Does learning in English cost you? What a randomised trial with 2,263 students actually found

In short: a randomised controlled trial with 2,263 students found that the same online course delivered in English rather than the students' own language pushed dropout from 57% to 71%. Comprehension among those who stayed barely moved. The language didn't make the material harder to understand — it made people leave. And believing your English is good didn't protect you. Below: the study itself, the country data around it, the honest size of the localisation effect, and what to do about all of it when the best material in your field is in English anyway.

Why this question is worth asking at all

You open an AI course and hit a fork almost immediately. The better material is in English: better-known authors, more frequent updates, more topics covered. The material in your own language is usually later, usually shorter, usually a translation. Common sense says: if your English works, take English — you'll lose maybe five percent of the comprehension and gain a lot in quality.

Common sense is wrong here, and it's wrong in the opposite direction from the one you expect. The loss doesn't happen in comprehension. It happens in whether you finish at all.

This is hard to check, because ordinary observation is useless. People choose the language of their course themselves, which means any measurement of "who learns better" measures not the language but who those people are. The confident go to English. The cautious stay native. Comparing their outcomes is like comparing the health of runners and non-runners and concluding that running is medicine. You need an experiment where the language is assigned rather than chosen. That experiment exists.

The experiment almost nobody in edtech quotes

Bälter, Kann, Mutimukwe and Malmström, Applied Linguistics Review, volume 15, issue 6, pages 2373–2396, 2024. A randomised controlled trial. n = 2,263 Swedish students, an online programming course, two language versions — Swedish and English.

The load-bearing word is "randomised". Which language version a person got was decided by lot, not by preference. That is the only way to strip self-selection out of the result: because assignment is random, both groups are on average identical in ability, motivation, age and every other thing you didn't think to measure. The one thing that differs is the language.

The scale matters too. Two thousand two hundred and sixty-three people is not a focus group and not a poll on social media. It is a sample on which a fourteen-percentage-point difference cannot be waved away as noise, which the statistics below confirm.

The result: comprehension survived, the people didn't

• Dropout in Swedish — 57%. More than half. That's an ordinary figure for a free online course, and it is a useful correction to every motivational speech ever given: even in your own language, fewer than half of the people who start will finish.
• Dropout in English — 71%. Fourteen percentage points more. The statistics: φ = 0.2; p < 0.00001.
• Mean score among active students: 16.9 in Swedish against 14.3 in English. Effect size δ = 0.085 — negligible.

Look at those three lines together, because the meaning is only in the combination. The dropout gap is large and solid. The score gap exists, but the effect is so small it carries no practical weight. The authors' finding reduces to one sentence: English didn't damage comprehension — it made people leave.

That's counterintuitive, and it's worth sitting with for a second. We are used to thinking of a foreign language as a quality filter: understand less, get less. What it actually acts like is a tax on effort. Every paragraph costs slightly more fuel. Every unfamiliar word is a micro-pause. None of those micro-pauses costs anything on its own, none of them stops you understanding. But summed across ten hours of course, they turn into the feeling that stops you opening the tab on a Wednesday evening. Not "I don't understand this". Just "not tonight".

That is what dropout looks like from the inside. Nobody makes a decision to quit a course. People simply stop coming back.

The uncomfortable part: confidence doesn't help

Then comes the finding that was the least pleasant to read. Self-assessed English fluency did not protect against dropping out. Students who rated their own English as good left too.

The logic is the same as for the effect overall. Your self-assessment answers the question "do I understand this?". The answer is honest — yes, you do. But you don't leave because of incomprehension. You leave because of the accumulated cost of effort, and self-assessment doesn't measure that at all. You don't feel the tax you're paying. You only feel the bill at the end of the month, in the form of an abandoned course and the conclusion that "I suppose I didn't need it that much".

Which gives the practical consequence that is worth the whole article: "my English is fine" is not an argument in this debate. It's an answer to a different question.

What this study does NOT say

Now, honestly, the limits — because quoting strong results is the fastest way to lie.

• These are Swedish students. Sweden is a country where 90% of the population speaks English (Eurobarometer 540, fieldwork September–October 2023). If the effort tax is measurable even there, it is unlikely to be smaller where proficiency is lower — but that is reasoning, not measurement. We have no direct replication elsewhere.
• It's a programming course. How the effect behaves on a marketing course, or a prompting course, does not follow from this study.
• It's an online course with free entry. A 57% baseline dropout hints at exactly that. People who have paid leave less easily — and we don't know whether that swallows the language effect or not.
• Fourteen points is about a group, not about you. "The English group lost 14% more people" does not translate into "your personal chance of finishing drops by 14%". A group difference and an individual probability are different quantities.
• φ = 0.2 is a strength of association, not an importance. It says the link is real and stable, not that language is the main reason you'd drop out. The main reason people drop out is life.

You can test that discipline right here. The prompt below isn't about language — it's about how to read any study someone is using to convince you. We put the numbers from this very article into it and asked the model to tear them apart:

Here is a research finding. A randomised controlled trial (Balter et al., 2024)
put 2,263 students through the same online course in either Swedish or English,
with the language assigned at random.
Dropout: 57% in Swedish, 71% in English (phi=0.2, p<0.00001).
Among students who stayed active, mean score was 16.9 in Swedish and 14.3
in English, an effect size of delta=0.085.

You are a hostile peer reviewer. Answer point by point:
1. List the ways this result could be entirely true and still not apply to me.
2. Explain what phi=0.2 does and does not tell a reader.
3. The comprehension gap is tiny while the dropout gap is large. Give three
   different mechanisms that would produce exactly that pattern, and say which
   single piece of evidence would separate them.
4. Name the one over-claiming sentence that a person summarising this study
   is most likely to write.
5. Finish with the strongest argument AGAINST your own criticisms.

Point 5 is mandatory. Criticism that doesn't know when to stop is no better than advertising.

The country picture: who actually studies online

Now widen the frame. If studying in English is a tax, you'd expect countries with weaker English to study online less. We pulled the Eurostat and Eurobarometer figures and looked.

Took online courses, % of population (Eurostat isoc_ci_ac_i, 2025): Norway 30.57 · Finland 28.75 · Netherlands 28.40 · Spain 27.21 · Sweden 23.20 · Italy 19.36 · EU-27 16.21 · France 11.34 · Poland 10.44 · Germany 10.09 · Romania 2.59.

Bought e-learning, % of population (Eurostat isoc_ec_ibuy, 2019 — indicator discontinued): Netherlands 15.14 · Denmark 12.81 · Norway 9.96 · Finland 8.75 · Spain 8.10 · Sweden 6.08 · Germany 4.48 · EU-27 3.93 · France 1.31 · Italy 1.18 · Romania 0.66.

Speak English as a foreign language, % (Eurobarometer 540, fieldwork September–October 2023): Netherlands 93 · Sweden 90 · Denmark 87 · Finland 81 · Czechia 41 · Portugal 41 · Spain 38 · Bulgaria 29 · Poland 27 · Romania 25.

The first thing that jumps out is the Netherlands. 93% speak English — the highest on the list. And they also buy more online learning than anyone in the EU: 15.14% of the population against an EU-27 average of 3.93%. Four times the average. A tidy story: they know English, so they buy courses.

The second thing that jumps out breaks that story. Spain: English at 38%, and 27.21% take online courses — fourth place, far above the EU average of 16.21%. With English proficiency less than half of the Dutch figure, Spain studies online almost as intensively.

It gets worse for simple explanations from there. Germany sits at 10.09% for taking courses — a third of Spain's rate and below the EU average — and Germany is not the EU's poorest or most isolated country by any measure. Sweden, with its 90% English, is at 23.20%, below Spain. Romania is at 2.59% with 25% English, and there the link does look like it's there.

Why you cannot draw a conclusion from those tables

Now stop, because past this point is where people break their legs. Country-level numbers prove nothing about causes. Nothing at all.

First, everything here is tangled with everything else. Countries with high English proficiency are also countries with high incomes, universal broadband, a culture of lifelong learning, employers who pay for training, and a dozen other things. Attributing the difference to language is like attributing it to the number of bicycles.

Second, this is the ecological fallacy: properties of a country do not transfer to a person inside it. The fact that Spain on average studies online more than Sweden tells you nothing about one particular Spaniard and his English.

Third, the two tables measure different things in different years: one is "took courses" in 2025, the other is "bought e-learning" in 2019, and Eurostat discontinued that second indicator entirely. You can place them side by side as two separate photographs — never as a trend.

So why show them? For exactly one conclusion, and it's a negative one: the hypothesis "weak English means people don't study online" is not supported by the country data. Spain, at 38%, sits in fourth place. English is not the demand switch. The causal power of language we know from precisely one place — the randomised experiment above. One study of causes outweighs ten tables of correlations, and that is the entire point of this section.

Why EF EPI cannot be used

If you have ever searched for a "ranking of countries by English proficiency", you have almost certainly seen the EF English Proficiency Index. Everyone quotes it, from newspapers to edtech startup decks. We don't use it, and here is why.

EF's sample is self-selected — EF says so itself. The index is built from the results of people who came to EF's website of their own accord and chose to take a free English test. That is not a country's population. That is the set of people interested enough in English to google a test for it. The mean age of test-takers is 26.

Look at what that does in practice. EF places Romania 11th in the world for English. Meanwhile Eurobarometer 540 (a probability sample, n = 26,523) and Eurostat AES 2022 report that Romania is the worst in the EU: 51.2% of residents speak no foreign language at all, and 25% speak English.

Eleventh in the world versus last in the EU isn't a methodological quibble, it's two different universes. The explanation is simple and unflattering: the fewer people in a country with middling English, the more EF's sample consists of the few whose English is excellent. The index does not measure the country. It measures visitors to EF's website in that country.

Keep this as a working move that transfers anywhere: the first question to ask of any ranking is who entered the sample, and at whose initiative. If entry was voluntary, the ranking measures the desire to enter.

What translation is really worth: the honest numbers

Suppose the language effect is real. How big is it, in reach and in money? Here the localisation industry promises "+40–70% conversion", and those promises are backed by nothing independent — we could not find a single verifiable measurement underneath them. There are, however, two measurements you can check.

• eBay: +10.9%. When eBay rolled out machine translation of listings, exports rose 10.9% (Brynjolfsson, Hui & Liu, Management Science 65(12), 2019). The important detail: the effect fell as the price of the item rose. The more expensive the purchase, the less the language of the description decided.
• Wikimedia: +3.64%. An A/B test on localisation produced a gain of three and a half percent.

Eleven percent and three and a half. The effect is real, it's measured — and it's modest. Nowhere that anyone has actually counted is there any sign of "+40–70%".

Notice how the two lines of data converge on one thought. Dropout in the experiment: plus fourteen points. Translation at eBay: plus ten and change, and less than that whenever the purchase matters more. Both numbers say the same thing: language is friction, not a wall. The friction is real, it is worth removing, and it cannot be sold as a miracle.

That eBay detail deserves its own paragraph, because it's about you. The language effect shrank as prices rose — the more serious the decision, the less the interface language weighed in it. Translate that to learning: the more you actually need the outcome, the less a foreign language will stop you. The reverse also holds, and this is the unpleasant half: if you only sort of need the material, a foreign language will almost certainly finish the idea off. It isn't the cause. It's the last straw.

And one more thing nobody thinks about: the price floor

Translation is not the only part of localisation, and it isn't the expensive part. There is also price, and price is structured more harshly than it looks.

We pulled Udemy's Price Tier Matrix V3 and converted it at ECB rates for 16 July 2026. The price floor in Brazil is R$24.90 ≈ €4.27. In Mexico it's MXN 129 ≈ €6.46. In Colombia, COP 34,900 ≈ €9.45. And in Argentina, ARS is not supported by Udemy or Coursera at all — locals pay rich-country prices.

From which follows the thing you can't see from a conversation about translation. Localisation is not "translate the buttons". It is also entering a market where the price floor is four times lower than you're used to, while in the country next door the currency isn't supported by the platforms at all. Translation is the cheapest part of the job. Which is precisely why everyone does it first, and is then surprised it didn't work.

What we honestly don't know

The rule is simple: if we had data on demand by language, we would have shown it here. We don't. Nobody publishes data on demand by language.

Counting courses per language on the platforms is not a workaround, and it's worth understanding why. Those counters measure supply, not demand. Zero Swedish courses on a topic means exactly two things at once: either it's an empty niche nobody thought to enter, or it's proof there is no market there and whoever entered left. One number, two readings, and the number cannot separate them. We're spelling this out precisely because the temptation was strong and we didn't take it.

So the balance sheet for this section: one strong causal result about dropout, two measured estimates of effect size, country tables that prove no causation, and zero data on demand. That's less than we'd like. But it is exactly what exists.

What to do about it when the best material is in English anyway

And it is in English, and that won't change soon. So the question isn't "which language is better", it's "how do I pay less of this tax".

First and most important. Since the effect hits dropout rather than comprehension, the thing to treat is dropout. Not comprehension. Not vocabulary. Finishing is the only metric worth fighting for here. Anything that lowers the friction of returning to the tab works; anything that improves your English over the next year has nothing to do with this course.

• Spend the foreign language on what doesn't exist in yours. If the material exists in your own language and isn't worse, take it. That's not weakness, it's saving fuel for something that matters more.
• Map before territory. Fifteen minutes on terminology before the course removes half the micro-pauses inside the course. There's a prompt for that below.
• Read rather than watch, where you can. Text runs at your pace and re-reads for free. Video runs at someone else's pace, and that is the single most common complaint we found in course reviews: "It is too fast for a beginner… until you try to see and understand it the next screen pops up" (Amit A., 29.04.2026, 2★), and "I am very frustrated that I can not easily access a transcript to reference later. The instructors are talking so quickly - I can not follow" (Vince D., 16.03.2026, 1★). Both complaints are about pace — and pace in a foreign language multiplies.
• Keep a model beside you as a translator on demand. Not translating the whole course, which defeats the purpose. Translating the one phrase you tripped on, in three seconds, without leaving the flow.
• Shorter sessions than you want. The effort tax accumulates. Thirty minutes four times a week beats two hours once a week, because two hours in a foreign language exhausts you into never having a fifth session.

Here's the prompt that does the second bullet for you. It's self-contained — just run it:

You are a bilingual technical editor. Answer in the language of this message,
but keep every English term in English.

Build a glossary of the 15 English terms a beginner meets in the first hour
of any AI course: prompt, token, context window, hallucination, fine-tuning,
temperature, system prompt, few-shot, zero-shot, RAG, embedding, inference,
chain-of-thought, guardrails, agent.

For each term give exactly four lines:
1. The term as it appears in the course subtitles.
2. One sentence of plain explanation.
3. The literal translation people reach for, and why it misleads.
4. One sentence using the term the way a lecturer would say it.

Then name the three terms whose everyday meaning fights hardest with their
technical meaning, and say what a beginner hears instead of the meaning.

The second prompt shows you the effort tax to its face. It builds a typical paragraph of course narration and dissects what you stumble on:

You are an online course scriptwriter. Do two things, in order.

First: write one 120-word paragraph IN ENGLISH exactly as an average online
AI course would narrate it — phrasal verbs, filler words, an idiom or two,
and one sentence that runs on far too long.

Then, in the language of this message:
1. Translate the paragraph.
2. List every construction a competent non-native listener is most likely
   to misparse, and explain why.
3. Ask me three comprehension questions about the paragraph.
4. Put the answers at the very bottom, under a line of dashes, so I can
   try first.

And this one is about where exactly you get lost. We wrote the study's conditions into it and asked the model to break dropout down moment by moment:

You are a dropout researcher for online courses. Answer in the language
of this message.

Situation: a person signs up for a 10-hour online course in their second
language. They finish the first video and never come back.

1. Reconstruct the 10 most likely moments of the drop, in the order they
   occur, from opening the page to closing the tab for good.
2. For each moment, mark whether language is the cause, the excuse,
   or irrelevant.
3. For each moment where language is the real cause, give one concrete fix
   that takes under two minutes and costs nothing.
4. Then argue against yourself: name the two moments where your fix probably
   doesn't work, and say what you would do instead.
5. Finish with one line: the cheapest single intervention with the widest effect.

The uncomfortable part about us

Now the honest bit, because without it everything above is advertising.

The English version of this site is the main one. The bulk of our material is made in English first — which is exactly the decision that the study in this article argues against. By its logic, some readers for whom English is a second language will not finish our courses: not because they won't understand, but because they'll pay the effort tax and one day not come back. We know the size of it: plus fourteen percentage points of dropout. We pay it knowingly.

Why do it anyway? Because the alternative is not "the same thing, in your own language". The alternative is less material, updated less often, at lower quality: a translation always trails the original, and in AI a six-month lag means an obsolete course. The competitor reviews we collected show how that ends: "The content is mostly from 2023. …I invested my 41 hours and Im learning content which is from 2023. Very disappointed" (Harsh A., 09.04.2026, 1.5★). Forty-one hours of your life on three-year-old material is no better for being in your native language.

So we do what we can: keep both versions, write the second one as a standalone text rather than a translation, and refrain from telling you that English is free. It isn't. It's fourteen points.

The whole thing in one paragraph

The language of instruction hits distance, not comprehension. A randomised trial with 2,263 students: 71% dropout in English against 57% in the students' own language, with a score difference of δ=0.085 — effectively zero. Confidence in your English doesn't protect you, because it answers "do I understand", and incomprehension isn't why you leave. The localisation effects anyone has genuinely measured are modest: eBay +10.9% (and smaller the more expensive the purchase), Wikimedia +3.64%; the "+40–70%" promises are backed by nothing. Eurostat's country tables prove no causation and, if anything, refute the "no English, no online learning" hypothesis: Spain, at 38% English, takes more online courses than Sweden at 90%. One practical conclusion: since the problem is dropout, treat dropout — short sessions, a glossary up front, text instead of video where possible, and your own language wherever it isn't worse. Next in this topic: free AI tools and their limits, what a prompt is, ready-made prompt examples and how to use ChatGPT.