"Young people are ruining the language" is a much older complaint than you think. In late Roman antiquity somebody copied out a list correcting how people were writing and saying things: write this, not that. The forms condemned on that list grew up to become Spanish and Italian. The person holding the pen was on the losing side.
To tell change apart from decay you have to see how change actually happens. Three things below: how one sound drops out of thousands of words at once; how a perfectly concrete word gets ground into a piece of grammar; and how a change spreads from a few people to everyone.
Start with the evidence. A 1679 English dictionary lists knave and nave, knight and night, knead and need as homophones — by then each pair sounded identical. Across the seventeenth century, k and g disappeared before n throughout the vocabulary. The spelling stayed where it was; the pronunciation walked off.
The point is not that they dropped but that they dropped together. No speaker sits down and decides word by word. What drifts is how the tongue handles a position — word-initial consonant plus nasal — and that position is the same in every word. This is why sound change looks so tidy: what it operates on is the sound, not the word.
The most famous set of correspondences works exactly this way. Where Latin kept the older shape, the Germanic branch swapped the whole inventory: pater against father, piscis against fish (p becomes f); tres against three (t becomes th); cornu against horn (k becomes h). Those are not four coincidences but three regularities, each having run over thousands of words. In the 1870s a group of German scholars pushed the observation to its limit: sound change has no exceptions — an exception means you have not found the conditioning yet.
The slogan was nearly falsified, then rescued. That set of correspondences did have a batch of disobedient words. In 1875 Verner showed that they depended on which syllable carried the older accent: where the accent came first the consonant stayed voiceless, and where it did not the consonant was voiced. English still carries a fossil of this — was and were, one verb, one form with s and one with r. That r descends from the voiced s. The exceptions were absorbed into the rule, and the rule came out harder.
None of this is confined to Indo-European. Middle Chinese syllables could end in -p, -t or -k, and Cantonese still carries them: 六 (luk6, "six"), 國 (gwok3, "country"), where the final stop is made by blocking the airflow and never releasing it. The Mandarin branch lost that whole position, and the same two words are now liù and guó. It is not that a few words dropped their endings; the position collapsed. (Neither branch "preserved things better" or "degenerated" — they took different roads.)
The ground is not perfectly level, though. Working on Chinese dialects in the late 1960s, William Wang and colleagues proposed a different picture: a change may be abrupt in itself, yet spread word by word, stall halfway, and leave residue. In Chaozhou one class of words has split into two readings with no phonetic condition that accounts for the split. Labov brokered a settlement in 1981: both mechanisms run, each with its own territory. Strength of evidence: wholesale correspondence rests on systematic matches across thousands of words, which is very hard evidence; lexical diffusion rests on the residue in dialect surveys, which is also real data. Nobody today takes only one side.
When you say I'm going to London, somebody is actually moving. When you say I'm going to tell you, nobody is walking anywhere. The second use starts appearing in the late fifteenth century; there are examples by 1483. A verb meaning "move" was conscripted as a marker of the future.
Three things happen at once: the meaning bleaches, and the motion is gone; the form erodes, and going to contracts to gonna; the slot hardens, and it is no longer freely a verb. The whole process is called grammaticalization — a contentful word turning into a grammatical element.
There is a clean diagnostic that English speakers can run on themselves: gonna takes only a verb after it. I'm gonna tell you is fine, I'm gonna the store is not — yet I'm going to the store is perfectly normal. One source has already split into two roads, and the eroded one did not inherit the motion use. Erosion marks the fork.
Chinese has taken the same road further. 了 (liǎo) was a verb meaning "to finish", still used that way in tenth-century vernacular texts; it was first reanalysed as a resultative element after the verb, then ground down into an aspect marker read as toneless 了 (le, the perfective particle), which cannot stand alone or take an object. 把 (bǎ) was a verb meaning "to grasp, to hold"; from the late Tang it bleached into an object marker, and in 把门关上 (bǎ mén guān shàng, "close the door") nobody is holding anything. 给 (gěi, "to give") took the same road, sliding from the act of giving toward the slots occupied in English by "by" in passives and "for" in benefactives.
This is not chance. Bybee, Perkins and Pagliuca surveyed 76 languages in 25 phyla in 1994 and found the path from motion verb to future marker recurring in languages with no relation to one another (French says je vais partir, literally "I go leave"). They also proposed unidirectionality: contentful to grammatical is the norm, and the reverse is rare.
The strength of unidirectionality should be stated plainly: it is a strong statistical tendency, not a law. Counterexamples exist but are scarce — the suffix -ism pulled out and used as a noun (all those isms), the preposition up pulled out as a verb (up the ante). Such cases have been collected on purpose, and there are far too few to overturn the trend.
The engine is frequency. Only combinations said often enough get ground down: going to contracts to gonna, while the structurally identical swimming to does not. Which incidentally explains something you have seen a thousand times without asking about it — why grammatical elements in language after language are so short. Usage wore them down.
That Roman list has a name: the Appendix Probi, 227 entries of the form "A, not B". One of them reads speculum non speclum — write speculum ("mirror"), not speclum. The condemned form, speclum, is the direct ancestor of Spanish espejo and Italian specchio. What the copyist saw as an error was a change in progress: an unstressed vowel in the middle of the word being squeezed out.
Change also has a fixed shape in time. In 1989 Kroch used Ellegård's historical corpus to look at English do: the meaningless do in Did you see him? was not always there. It took centuries to travel from scattered thirteenth-century uses to obligatory status in modern English. Plot the proportion of use against time and you get an S: slow at first, steep in the middle, slow again at the end.
What he found next is the interesting part. The same change runs in several syntactic environments at once — questions, negatives, emphatics — and their starting points are far apart, some well under way while others sit near zero. But the slope at which they rise is identical. He called this the constant rate effect. The weight of that sentence lies here: how fast the change rises is not set by the environment. The environments merely board at different times.
So who boards first? Labov went to Martha's Vineyard in 1963. Two island vowels had been drifting toward the mainland pronunciation, and he found a subset of speakers going the other way, pulling them back toward the more island-sounding variant — a subset concentrated among younger fishermen who meant to stay and had opinions about the summer tourists. Pronunciation had become a marker of identity, though nobody was using it consciously. In 1966 he counted whether clerks in several New York department stores pronounced a final r: the proportion differed by store, and younger speakers used it more.
That handed historical linguistics a convenient method called apparent time: compare people of different ages at a single moment and treat the result as a snapshot of the change. Its premise should be put on the table — that pronunciation is essentially fixed after adulthood — and that is only an approximation.
So "the language is getting worse" is really a statement about something else: you have taken the version you learned as the yardstick. The riverbed keeps moving, and no stretch of it is where it was supposed to be.
Because a second machine runs in the opposite direction: analogy. After a regularity has passed through, it leaves behind a litter of forms that no longer match each other, and speakers — children learning to talk above all — flatten them back toward the commonest pattern. That is how English past tenses migrated wholesale to the -ed ending. The irregular forms that survive are almost all high-frequency, because you hear them often enough not to need a rule to derive them. The division of labour is clean: erosion eats the frequent first, analogy cleans up the infrequent.
Both, but about different kinds of change. Erosion of sound tracks how often a form is used, so the most frequent forms are worn hardest: greetings, address terms, auxiliaries, worn down past recognition. Regularization of morphology tracks strength of memory, and a form recalled that easily resists the rule. The same frequent word can therefore be the fastest to collapse phonetically and the most stubborn morphologically. You can check this yourself: take the handful of words you say most often in your own language and see whether both hold.
The difficulty is in how it is stated. "Counterexamples are rare" is hard to knock down with a single counterexample, and what counts as one depends on where you draw the line between contentful and grammatical — some treat -ism used as a noun as a counterexample, others as a one-off conversion. To make the claim falsifiable you would first need a criterion for "moving backwards" that does not rest on intuition: whether the form gets longer, say, or whether it can freely take arguments again. There is still no agreed answer to that step.