You have probably heard that Chinese is one of the hardest languages in the world, and also the opposite, that it barely has grammar. Both remarks say the same thing: it is not like English. But "not like English" is a judgement about the comparison, not about Chinese — of the seven thousand or so languages on earth, almost all of them are not like English.
Put Chinese back among those seven thousand and measure again, and the four features people name most often all shift position: three move from rare to common, and only the fourth is genuinely alone.
Take it apart by ear first. One syllable, ma, four tones, four unrelated words: the first stays flat and high (妈 mā, "mother"); the second slides up from the middle, the way an English speaker's voice rises at the end of a question (麻 má, "hemp"); the third dips very low, often with a squeezed, creaky edge (马 mǎ, "horse"); the fourth drops hard from high, like an impatient "No!" (骂 mà, "to scold").
That sounds exotic until you count. The cross-language database WALS samples 527 languages: 307 have no tone, 132 have a simple system (a two-way high-low contrast), and 88 have a complex one (three or more levels, as in Mandarin). So more than four in ten languages use pitch to tell words apart, in dense belts across sub-Saharan Africa, Mesoamerica and mainland Southeast Asia.
The interesting part is in the brain. In 2000 a group played the same Thai tones to native Thai listeners and to native Chinese listeners. In the Thai speakers, the left hemisphere — the side that usually handles language — lit up while they processed those pitches; in the Chinese speakers it did not, because Thai tones do not distinguish words for them and were handled as plain pitch. Whether pitch counts as language or as melody depends on whether it does any work in your language. Evidence strength: small-sample imaging, with a converging result from a different method in Cantonese speakers.
There is a stranger layer still: this tone system was not always there. In 1954 someone showed, using Vietnamese, that tones are the shadow of final consonants that fell away — words once told apart by their endings lost those endings, leaving only the pitch trace the ending had left on the vowel, so pitch was forced to take over the job of distinguishing words. Chinese took the same road: its rising tone comes from a lost glottal stop, its departing tone from a lost final s. The most "Chinese" of features is only some fifteen centuries old, and a language from a different family on the same stretch of land walked into the same shape at roughly the same time.
English says "three fish"; Mandarin says 三条鱼 (sān tiáo yú, "three long-thin-thing fish"), and the extra word 条 (tiáo) is a numeral classifier. It is not optional decoration — without it the phrase is ungrammatical. Nor is the choice free: 条 (tiáo) covers long flexible things (fish, roads, snakes, rope), 张 (zhāng) covers surfaces you can spread out (paper, tables, beds, mouths), 把 (bǎ) covers things with a graspable handle (umbrellas, knives, keys). Before you can report a number, you have already committed to a shape.
This too is unremarkable. In a separate WALS sample of 400 languages, 260 have no numeral classifiers, 62 use them optionally and 78 require them as Mandarin does, clustered across East and Southeast Asia, Mesoamerica, Amazonia and West Africa.
What is more interesting is how these boxes grow. They start geometric and quickly leave geometry: 条 (tiáo) also counts news items and text messages. Japanese 本 (hon) originally covered long thin objects (pencils, neckties, trees) and today also counts telephone calls, judo bouts and baseball hits. Both languages pull the same "long and thin" outward, and in different directions.
In 2020 someone turned this into a computable question: take the nouns that emerged over the past half century and ask which model predicts the classifier each one takes. The traditional story is that a category revolves around a typical member and newcomers are seated by how close they are to the centre; what won instead was a weighted-exemplar model, in which a new word follows the handful of old words closest to it in meaning. Categories grow link by link rather than spreading out from a centre. Children fit this: corpus work finds the numeral-classifier-noun frame adult-like around two and a half, while which noun takes which classifier is still being tuned past six, with the general-purpose 个 (gè) standing in when nothing else comes to mind — exactly as the topic on what children get wrong argued, heavy use of the default is what shows the system is installed.
One large claim often hung on classifiers deserves a note. Some have argued that speakers of classifier languages lean toward sorting things by material, and speakers of non-classifier languages by shape. A set of six experiments in 2009 went looking: five replicated the effect, but the sixth — the only one that did not require interpreting what a noun meant — found nothing at all. So the effect looks more like how you read a word than how you see the world.
那场火 (nà chǎng huǒ) 幸亏消防队来得快 (xìngkuī xiāofángduì lái de kuài) — "that fire — luckily the fire brigade came fast" — is perfectly ordinary Mandarin, and very strange once you take it apart: the opening phrase has no relation at all to the only verb, 来 (lái, "come"). The fire does not come; the brigade does. The phrase is not the agent of coming, not its patient, not even a time or place. It is simply placed first, to announce what the stretch of talk is about. Linguistics calls that opening a topic.
English cannot place it that way. You have to attach it to something: "As for that fire, ..." or bury it in a clause. Mandarin needs no attachment, because position itself is the role: sentence-initial means topic. In 1976 two linguists used this to split languages into the subject-prominent (English, French) and the topic-prominent (Mandarin, Japanese, Korean). The dividing line is not whether a language has subjects; it is what the sentence organises itself around first.
Two more things fall out of this. One is 下雨了 (xià yǔ le, "raining"): English must supply an "it" that refers to nothing, while Mandarin does not even need the placeholder, because there is no slot demanding to be filled. The other is 你看见他了吗 (nǐ kànjiàn tā le ma?) — 看见了 (kànjiàn le), roughly "did you see him? — saw", where subject and object are both dropped and recovered from context, as Japanese also does.
Even this is not a lone case. What is worth singling out is a smaller difference: Japanese and Korean give topics a dedicated marker (Japanese は wa, Korean 는 neun) that tells you outright "this is the topic", while Mandarin supplies no marker at all, only position and pause. Mandarin topics have to be heard — which is also where they most often go wrong in cross-language settings.
First, a misconception to clear away: Chinese characters are mostly not pictures. A 2003 study went through all 2,570 characters taught in Chinese elementary school and found about 72% are semantic-phonetic compounds — one half hints at the meaning, the other at the sound. In 妈 (mā, "mother"), the left half 女 (nǚ, "woman") carries the meaning and the right half 马 (mǎ, "horse") the sound; in 铜 (tóng, "copper"), the left half is 金 (jīn, "metal") and the right is 同 (tóng, "same"). The system was half phonetic from the start.
The trouble is that sound changes and graphs do not. 江 (jiāng, "river") carries 工 (gōng, "work") as its sound hint and 河 (hé, "river") carries 可 (kě, "may"); two thousand years ago these were probably close, and now they are not. The same study found a counterintuitive arrangement: the characters taught in the first two grades have the fewest strokes and are also the ones whose sound hints are least reliable and whose meaning hints are least transparent, because the commonest characters tend to be the oldest and the most worn. What a child meets first is the least regular part of the system.
The cost is countable. A standard frequency list found that 2,500 common characters cover 97.97% of modern text, and another thousand bring it to 99.48%. That looks like a quick ceiling, but the missing two per cent is precisely where the cost sits: without them the other 98% does not read smoothly, and they can only be learned one at a time. With an alphabet, twenty-odd letters let you pronounce a word you have never seen, though pronouncing is not understanding.
The payoff is the mirror image: you can understand what you cannot pronounce. Many readers can tell from its components and context that 饕餮 (tāotiè) names a ravenously greedy mythical beast without being sure how to say it; English runs the other way, where "otorhinolaryngology" can be sounded out letter by letter and still mean nothing to you until someone says "ear, nose and throat". One credit here is usually misfiled: 电脑 (diànnǎo, literally "electric brain", computer) and 手机 (shǒujī, literally "hand machine", mobile phone) are easy to understand thanks to compounding, not to the graphs — written as diànnǎo and shǒujī they would be just as transparent.
So what is genuinely rare? Writing appears to have been invented independently only about four times: Mesopotamia, Egypt, China, Mesoamerica. Of those, the line that maps a sign onto a morpheme rather than onto sound alone, and is still in large-scale daily use, is down to this one system (plus the kanji borrowed from it into Japanese). Tone, classifiers and topic prominence all have company elsewhere. This one does not.
retry_budget_exhausted is a direct hit; at standup, three people pronounce the same abbreviation three different ways and it takes ten minutes to discover they mean one thing. This is the ledger from "an alphabet lets you pronounce what you cannot understand": pronounceability and searchability are two separate payoffs and you buy them separately. What to change: give every name that gets said aloud a "reads as" line in the README, and stop using names distinguishable only on the page (l1 versus I1), which are homophones on the spoken channel. Measure how often "which one do you mean?" interrupts a standup.One workable approach is rarity in a cross-language sample: in how many languages does the feature appear, and are those languages crowded into one geographic region? Crowding suggests the feature spread rather than arising independently. A second ruler is comparison with a language's own relatives: Classical Tibetan is related to Chinese, yet it had consonant clusters and no tone. When both rulers agree, you can be reasonably confident the feature belongs to that language.
It usually means "no inflection" — verbs do not change shape for tense, nouns do not change shape for number. But grammar is not inflection: word order, function words and topic structure are hard rules, and breaking them makes a sentence impossible. Narrowing "grammar" to "word endings" is a measuring stick that takes Indo-European as the default; change the default and the sentence stops being true.
Because differences only show up in contrast. A native speaker does not feel they are "using tone", just as an English speaker does not feel they are forced to choose a tense in every sentence. That is exactly what the imaging study delivered: the same physical quantity falls back to being mere sound when it is not part of your own system. So "what is special about my language" is nearly unanswerable by introspection.