The Distance to English

Two kinds of effort

“Is Japanese harder than Spanish?” has no answer until you say harder at what. Effort points make it answerable. And the model keeps two piles apart, because you put the two kinds of effort in completely different ways.

Across all 67 the split is 63% memory, 37% drill, and the ratio is often more informative than the total. Mandarin is 76% memory: its grammar is light and its workload is characters and vocabulary, stored one at a time. Korean is 55% drill, the highest here. Turkish is close behind at 53%, on the fewest exceptions in the set, tied with Indonesian and Malay: almost nothing to memorize, an enormous amount to make automatic.

A point is not an hour and it does not convert into one. Two people burn through points at very different rates, and the same person burns them faster on a good day. What the point does is let vocabulary, grammar, sound and script be added together at all, which is the thing that makes languages comparable.

The second pile is the one learners underestimate. Memory points you can put in anywhere: a deck on the train, a page before bed, a film with subtitles. Drill points only come off the total while you are producing the language in real time, at speed, with someone waiting for you to finish the sentence. That gap is why you can know a language on paper and still freeze the moment you have to speak it.

What each language takes

Least effort first, with every total split into the six things that take it. Click any language to open its own breakdown.

Afrikaans is the lightest at 2,498 points and Japanese the heaviest at 10,097, a spread of about four to one. The top is Romance and Germanic, joined by Indonesian and Malay, whose grammar is genuinely minimal even though almost none of their vocabulary is reachable. The bottom is Japanese, Cantonese and Mandarin, where a writing system is most of the work, alongside Burmese, Korean and Tamil.

Compare two

The totals are only half the story. Pick any two and the panel shows where the gap between them actually comes from, and which parts run the other way.

Where the effort goes

Each one is measured, then converted into points. The colors are the ones stacked in the ranking above. The components are not weighted equally, and no weight was chosen by hand: each one takes whatever its measurements add up to. Summed across all 67 languages, vocabulary takes nearly half of every point in the set; pronunciation, for all the attention it gets, takes three percent. The ranges tell the sharper story: the same component can be almost nothing in one language and the main event in another.

And one head start

Cognates do more than save you the trouble of learning a word. They make part of the language comprehensible before you have learned anything, and comprehensible input is how the rest gets acquired. So the model measures, from real frequency distributions, how many words you must actually be taught to understand 95% of running text once cognates are free. A French learner never has to be taught 30% of the vocabulary. A Mandarin learner is given nothing.

Why vocabulary is half the bill

Some parts of a language cost you far more than others. We did not pick those shares. We picked the prices, and the shares came out of the shopping.

Think of a grocery bill. Rice is cheap per bag. Saffron costs a fortune per gram. But you buy rice by the kilo and saffron by the pinch, so the rice still costs you more at the till. Learning a language works the same way.

Grammar is the saffron. One grammar system costs sixty times what one word costs. But a language only asks you to learn about eleven of them, and it asks you to learn two thousand words. Grammar is not cheap here. It is rare.

Vocabulary is the rice. It lands near half whatever we do, because of how we measure. One point is one word. Set the target at two thousand words and vocabulary is half the bill before anything else is counted. That is our ruler talking, not the language.

If 14% for grammar looks too low, you are reading the wrong line

What you feel as grammar is spread over three of the six: the systems themselves, the exceptions that follow no rule, and the part where you have to do it all at speed. Add those together and you get 42%, second only to vocabulary.

Turkish vowel harmony is the example. The rule takes a minute to understand. It takes a hundred hours before it comes out of your mouth on time. So we count it under real-time processing, because that is where the work actually is.

We checked the whole thing against the Foreign Service Institute, who have taught 52 of these languages to diplomats and know how many hours each one took. Our totals line up with their hours at 0.81. Where we disagree, we usually charge too little for writing systems. That is the weakest part of this model, and the part most likely to change.

None of this knows you

These numbers describe an average learner, and nobody is an average learner. Maybe you pick up sounds by ear and grammar never sticks. Maybe you can read a new alphabet in a week but cannot hold a new word in your head for an hour. Then the split above is wrong for you, and no amount of measuring fixes it.

So we ask you instead. The panel above the ranking asks four questions about what comes easily to you, and re-sorts the list on your answers. Tell it a new writing system does not scare you, and Mandarin climbs about a dozen places. Tell it that grammar and speaking at speed are both hard, and Turkish, Hungarian, Georgian and Korean slide the other way. Everywhere else on this page, the measured numbers stay exactly as they are.

Hardest at what

The ranking is the least interesting output. Sort by a single dimension and the picture changes completely. Click a row to open that language's breakdown.

The map

Effort radiates outward from the Germanic and Romance core, and the two most demanding languages sit at the far end of the same line.

Where the numbers come from

Five public datasets, plus a small number of declared judgements. The aim was to count wherever counting was possible.

How a word is scored

Each word in the comparison lists is classified by how a learner could actually reach it, and the route sets the score.

The distinction matters more than a raw cognate count would suggest. A resemblance is only a cognate if the two words share an ancestor, so the model checks the etymology rather than the spelling: German nicht is related to English not, not to night, and Vietnamese truong is not related to strong at all, however much it looks it.

What is judged rather than counted

Six or so conversion rates, turning sounds, rules and characters into points. A handful of weights for how alien a kind of system is to an English speaker. Counts of irregular forms. These are stated openly in each language's panel, and the ordering is robust to moving them: when every exchange rate was moved by half in each direction, the ranking held with a correlation above 0.98. The order is driven by the counts, not by the judgements.

What this cannot tell you

• It is not a prediction about you. There are no learner variables in the model: no aptitude, no motivation, no prior languages. Someone who already speaks Spanish faces a different Italian than the one described here.
• Points are not hours. They do not convert at a fixed rate, and the model deliberately refuses to pretend otherwise.
• Differences under about 3% are ties. The Romance cluster at the top should be read as one group, not a ranking.
• Some inputs are estimates, and they are labeled. Counts of irregular forms, the weights for how alien a system is, and a few transparency terms are judgements rather than measurements. Each language's panel shows which of its numbers came from data and which did not.
• Coverage is uneven. Frequency corpora exist for 39 of the 67; the rest rely on concept lists, and where the evidence is thin the estimate is pulled toward its language family and marked as such.
• Tone is counted, not bucketed. WALS sorts tone into none, simple and complex, and leaves 23 of these languages uncoded. That had charged Icelandic, Dutch and Afrikaans for tone they do not have, and scored Yoruba's three tones the same as Cantonese's six. Every language now carries its actual inventory, and each panel names the source for its own figure.

This is not here to settle arguments about which language is hardest. It is here to show you what kind of hard your language is: whether the work is vocabulary or grammar, whether you pay it by memorizing or by drilling, and therefore what the next year of learning is actually going to feel like.

Lucas BighettiTeacher AI