Someone has probably corrected the way you speak. Maybe a crossed-out preposition at the end of an English sentence, maybe a remark that your accent is a little heavy. The correction sounds like it is pointing at an error, the way an arithmetic mistake is an error.
But a surprising share of these rules have traceable origins: a year, a book, one person's preference. How the standard variety got picked in the first place, where the line between "a language" and "a dialect" is actually drawn, and how much of what you hear in someone's voice you put there yourself.
Start with Italy. At unification in 1861 the number of people in the new country who could use Italian was startlingly small: De Mauro put it at 2.5%, and Castellani, recounting in 1982 on broader criteria, arrived at about 9.5%. The two figures differ by more than a factor of three and the dispute is still open, but they point the same way — this "national language" was at the time the written language of a tiny minority. Its basis is fourteenth-century Florentine, the variety Dante, Petrarch and Boccaccio wrote in. What selected it was not clearer sounds or tidier grammar; it was that those three men's books were copied most.
Now back to that meeting. The 1913 Commission on the Unification of Pronunciation produced what is now called the Old National Pronunciation: Beijing sounds as the base, but with the entering tone kept (a syllable cut extremely short, as if the end were clipped off) and with a distinction that separated the initial consonants of 尖 (jiān, "sharp") and 坚 (jiān, "firm"). A careful compromise; north and south both got something. The trouble was that it had no native speakers — announcers had to be trained specially to produce it, and teachers could not hold it steady. In 1932 the Ministry of Education issued a new glossary of standard readings and switched the standard to Beijing pronunciation.
Laid side by side, the procedure becomes visible: select one variety out of those already in use, codify it (dictionaries, grammars, something fixed to point at), extend it into new functions (law, textbooks, news all written in it), and then wait for acceptance. Not one of those four steps turns on any property of the variety itself. They turn on whose books travelled, who sets the curriculum, who is on the radio.
Nor is step one required. Norway maintains two official written standards, each municipality choosing which to use, and no spoken standard at all: people use their home dialects in parliament and on the evening news. So the difference between a standard language and a dialect is not a linguistic difference. Hand someone a transcript and they cannot tell from the grammar whether it is standard; they can only check whether it was written into a dictionary and taught in a school.
The Dutch–German borderland is the old example. Before standard languages spread, people living in the eastern Netherlands understood the villages just across the border more easily than they understood the far west of their own country, and the German side was the same in reverse. Dialects vary continuously: go another fifty kilometres and something shifts, neighbours always understand each other, only the distant ends fall apart. But the two sides were not taught the same written language in school — one pulled toward Standard Dutch, the other toward Standard German.
Scandinavia is the mirror image. Danish, Norwegian and Swedish each have their own orthography, dictionaries and regulatory bodies, so they count as three; yet they sit in one continuum, and speakers of all three can broadly get through a conversation each using their own. Intelligibility is asymmetric, too: Nordic surveys find Norwegians understand their neighbours most easily, while Danes and Swedes struggle most with each other. Count by intelligibility and these are one language; count by states and orthographies and they are three. No clean criterion separates the two labels — and calling a variety a dialect says nothing about it being simpler or less complete, only that it has not been through those four steps.
English has a rule against ending a sentence with a preposition, and its source is on record. In 1672 the poet Dryden faulted his predecessor Ben Jonson for putting prepositions at the end, calling it a common fault with him, without saying why it was one. Ninety years later Bishop Lowth's 1762 grammar was in fact quite restrained about the construction: it is an idiom, he wrote, which our language is strongly inclined to; it prevails in common conversation and suits the familiar style in writing, though placing the preposition first is more graceful and better suits the solemn and elevated style. The charming part is that the sentence he wrote that in ends with a preposition. A remark about register was turned into a prohibition by later textbooks.
The ban on splitting an infinitive is younger still: proposed in 1803, popularised by Alford in 1864, and part of a nineteenth-century enthusiasm for importing Latin constraints into English — a Latin infinitive is one word, so nothing can be wedged inside it. Double negatives went the same way: ordinary in Chaucer and Shakespeare, ruled out by eighteenth-century grammarians who applied arithmetic, two negatives making a positive, to speech.
In Chinese, the written division of labour between 的 (de, marking an attributive) and 地 (de, marking an adverbial) was settled in the early twentieth century. In the decades when vernacular writing took hold and translated European works poured in, translators deliberately split the two characters in imitation of the adjective and adverb endings of English and French; a fourth character, 底 (de, marking possession, as in 我底书 wǒ de shū, "my book"), was also in use and did not survive. In speech all of them are pronounced alike, so the distinction does not exist there at all — it was a coordination convention among people who write, from the start.
The other half matters more: varieties that get marked as substandard usually turn out to be tightly rule-governed. African American English often omits the copula — He is nice comes out as He nice. Labov showed in 1969 that where it is omitted is anything but arbitrary: wherever Standard English can contract is to 's, this system can delete it; wherever Standard English cannot contract, it cannot delete either. He's nice contracts, so He nice is well formed; the final is in I don't know where he is cannot contract, so I don't know where he is not available. That is not a loosening of the rules but a different set of them — and to know where deletion is possible you have to be holding both systems at once.
An experiment run in Montreal around 1960 had a neat design: a set of English–French bilinguals each recorded one passage in English and one in French, the recordings were shuffled, and listeners were asked to rate "these speakers" — not knowing they were hearing the same people twice. The English guises scored higher on intelligence, dependability and ambition, and the gap was not opened only by English-speaking listeners; French-speaking listeners rated the French guises lowest of all. The result is not about the two languages but about who held economic position in that city at that time; run elsewhere, its direction follows local power relations.
Rubin's 1992 design is sharper. American undergraduates heard one recording — the lecturer was a native speaker of American English — while half of them saw a photograph of a white face and half a photograph of an East Asian face, the two matched in advance for attractiveness. The group shown the Asian face reported hearing a foreign accent, and scored lower on the comprehension test that followed. The audio was identical.
That result has since been chased down carefully. Zheng and Samuel in 2017 first reproduced the same shift with static photographs; but once the photographs became dubbed videos, once items with unambiguous accents were mixed in, or once the two faces were interleaved at random rather than blocked, the shift shrank, disappeared, or reversed. Their reading is that what moves is the rating, not the hearing — you did not actually hear a different voice, you gave the same voice a different score. That distinction makes the finding more worth taking seriously, not less: the score is the end that lands on a transcript or a performance review.
Attitudes also run backwards into how people speak. In Labov's stratified survey of New York, the more formal the setting the closer people moved to the line they carried in their heads — and the ones pushing hardest were not the top group. On the most formal task of all, reading word lists, the lower middle class overshot the upper middle class in pronouncing r. They were the most aware of which form counted as respectable, and the most invested, so they went past it.
A standard delivers two things: written intelligibility across dialects, and a single entry point for education (one set of textbooks). Norway shows both are obtainable from written standards alone, without also selecting a spoken one. The cost lands on a third thing — once one variety is set up as the measure, the others get ranked against it, and that step is not required by the first two.
Separate two kinds of sentence first. "Native speakers do not say this" is a factual claim and observation can overturn it. "This is inappropriate here" is a convention about setting; it holds in its setting, and it really does let large numbers of people write things others can read. The trouble has never been that norms exist, but that norms get restated as facts — "documents in our field aren't written that way" becoming "that sentence is wrong".
They are trained on written standards and sit in nearly everyone's writing channel, reaching more people than any language academy in history. But they differ from schools and exams in one key respect: no sanction, only a default, and a default can be ignored with one keystroke. If that difference is real, we should see a split — variance in written forms narrowing, while people's judgements about which sentences sound human stay where they were.