.

What two thousand years do:  the historical cases.  The documented cases of one language becoming many are the best control the young-earth side has, and they all point the same way.

The Romance case is the standard.  Coleman:  "by the ninth century AD written Latin and the diverse spoken forms of Latin had ceased to be registers of one language ... but were now quite separate, if closely related, languages. Caesar and Livy would have recognised the Latin of Nithard's Historiae as a form of their own language; they would have found his citations of the Strassburg oaths as baffling in the romana lingua as in the teudisca."²⁴  Jerry Norman's Chinese is the best non-Indo-European parallel:  "The Chinese dialectal complex is in many ways analogous to the Romance language family in Europe: both have their roots in a large-scale imperial expansion that took place in the centuries just preceding and just following the birth of Christ ... there is probably as much difference between the dialects of Peking and Chaozhou as there is between Italian and French; the Hainan Min dialects are as different from the Xian dialect as Spanish is from Rumanian."²⁵  Alan Kaye on Arabic:  "the differences between many colloquials and the classical language are so great that a fallah who had never been to school could hardly understand more than a few scattered words and expressions in it", and Maltese "developed into a new Semitic language in its own right".²⁴²⁴  Against these, Adelaar's Malagasy:  about 1,300 years on a large island with heavy Bantu contact, and the dialects "are also relatively homogeneous", still recognisably Southeast Barito beside Ma'anyan in Borneo.²⁶²⁶

Reading the table:  in 1,000 to 2,200 years one attested language becomes a group of mutually unintelligible but plainly related languages, keeping 70 to 85% of its basic vocabulary per lineage per millennium, 100% in Icelandic and about 70% in English and Cypriot Greek.  The fast cases have causes:  heavy borrowing in English, creolisation in Tok Pisin, taboo in a group of 400 in East Greenland.  None of the eight produced an unrelated family, and in every one the descendants are recognisable as a family from the basic vocabulary alone.  Oller is right that dialects "can become mutually unintelligible within about 1,000 years";¹¹ Hammarström says the same, "approximately 1,000 years for it to reach unintelligibility with its former self".²⁷  But unintelligibility is not the measure that limits the family count.  Cognate retention is, and that decays far more slowly.  A Masoretic Babel lies about 4,250 years before AD 2000 and a Septuagint Babel about 4,770, two to four times the depths in the table.  At the fastest millennium-scale rate recorded, 0.694, two lineages separated for 4,250 years share about 4.5% of a basic list, near the 7% chance floor;  at the median rate, 0.779, they share about 12%, which is detectable.²⁸  So the question for the young-earth side is not whether languages diversify fast.  They do.  It is whether two to four times the Romance depth can turn one language into families that share nothing, and on these rates it cannot.

Creoles and mixed languages.  A creole is a new language made in a generation, and the young-earth side can fairly cite that for speed.  What it does to basic vocabulary is the opposite of what a Babel-by-creolisation story would need.  Thomason and Kaufman:  "Most languages of all three nongenetic types will derive most of their basic vocabulary from a single source."⁶  Ma'a "has perhaps 50 percent Cushitic vocabulary (including most of the basic vocabulary) and a few Cushitic structural features, while its remaining vocabulary, together with almost all its grammar, is of Bantu origin".⁶  Crowley's datum for Tok Pisin:  "out of 200 items, the two languages have common sources for 146 items", after about a century;²⁹ that is 27% of the list replaced in 130 years, six times what the clock allows, and 73% still plainly English.  Dyen, Kruskal and Black make the same point from the other side, counting those shared words as "pseudo-cognates" because the creole was "mutually unintelligible with English at the moment that its first native speaker came into existence", and noting that "There is no evidence for creole languages prior to the post-Columbian colonial period".³⁰³⁰³⁰  Thomason and Kaufman's warning is the one that matters for Part 2:  "after a few hundred years with no documentation, a language like this might look like a distant relative of its lexical source language".⁶  Creolisation fast-forwards a language to the status of a close or distant relative.  No documented process turns one language into several whose basic vocabularies share nothing, and no young-earth writer found has proposed one.

Part 2: The divergence-time test

The Bayesian dates of How Old Are the Language Families? come from models with dozens of parameters, and a skeptic of either side is entitled to distrust them.  So this article asks a simpler question with a method a reader can check.  Given only calibrations a young-earth chronologist accepts, how much time does the divergence inside Indo-European require, and does it fit the time since a Babel date?  The code is Deno TypeScript in research/code/ beside the parent article, its outputs are in research/data/, and every number below is read from those outputs.

What was measured.  For two languages, the share of meanings for which they have cognate words, the "lexicostatistical percentage" of Dyen, Kruskal and Black.  The data are IE-CoR, the Indo-European Cognate Relationships database behind Heggarty and colleagues' 2023 paper:  160 languages, 170 meanings, 25,741 cognate judgements made by specialists, with loanwords flagged by cognate set and 52 ancient languages dated by their texts.³¹  For Austronesian, the ABVD database of Greenhill, Blust and Gray, 2,040 word lists over 210 meanings.³²  The rules, applied identically to calibration and inference:  a meaning is shared if any form of one language shares a non-loan cognate set with any form of the other, which is Dyen's rule;  a loanword stays as a present word that matches nothing, so borrowing counts as replacement of the native word, as in Dyen's coding;  and a sensitivity run drops loans as missing data instead.³³  Uncertainty is a bootstrap of 1,000 resamples over the meanings both languages have.

Validation.  Before the code touched IE-CoR it was run on Dyen's raw 1997 data file and asked to reproduce the thirteen percentages Dyen, Kruskal and Black printed in 1992, such as "the smallest percentage between any two Baltoslavic lists is 30.6, between Latvian and Bulgarian" and "Macedonian's highest percentage is 83.5 with Bulgarian".³⁰³⁰  Eleven of thirteen match to the printed decimal and two miss by a little, 17.6 against 17.5 and 91.5 against 91.0, the second on a file that postdates the book.⁸  Then the input was deliberately broken:  30% of forms reassigned at random to unique classes, which should cut every shared fraction to about half.  Macedonian and Bulgarian fell from 83.5 to 42.1, Icelandic and Faroese from 92.2 to 49.0.  The check can fail, and fails by the predicted amount.⁸

Calibration.  The rate was learned from 59 strict ancestor-to-descendant pairs in IE-CoR:  Latin to each Romance language, Old English to English, Old Icelandic to Icelandic, New Testament Koine to the modern Greek dialects, Old Irish to Irish, Scottish Gaelic and Manx, Old Church Slavonic to Bulgarian and Macedonian, Classical to modern Armenian, Old to Middle to New Persian, Old to Middle to modern Welsh and Breton, and others.  The old end of every pair is dated by its texts:  Cicero's Latin 50 BC, Ælfric's Old English about AD 950, the Old Icelandic of about 1200, and so on.  Every calibration lies within the last 2,450 years and none uses radiocarbon.⁷  Retention per lineage per millennium, r:

With loans treated as missing rather than as replacements every rate slows:  fastest 0.726, median 0.821, slowest 0.919.  The fastest short-interval rates all come from borrowing:  dying Cornish flooded with English, Breton with French loans, Arbëresh with Italian.⁷  The median, 0.779, sits where Lees's 0.805 on a different list would lead one to expect;  IE-CoR was designed to include less stable meanings, so its rates must be calibrated on itself, as they are here.  Sister splits with historically dated separations give slower figures still:  European and Brazilian Portuguese 0.900, Icelandic and Nynorsk 0.888, Icelandic and Faroese 0.924, Cappadocian and standard Greek 0.843.⁷

Is the fastest rate realistic for a whole lineage?  No, and the dataset itself shows it.  Applied to Mycenaean Greek of 1350 BC and Attic of 400 BC, whose common ancestor must be older than the Linear B tablets, r = 0.694 puts that ancestor at 1062 BC, with a bootstrap range of 935 to 1203 BC, three centuries after one of the two was written down.³⁴  The fastest rate gives ages that are too young even where history can check them.  It is a floor, not an estimate, and it is used below as the most generous rate a young chronology can claim.

Inference.  For two languages attested at dates dX and dY with shared fraction S, the two lineages together run L = ln S / ln r millennia, and the common ancestor sits at T = (dX + dY - 1000 L) / 2.³⁵  The attestation dates are IE-CoR's own tip dates:  Hittite 1450 BC, the generous young end of the Hattusa archive;  Luvian and Mycenaean Greek 1350 BC;  Early Vedic 1300 BC;  Younger Avestan 550 BC;  Attic 400 BC;  Latin 50 BC;  Gothic AD 370;  Tocharian B AD 650;  Old Prussian AD 1500;  modern lists AD 2000.  Common-ancestor dates at four rates, with the bootstrap interval at the fastest millennium rate:

The Proto-Indo-European root is the split of Anatolian from the rest.  At the fastest millennium rate applied to every lineage for its whole history, the root pairs put it between about 2450 BC and 3500 BC, with the earliest-attested pairs giving the oldest dates because they carry the least post-attestation change.  At the median rate the root is 3600 to 4900 BC.  The Bayesian analysis of the same dataset, 8,120 years before AD 2000 or about 6100 BC, lies between the median and slowest columns.  Setting loans aside as missing barely moves it:  Hittite and Vedic at the loans-missing fastest rate give 3777 BC.⁹

What the Bronze Age pairs say.  Hittite and Vedic Sanskrit, written down 800 to 950 years after a Masoretic Babel, share 21% of IE-CoR's meanings.  That is about as many as modern Spanish and Russian share, 21.8%, or English and Latin, 21.8%, and fewer than Attic Greek and Latin, 26.5%, or Lithuanian and Russian, 35.3%.⁹  Two languages that far apart cannot have been one language a few centuries earlier at any recorded rate.

Rate variation makes the constant-rate figure a floor.  Dyen, Kruskal and Black said it in words:  "different meanings are subject to replacement of their words at different rates ... so that the usual glottochronological rule for deducing age from lexicostatistical percentage is not valid. Instead, similarity decays in the manner of radioactivity from a mixture of numerous radioelements".³⁰  The direction of the error is provable.  If each meaning is replaced at its own rate, the effective rate over a long interval can only fall with depth, because the fast meanings are gone early and the slow ones remain;  so a rate calibrated over 2,450 years is at least as fast as the true rate over 5,000, and inverting a deep S with it gives a time no longer than the true time.³⁶  On the dataset, a gamma mixture of rates beats a single rate by a likelihood-ratio statistic of 208 on 23 disjoint lineage segments and 2,222 on all 59 pairs, each on one degree of freedom, and the per-meaning rates run from zero, for 19 meanings never replaced in any dated segment, to 1.45 per thousand years.³⁷  At the Hittite-Vedic level of similarity the correction multiplies the required time by 1.65 on the disjoint fit and 2.42 on the full fit, which moves the Hittite-Vedic ancestor at the fastest rate from 3500 BC to about 4880 BC.  The gamma extrapolation rests on the assumed shape of the slow tail, so the corrected figures are estimates;  the uncorrected ones are floors.

The Babel windows.  Babel is placed at Peleg's birth, the earliest moment "in his days" and so the most time each text allows:  Masoretic 2242 BC, Ussher's dispersion as Answers in Genesis prints it, or 2417 BC with Abraham born in 2166 BC;  Samaritan 2707 BC;  Septuagint 2767 BC.  The last three are from the table in the Old Testament text and chronology, and the choice among them belongs to the Tower of Babel.  If a pair's family began at Babel, its two lineages have between them the years from Babel to each attestation.  The speed-up k is how many times faster than the stated rate every lineage would have had to change, continuously, to produce the observed S in that time.  Below 1 fits.

"Cornish" is r = 0.556, the rate of a dying language over 400 years;  "fastest 1 ka" is 0.694;  "gamma" is the disjoint gamma fit scaled to the fastest millennium calibration.  The full-fit gamma gives larger values, 4.80 for Hittite and Vedic under the Masoretic date.¹⁰  Under a Masoretic Babel, Hittite and Vedic need every lineage to change one and a half times as fast as dying Cornish, 2.45 times as fast as the fastest lineage recorded over a millennium, and 3.5 to 4.8 times as fast once rate variation among meanings is allowed for.  Under the Septuagint, 0.95, 1.53 and 2.2 to 3.0.  Oller's claim that observed rates "could easily have been produced within a space of 4,000 years" is the base claim this table tests.  Under the Masoretic date it fails before rate variation by a factor of 1.8 to 2.5 for Hittite against Vedic and Mycenaean, and 1.4 for Vedic against Mycenaean;  only the two Anatolian languages, Hittite and Luvian, fit.¹¹

The escape routes, and who proposes them.  A young chronology can be rescued in five ways.  The corrections log of this project requires that each be labelled by whether any young-earth writer proposes it, and the audit found that most do not.

Route (a), faster change in small post-Babel populations.  Duursma proposes it, in two sentences and without numbers:  isolation "could result in greater word loss" and "some languages have a much quicker word loss, for cultural reasons."¹²  The founder sizes are published young-earth positions:  Osgood's "a total of around 25,000 human beings" at Babel, and Carter and Hardy's model giving "greater than 1,000 individuals" a century after the Flood, and "up to 40,000", only at early ages of first childbearing that they call "not likely scenarios".³⁸³⁹  The number of Babel languages is also published:  Duursma's "about a dozen or so", Hodge's 78 at least.  Their combination with Bromham's measured slope and the IE-CoR rates is this article's construction, and it is labelled so.  Founders per lineage are the people divided by the languages, halved because splitting Proto-Indo-European in two at Babel itself gives the root pair the most time:  6 for Carter and Hardy's 1,000 with Hodge's 78;  160 for Osgood's 25,000 with Hodge's 78;  1,042 for Osgood with Duursma's 12.  Each lineage then grows exponentially to the population of its attested community, and its loss rate at every moment is scaled by Bromham's slope relative to the calibration lineages, capped at the smallest population Bromham measured so that nothing is extrapolated below 200 speakers.¹³  The residual speed-up still needed, for Hittite and Vedic:

Four things the population computation shows.  It is the strongest escape route:  with Osgood's and Hodge's numbers the mean multiplier over the interval is 1.5 to 2.0, which halves the Masoretic gap at the fastest rate and closes the single-rate gap under the Septuagint.  Adding Atkinson's bursts, which cut the time-dependent part by at most a fifth, brings the central Septuagint case to 0.69 at one rate and 0.96 with the gamma correction.¹³  It does not close the Masoretic gamma-corrected gap, which stays at 1.7 to 2.3 in the central case and 1.6 with bursts, and what it does close rests on stacked favourable choices:  the fastest recorded lineage as the baseline with a population multiplier on top of it;  a slope that was significant for only one of Bromham's three population measures and for loss but not total change;  and both lineages splitting at Babel itself.  The effect is front-loaded and then gone:  almost all the extra loss happens while a lineage is small, and once it reaches tens of thousands the multiplier is near 1.  The Hittite kingdom, the Mitanni state and the Mycenaean palaces were large for centuries before their texts, which is why the one variant that clears the Masoretic gamma gap, holding every lineage at founder size for the whole interval, is not a real option.  Across every scale assumption tried, with founders of 160, the Masoretic date and Bromham's measured slope, the residual runs 1.09 to 1.83 at one rate and 1.54 to 2.58 with gamma.  At the upper end of the slope's 95% interval it runs 0.70 to 1.51 and 0.99 to 2.13, so the Masoretic single-rate gap can close if the slope is taken at its upper bound and the population scales at their most favourable.¹³

Route (b), Babel created internally diverse proto-languages.  No young-earth writer proposes this, and the published model says the opposite:  Wieland's one stem per family, Oller's "radically distinct" languages, Hodge's Genesis 10 tally of families.  Every published model fixes the similarity of two lineages at Babel at 100%.  It is this article's hypothetical, and the computation is better read as a measure of how far the data sit from the published model.  The S0 column asks what share of basic vocabulary two lineages must already have shared on leaving Babel for the fastest millennium rate to produce today's S.⁴⁰

Under a Masoretic Babel the ancestors of Hittite and of Vedic would have had to leave Babel sharing only about 40% of their basic vocabulary, about as different as Lithuanian and Russian are today, and still be related by regular sound correspondences.  Under the Septuagint, about 60%.  A model that allowed this could absorb any S and would not be testable by lexicostatistics;  it would also have to say why created languages show the nested structure of descent.  Nobody has proposed it.

Route (c), the attestation dates moved later.  The chronology exists;  the application does not.  Anne Habermehl's 2023 Answers Research Journal synchronisation sets Solomon and Thutmose III at about 1000 BC against the standard 1470 BC, and the Bronze Age collapse at about 600 BC against the twelfth century, a shift of about 470 to 550 years.¹⁴  The Hittite archive, the Linear B tablets and the Mitanni treaty that anchors early Vedic all sit in the Late Bronze Age and are dated through Egyptian and Hittite synchronisms, so her scheme would move them by about the same amount.  She does not apply it to language;  the application is this article's.  With every date before 700 BC moved 550 years later, Hittite at 900 BC and Vedic at 750 BC:

The 300-year row is the scale of the non-creationist revisionists James and Rohl, and no young-earth writer proposes it either.⁴¹  Lowering the Late Bronze Age helps a young chronology because it lengthens the time between Babel and the first texts, but it does not close the gap at the fastest millennium rate under the Masoretic date, and it has costs this article does not price:  a Rigveda at 750 BC collides with the Vedic literature's own internal sequence down to the early Buddhist texts, and the Assyrian eclipse and tree-ring anchors of Mesopotamian absolute chronology must also be overcome.  The opposite sensitivity was run too:  IE-CoR dates Hittite at 1450 BC because most of its list comes from the empire-period archive, but the Old Hittite texts are of the seventeenth and sixteenth centuries;  at 1650 BC the Masoretic speed-up rises from 2.45 to 2.77.⁴²

Route (d), creolisation or mixed languages.  No young-earth writer mentions creoles, pidgins or mixed languages in connection with Babel.  Part 1 prices it anyway:  creoles keep most of the lexifier's basic vocabulary and mixed languages keep one source's, so the route cannot make unrelated families from one language.  It is this article's hypothetical, and it fails.

A gap reading of Genesis 11.  Morris allowed that gaps might raise his dates by "up to about one thousand years or more (possibly five thousand years at the very outside)".¹⁵  No young-earth writer applies that to language, and the later literature found here does not take it.  Had a thousand years been taken, a Babel at 3358 BC would leave Hittite and Vedic needing 1.07 times the fastest millennium rate, and 1.5 to 2.1 times with rate variation:  below every Septuagint figure, but still not a fit.¹⁶  Only the far end of Morris's range reaches the text-dated family ages with no change to the rates:  at 7358 BC the factors are 0.36 at one rate and 0.5 to 0.7 with rate variation.  So this is the one young-earth-compatible reading that can reach those ages without faster change, and it is proposed in principle by the founder of the movement.

Austronesian.  The same test on ABVD gives the young-earth side a real point.  The calibration is the East Polynesian settlement, on radiocarbon dates after AD 1000 that no young-earth model disputes:  Wilmshurst and colleagues' 1,434 dates put the Society Islands at AD 1025-1120 and the remaining islands at AD 1190-1290.⁴³  Twenty-eight East Polynesian sister pairs give rates from 0.673 per millennium for Tahitian and Marquesan, the fastest, to a median of 0.795, with Tahitian fast as expected for a language with recorded word-taboo replacement.¹⁷  The root is the split of Atayalic from Malayo-Polynesian, both primary branches of Austronesian in Glottolog;  over 56 such pairs the median shared fraction is 0.098.  At a single rate the result fits a post-Babel origin:  the fastest Polynesian rate puts Proto-Austronesian at 990 BC, a speed-up of 0.70 under the Masoretic date, and the median rate at 3110 BC, 1.21.  Only the rate-variation correction pushes it back, and at S near 0.1 that correction is unstable:  the two gamma shapes fitted on IE-CoR give 4140 BC and 10550 BC from the same calibration.¹⁷  Gray's 5,230 and Greenhill's 5,117 years, about 3200 BC, lie between the single-rate and the gamma figures.  Austronesian has no Bronze Age texts;  nearly every list is modern, so a Masoretic Babel leaves each lineage about 4,200 years, and lexicostatistics on modern lists cannot exclude a post-Babel origin at that depth.  The evidence against a young Austronesian is archaeological, the Taiwan Neolithic and the Lapita horizon, and those dates are radiocarbon.

Limits.  Glottochronology's reputation is deserved, and Part 1 shows why:  rates vary by a factor of two or more between documented lineages, and this test uses that fact by reporting every result at the fastest recorded rate as well as the median.  Borrowing is handled two ways, and the answer barely moves.  Coding choices, which philologist judged which words cognate, are IE-CoR's and not this article's;  a missed deep cognate makes the estimate older and works against the floor reading, and that effect cannot be priced from the data.  List noise makes short calibrations look fast:  Parkinson's 1770 Maori and Tahitian lists compared with the modern ones give absurd rates of 0.03 to 0.22 per millennium, a fixed per-list penalty that favours a short chronology and was excluded.¹⁷  IE-CoR's 170 meanings are not the Swadesh list, so its rates are calibrated on itself and are not interchangeable with Lees's.  The attestation dates are IE-CoR's tip dates with standard handbook anchors, and a historian's citation for each has not been assembled.  And the whole test measures one family in depth, Indo-European, and one at the root, Austronesian;  Sino-Tibetan was not run.

Objections and responses

"Glottochronology is discredited.  Bergsland and Vogt destroyed it in 1962.  So the divergence-time test proves nothing."  Bergsland and Vogt showed the rate is not constant:  Icelandic keeps 100% of the 100-word list over 800 years and East Greenlandic far less than the index over a few centuries.  The test in Part 2 does not assume a constant rate.  It reports every result at the fastest rate recorded over a millennium, at the median, and at the slowest, and it treats the fastest as a floor.  The young-earth case needs a rate faster than any of them, in every Indo-European lineage, for the first thousand years after Babel and never again.  Rate variation widens the error bars of a date;  it does not, by itself, make the families young.  And the fastest documented cases have known causes, borrowing, creolisation, taboo, that are hard to posit for every lineage at once.

"Small, isolated post-Babel clans would have changed language far faster than Latin or Old English did.  Bromham and Nettle show small populations change faster."  This is the strongest objection, and Part 2 computes it from published young-earth founder sizes rather than dismissing it.  Bromham's loss slope, applied to lineages growing from Osgood's or Carter and Hardy's numbers, gives a mean multiplier of 1.5 to 2.0, which halves the Masoretic gap and closes the single-rate Septuagint gap.  Three limits remain.  Bromham found no population effect on total change and no improvement from modelling founder effects;  the effect is front-loaded and vanishes once a lineage reaches tens of thousands, which the Hittite, Mitanni and Mycenaean states did centuries before their texts;  and the Masoretic gap with rate variation stays at 1.6 to 2.3 even in the central case.  Bergsland and Vogt's Icelandic is the standing warning that isolation can also freeze a language.

"The Septuagint date fits.  Why insist on the Masoretic?"  This article does not insist.  Under the Septuagint's 2767 BC the single-rate gap for Hittite and Vedic is 1.53, and with Duursma's faster loss in small groups, computed here from published founder sizes, it closes.  The young-earth reader who takes the longer chronology, or Morris's gap allowance, has a tenable position on the language dates.  The cost is that the longer chronology is a textual decision argued in the Old Testament text and chronology, and the writers who print 2242 BC have not taken it.

"Creoles form in a century.  Post-Babel contact could have scrambled vocabularies beyond recognition."  Creoles keep most of their lexifier's basic vocabulary.  Tok Pisin keeps 146 of 200 English items after a century;  Ma'a keeps most of its Cushitic basic vocabulary under a Bantu grammar.  Creolisation makes a language that looks like a relative of its source, not an unrelated family.  To erase a relationship this way, every family would need repeated creolisations with lexifiers from somewhere else, and that vocabulary would still be traceable to its source.  No young-earth writer proposes it.

"Austronesian fits a post-Babel origin on your own test.  Why is that not the headline?"  It is reported as a point for the young-earth side, and it is.  The reason it is not the headline is that Austronesian has no Bronze Age texts, so the test has far less power:  every lineage gets the full time since Babel, and the single-rate fit at the fastest Polynesian rate is a floor on a family whose root the Bayesian analyses put at 5,100 to 5,200 years.  What Austronesian shows is that lexicostatistics on modern lists cannot exclude a young family at that depth.  What Indo-European shows is that it can, when the texts are old enough.

Issues and unanswered questions

• No historian's citation for each attestation date in the computation. The tip dates are IE-CoR's own, with standard handbook anchors. A verifier should pin Hittite 1450 BC, Mycenaean 1350 BC and Early Vedic 1300 BC to a primary chronological source each.
• Sino-Tibetan was not run. The constant-rate test covers Indo-European in depth and Austronesian at the root. A run on a Sino-Tibetan cognate database with the oracle-bone Chinese as the old tip would show whether the 2.45 factor is an Indo-European peculiarity.
• Creole cognacy rests on one textbook exercise. Crowley's 146 of 200 for Tok Pisin is given without a source for the count. A primary Swadesh-list count for a creole against its lexifier, and for a mixed language such as Ma'a or Michif, would settle route (d) with numbers rather than one datum.
• Bromham's slope is thin. One significant population measure of three, for loss only, over ten Polynesian pairs. A replication on Indo-European or Bantu, where population histories are better known, would test whether the population escape route holds outside Polynesia.

Issues that bear on more than one of these articles, among them the unread source behind the Muyuw figure and which young-earth model is being tested, are listed in the parent, Language Families and the Tower of Babel.

Verdict

Languages diversify fast.  Oller is right that dialects become mutually unintelligible within about a thousand years, and in 1,000 to 2,200 years one attested language becomes a group of related languages.  But cognate retention, which decides whether a family can be recognised, decays slowly, and no documented case turned one language into unrelated families in that time.

Under a Masoretic Babel the Bronze Age texts require every Indo-European lineage to have changed 2.45 times faster than the fastest lineage recorded over a millennium, and 3.5 to 4.8 times faster once rate variation among meanings is counted, for the eight to nine centuries before Hittite and Vedic were written.  Faster loss in small founder groups, the one escape route a young-earth writer proposes, halves that gap when computed from the movement's own founder sizes.  It does not close the gap with rate variation unless every favourable assumption is taken at once.  The Septuagint date with the same founder groups comes within reach of the data, and with Habermehl's lowered Late Bronze Age, applied here to the texts though she does not apply it herself, it fits even with rate variation.  Morris's gap allowance at a thousand years leaves Hittite and Vedic just short at a single rate, 1.07, and only its far end, five thousand years, fits with no change to the rates.  Austronesian, with no Bronze Age texts, fits a post-Babel origin at the fastest Polynesian rate, and that is a real point for the young-earth side.

So the honest young-earth position on the language evidence takes the longer chronology, or Morris's gap.  It says plainly that Indo-European needs a burst of early change, in founder groups whose later states were not small, that no document records.

Companion pages

• The code and data: the divergence-time computation, with its outputs in research/data/.
• The literature: every source consulted, what it measured, and which way it cuts.
• Quotations: the best statements by linguists and by young-earth writers, with locators.
• Quarantine: quotations and attributions held back, and why.
• Language Families and the Tower of Babel: the young-earth model in its own words, the count of families, and the verdict on every test.
• Did Language Begin Once?: the comparative method's horizon, Proto-World and the macro-families, and the phoneme gradient.
• How Old Are the Language Families?: the dated family trees, and the languages written down before each Babel date.

Sources