필사 모드: What People Good at Languages Have in Common, and How Much of It Is Verified

Introduction

Pieces titled something like "the seven habits of language geniuses" are among the most reproduced genres on the internet. And most of them are impossible to verify. What supports the claims is usually one person's interview, that interview gets cited in another piece, that piece gets cited again, and at some point the phrase "according to research" is attached.

The purpose of this post is not to mock the genre. It is to pick out the real thing mixed in with it. I will take six common claims one at a time and mark the class of evidence each of them has.

First, the most important premise of this post. What is actually known about hyperpolyglots is very thin. And a substantial part of what is known comes from reporting, not from experiments.

1. What is known about hyperpolyglots is thinner than you would think

The most widely read book on this subject is Michael Erard's Babel No More: The Search for the World's Most Extraordinary Language Learners. It was published by Free Press on 10 January 2012. Kirkus Reviews describes it as narrative non-fiction weaving together historical investigation of the nineteenth-century polyglot priest Giuseppe Mezzofanti, interviews with contemporary hyperpolyglots, a review of brain research, and observation of a polyglot gathering in Belgium. That is to say, the book is excellent reporting, not experimental research. Evidence class: journalism.

On the experimental side, the two most frequently cited items are the papers from Fedorenko's lab. The previous post already covered them, so only the numbers are repeated here. The 2021 study had 17 polyglots (17 controls); the 2024 study had 34 polyglots. The definition of polyglot in the 2024 paper is by self-report: at least 4 out of 20 in five or more languages, and at least 10 in one of the non-native languages. The authors noted as limits that proficiency was self-reported rather than objectively measured, and that a sample of 34 is small for individual-differences research.

So the experimental foundation of this field runs to a few dozen people. Generalising "the common habits of polyglots" on top of that goes far beyond what the data can carry.

2. Survivorship bias: the structural defect of the genre

This is what I most want to say in this post.

Almost every learning method we encounter belongs to someone who succeeded. The person who agrees to an interview, opens a YouTube channel, writes a book, is already someone who succeeded. The person who failed never gets a place to explain their method.

The problem does not stop there. A substantial number of the people who failed may well have used exactly the same methods as the people who succeeded, and we cannot observe that fact. So the mere fact that a method appears in common among the successful does not let us say the method caused the success. It may have appeared at the same rate among those who failed.

This is survivorship bias. And this bias is not an incidental defect of the "what polyglots have in common" genre; it is the structure of the genre itself. That genre is designed to extract traits from a survivor-only sample.

There is only one way to avoid this bias: put the people who did not succeed into the sample as well. That is, you need research that tracks a randomly recruited group of learners and looks at which traits or behaviours are associated with later achievement. Such research does exist, but it is done on ordinary learners rather than polyglots, and the variables it handles are far more mundane. The aptitude tests and the anxiety research covered below are of that kind.

There is one more thing: the problem of retrospective self-report. That is exactly what Macnamara and Maitra's replication study dealt with. The design relied on values people estimated by recalling their past amount of practice, and the explanatory power that had been 48 percent in the original fell to 26 percent in the replication. People do not accurately remember how they came to be good at something. That is why the recollections of successful learners are interesting but weak as evidence.

3. Claim by claim

3-1. "They started young"

Here there is large-scale data. It is the paper Hartshorne, Tenenbaum and Pinker published in 2018 in Cognition. I read the abstract of this paper through the Semantic Scholar metadata API. The publisher page was blocked, so I could not read the full text.

According to the abstract, the study used data from 669,498 native and non-native English speakers and estimated the trajectory of learning ability with a computational model separating current age, age of first exposure, and years of experience. The result was that grammar-learning ability holds roughly steady until an adulthood threshold of 17.4 years and declines steadily thereafter. This pattern appeared not only in difficult syntactic phenomena but also in easy syntactic phenomena that are normally acquired early. The authors summarise the result as supporting the existence of a critical period for language acquisition, but with an end point far later than previously estimated.

What matters here is what this paper does not say. First, it is not saying that you must not learn after 17.4. It estimates the point at which the decline in learning ability begins; it does not set a point of impossibility. Second, what the study dealt with is grammar. It does not apply as-is to vocabulary, pragmatics or pronunciation. Third, the authors open the paper by pointing out that previous research lacked statistical power, and that measuring only the final attainment of learners who started at different ages cannot tell you how learning ability changes. In other words, this field was methodologically weak up until this paper.

In addition, I confirmed from search results that an academic commentary on this paper appeared in the same journal in 2020, but I could not read the full text. So the content of the commentary is not covered in this post. I do note, though, that this result was not accepted uncritically in the field.

Evidence class: large-scale single study. The direction is fairly solid, but it is a different story from the popular proposition that you have to start young to be good.

3-2. "They are sensitive to sound and have language aptitude"

Language aptitude is one of the few areas where a measuring instrument actually exists. MLAT is the long-standing standard, and more recently LLAMA is widely used.

The open-access paper Rogers and colleagues published in 2017 in the Journal of the European Second Language Association, volume 1, issue 1, examined the LLAMA tests. Reliability was inspected on data from 240 people between the ages of 10 and 75, and multiple regression was applied separately to data from 404 people. Some of the results: people who had learned a second language through formal instruction scored significantly higher than monolingual speakers. Prior second language learning experience explained 6 percent of the variance in LLAMA_B results. There was no significant difference between people whose native language was English, Arabic or Chinese, which supported the language neutrality of the tests. And the group aged 10 to 11 scored significantly lower than older groups on both the vocabulary task and the sound recognition task.

That last result is interesting. It points in the opposite direction from the frequently repeated line that the younger you are the better your feel for language. Of course scores on this test are not the same thing as actual language acquisition outcomes, so it should not be hastily inverted. And the authors write that these tests still need further validation in terms of their ability to predict language learning outcomes.

Evidence class: the existence and reliability of the instrument are confirmed at the level of a single study. How much aptitude predicts achievement is undetermined. As for the claim that being sensitive to sound is itself a characteristic of polyglots, within what I was able to check there is no direct evidence for it.

3-3. "They tolerate ambiguity well"

The story that tolerance of ambiguity is an advantage in language learning has been repeated for a long time in the learner-characteristics literature. In this round of investigation I did not reach a primary source supporting that claim. Searching showed that relevant literature exists, but I could not find something I could open, read and cite.

So I leave this item as follows. It is a claim people often make, and this post did not verify it. Evidence class: unverified.

Leaving it blank like this fits the purpose of this series better than pretending to have read literature I did not read and citing it anyway.

3-4. "Overwhelming volume of input"

The same applies here. The claim that large amounts of listening and reading are the core of language acquisition is extremely widespread in language teaching, and the associated theory is famous. But in this round of investigation I did not reach a study that directly measured whether polyglots take in more input than ordinary learners.

There is an indirectly related result. In the 2024 study by Malik-Moraleda and colleagues, the higher the self-reported proficiency in a language, the stronger the language network response. But this is a relationship between proficiency and brain response, not between input volume and proficiency. Moreover, the same paper noted that proficiency correlates with age of acquisition, which introduces a confound into the interpretation.

Evidence class: unverified. That input is necessary is close to common sense, but on "how much" and "what kind" there is no number this post can cite.

3-5. "They tolerate being publicly bad at it"

This item has usable evidence. It just does not point in exactly the same direction as the claim.

The open-access meta-analysis Botes, Dewaele and Greiff published in 2020 in The Journal for the Psychology of Language Learning, volume 2, issue 1, dealt with the relationship between foreign language classroom anxiety (FLCA) and academic achievement. It pooled 99 effect sizes covering 14,128 people using random-effects models. The result was a moderate negative correlation between anxiety and achievement, with a coefficient of -0.39 for general academic achievement (59 effects, 12,585 people).

Two things call for caution here. First, this is a correlation. Whether anxiety lowers achievement, or low achievement produces anxiety, or a third variable acts on both, cannot be determined from this data alone. Second, having low anxiety and adopting an attitude of volunteering to be publicly bad are not the same thing. The latter is a behaviour, the former an emotion.

Even so, this result shows that the item is not entirely folklore. That emotional factors in language learning are systematically associated with achievement is a fact confirmed with large-scale data.

Evidence class: a correlation confirmed by meta-analysis. Causation and direction undetermined.

3-6. "They use special memory techniques"

Techniques like spaced repetition, the memory palace and image association are frequently introduced as the secret of polyglots. In this round of investigation I did not reach a study measuring the use of memory techniques among polyglots. Within what I was able to check, the source of this claim is personal interviews and self-report.

Evidence class: self-report or folklore.

One thing to note is that the effectiveness of memory techniques in general and the characteristics of polyglots are separate questions. The literature on the former exists elsewhere and is outside the scope of this post. What has not been verified here is the causal claim that polyglots became polyglots because of it.

4. Evidence classes at a glance

I think this table is the most useful object in the series. Three of the six items are unverified or self-report.

5. How to read self-report

This is not to say you should ignore interviews with successful learners altogether. They are an excellent source of hypotheses. But there is a way to read them.

First, you have to ask about what that person did not do. A success story is a list of what one person did, and which item on the list was decisive is unknown even to them.

Second, you have to ask how many people used the same method and failed. Even if you cannot get an answer, the habit of asking the question blocks about half of survivorship bias.

Third, you have to look at the conditions. As the earlier posts in this series confirmed, interpreting graduate students are people who passed an entrance examination, and they spend a substantial part of the day in training. Transplant the method with those conditions stripped out, and what has been transplanted is not the method but the name of the method.

Fourth, when a number appears you have to look for its source. The figure of "80 percent of the neuro-linguistic operations" from the first post is the representative case. The more impressive the number, the more often it has no source.

What has not been verified

Here again is what this post could not verify.

First, I could not find a citable source on the relationship between tolerance of ambiguity and language achievement.

Second, I could not find a study measuring the input volume of polyglots.

Third, I could not find a measurement study supporting the claim that the use of memory techniques is a characteristic of polyglots.

Fourth, I read only the abstract of the 2018 paper by Hartshorne and colleagues. I could not check the composition of the sample, the details of the model, or the content of the commentary raised afterwards.

Fifth, the definition of hyperpolyglot itself differs from study to study. The 2024 paper took five languages as its criterion, while the widely cited popular book mentions six as a criterion. Different definitions mean different samples, and different samples mean the results cannot be combined.

Sixth, none of the research in this post directly addressed English learning by Korean speakers.

Closing

If what remains after reading this post is "so nothing is certain in the end", only half of it got through. Some things are certain.

There is large-scale data on the trajectory of grammar-learning ability, instruments for measuring language aptitude really exist, and the association between emotion and achievement has been confirmed with data on more than ten thousand people. The list of polyglot habits, by contrast, is largely unverified. Not mixing these two layers is the whole of what this post proposes.

And one question remains to ask yourself. The method I am about to imitate: is it the reason that person succeeded, or a habit the successful person happened to have? If you cannot answer that question, it does not mean do not try the method. It means do not blame yourself when it does not work.

• Previous post: What the Research Shows
• Next post: What an Ordinary Learner Can Actually Take from Interpreter Training

If you want to manufacture chances to speak and increase your exposure to anxiety in small steps, /tools/conversation-lab helps.

References

Every link was opened and checked directly on 16 August 2026.

• Hartshorne, Tenenbaum & Pinker (2018), A critical period for second language acquisition: Evidence from 2/3 million English speakers, Cognition, DOI 10.1016/j.cognition.2018.04.007. Publisher page inaccessible, so only the abstract was checked via the Semantic Scholar metadata API: https://api.semanticscholar.org/graph/v1/paper/DOI:10.1016/j.cognition.2018.04.007?fields=title,authors,year,venue,abstract
• Rogers, Meara, Barnett-Legh, Curry & Davie (2017), Examining the LLAMA aptitude tests, Journal of the European Second Language Association 1(1), DOI 10.22599/jesla.24. Open access: https://euroslajournal.org/articles/10.22599/jesla.24
• Botes, Dewaele & Greiff (2020), The Foreign Language Classroom Anxiety Scale and Academic Achievement, The Journal for the Psychology of Language Learning 2(1). Open access: https://www.jpll.org/index.php/journal/article/view/botesetal
• Jouravlev, Mineroff, Blank & Fedorenko (2021), The Small and Efficient Language Network of Polyglots and Hyper-polyglots, Cerebral Cortex 31(1), 62-76: https://pmc.ncbi.nlm.nih.gov/articles/PMC7727365/
• Malik-Moraleda et al. (2024), Functional characterization of the language network of polyglots and hyperpolyglots with precision fMRI, Cerebral Cortex 34(3), bhae049: https://pmc.ncbi.nlm.nih.gov/articles/PMC10928488/
• Macnamara & Maitra (2019), The role of deliberate practice in expert performance, Royal Society Open Science 6:190327: https://pmc.ncbi.nlm.nih.gov/articles/PMC6731745/
• Kirkus Reviews, review of Babel No More (Michael Erard, Free Press, 2012-01-10): https://www.kirkusreviews.com/book-reviews/michael-erard/babel-no-more/