title: "2,854 People Read My Articles. Zero Commented. Then I Translated Them Into English."
tags:
- AI
- CulturalAnthropology
- Linguistics
- Alignment
- Communication
private: false
updated_at: ''
id: null
organization_url_name: null
slide: false
ignorePublish: false
Abstract
The same author published the same content in Japanese and English simultaneously.
Japanese (Zenn): 2,854 readers over 2.8 months. Zero comments. Six followers.
English (Medium/X): Within one week — a physician with 12.4K followers, an AI analyst with 5K+ Substack readers, and an AI researcher with 14,000 followers all responded.
Same content. Different language. Dramatically different reactions.
Why?
This article dissects the asymmetry through cultural anthropology. It rejects the simplistic explanation of "Japanese people are introverted" and cross-analyzes platform norms, language structure, cultural logic of knowledge sharing, and epistemic risk distribution.
The final destination is not a binary "which culture is better" — it's a question beyond that:
What emerges when we merge the "silent deep-readers" of Japanese-speaking communities with the "reactive consumers" of English-speaking communities?
Prologue: An Anomalous Silence
On February 12, 2026, I published an article titled "I Calculated the Probability That 2,854 People Read My Work Without a Single Response."
The content was a statistical analysis using my own analytics data. Using a Poisson distribution, I calculated the probability that "zero comments across 96 articles" was random. The result: 0.0068% or less.
The data told a strange story:
It was being read. 2,211 Japanese readers spent an average of 1 minute 35 seconds per visit. One reader in Okayama: 7 minutes 37 seconds. Kochi: 4 minutes 17 seconds. Niigata: 5 minutes 30 seconds. The referrer data showed statics.teams.cdn.office.net in 22 sessions — evidence that someone shared the link internally via Microsoft Teams. Edge browser users (likely corporate PCs): 3 minutes 31 seconds.
No visible reaction. Zero comments. Six followers. 0.43 likes per article — less than one-seventh of benchmarks.
This couldn't be dismissed as "no one was interested." There was interest. There was deep reading. There was internal corporate sharing. But no visible response.
Eleven days later, one reader finally broke the silence. They honestly explained why they had never commented before. Their testimony corrected and refined my hypothesis.
Two weeks after that, results from the English-language experiment came in.
When I translated the same content into English and posted it to Medium, within one week: a physician with 12.4K followers, an AI analyst with 5K+ Substack readers, and an AI researcher with 14,000 followers all responded.
Same content. Different language. Dramatic difference in reaction.
Explaining this asymmetry is the purpose of this article.
Part 1: Dissecting the Japanese Silence
The Data
Basic statistics (November 20, 2025 – February 12, 2026):
| Metric | Value |
|---|---|
| Articles published | 96 |
| Active users | 2,854 |
| Japanese real readers | 2,211 |
| Average session duration (Japan) | 1 min 35 sec |
| Engagement rate | 50.96% |
| Comments | 0 |
| Followers | 6 |
| Likes per article | 0.43 |
The Poisson analysis: even with the extremely conservative assumption of 0.1 comments per article (one comment per 10 articles), the probability of zero comments across 96 articles is 6.8×10⁻⁵. Statistics confirm this is structural, not random.
The anomaly: quality of reading vs. absence of reaction
| Region | Users | Engagement Rate | Avg Session |
|---|---|---|---|
| Okayama | 27 | 64.29% | 7 min 37 sec |
| Niigata | 16 | 52.83% | 5 min 30 sec |
| Kochi | 18 | 66.67% | 4 min 17 sec |
| Edge (corporate) | 284 | 63.56% | 3 min 31 sec |
The 27 people in Okayama spent 7 minutes 37 seconds reading. This is not accidental browsing — this is the sustained concentration of reading a research paper. Yet not one comment emerged.
Read → Deep read → Share with colleagues → But silence.
Some powerful structural force was suppressing response.
A Reader's Testimony
The comment that arrived 11 days later was testimony explaining this phenomenon from the inside.
A Zenn technical reader explained why they hadn't commented (paraphrased to avoid direct quotation):
- Not because I'm a non-engineer — on the contrary, the Abhidhamma perspective on AI deserves more attention
- I didn't comment because I felt I had no "valuable opinion or knowledge" to contribute
- On a technical site like Zenn, "This was interesting!" feels like it has no value. On X that's fine, but on Zenn it feels inappropriate
- Those who can identify errors in AI content are probably less than the top 1%
- Japanese engineers are good at pointing out errors with data, but not proactive about exchanging opinions openly
- Publishing in an academic journal would have the most impact — there's a mismatch between Zenn and this type of content
This testimony corrects my original "armor hypothesis."
My original proposal was: "有身見 (sakkāya-diṭṭhi) = the armor of an account" — the idea that pressing like creates a visible record, so there's a calculation about what colleagues would think. This was half right but incomplete.
What this reader's explanation adds: platform norms.
Zenn has an implicit norm about appropriate comment quality — it is the provision of technical value or pointing out errors. Expressing "this was interesting" doesn't fit that norm. Writing a norm-violating comment wounds one's sense of belonging to the Zenn community.
So they stay silent.
This is not just "armor" (self-protection). It's the logic of adaptation to community norms — sincerity to the space.
The reader also observed: "I'm unaware of a space online where free and open opinion expression and discussion on technical topics is possible in Japan. If it exists at all, it would be within university research labs."
This is not about individual personalities. It's about the structural absence of infrastructure for free public opinion exchange in Japan's technical community.
Part 2: The English-Language Experiment
One Week of Reactions
In mid-February 2026, I resumed activity on Medium. Same content — Japanese articles translated into English.
Response 1: A physician (12.4K followers)
Article: "The Language You're Using to Build AI Might Be the Reason It Keeps Lying to You"
Why did a physician respond to an AI × linguistics article?
Their writing consistently circles the problem: "language determines patient behavior." "You are obese" versus "your BMI suggests we explore some options together" produce completely different patient responses. Language structure creates reality — learned in the body over 15 years of clinical practice.
My article's claim: "AI's tendency to confabulate might be caused by language structure" — structurally identical to the core of their expertise.
"Structure is the problem" — this recognition frame cuts across medicine, language, and AI.
Response 2: An AI analyst (5K+ Substack readers)
An AI/Tech analyst running "Neural Foundry" on Substack started following me. Their readership deeply analyzes technology trends. An article at the intersection of AI × Buddhism × language structure reaching this cluster shows they're carrying the question: what are the structural causes of AI's limitations?
Response 3: An AI researcher on X (14K followers)
My X post: "Read this thread before commenting on AGI benchmarks"
Content: pointing out measurement problems with the Einstein test proposed by Demis Hassabis. "A 1911 cutoff doesn't remove Einstein's reasoning structure from the training data. The influence of Mach, the tension with Maxwell's equations, the equivalence principle — all of it exists as text. You're not testing independent discovery. You're testing whether the model can reconstruct the path from compressed traces of that same path."
A verified AI researcher with 14K followers liked it — signaling: "this person understands measurement-theoretic problems."
Response 4: Ship series
"The Mathematics of Why Ships Float — And Why Getting It Wrong Kills People" — 29 likes and 1 comment in two days.
Same content in Japanese on Zenn: a few likes, zero comments.
The content didn't change. Only the language changed.
What Changed
| Changed | Didn't Change | |
|---|---|---|
| Language (Japanese → English) | Content and logic | |
| Platform (Zenn → Medium/X) | The author (me) | |
| Readership demographics | The underlying ideas | |
| My attribute visibility |
The most important change: my attribute filter disappeared.
My Japanese profile says: "50-year-old househusband, non-engineer, graduate of Bibai Technical High School." In Japan's technical community, this profile functions as a credibility proxy. It answers "can this person's technical claims be trusted?" with: "possibly not a technical person."
In English-speaking communities, there's just the account "Dosanko Tousan." My age, hometown, education, work history — invisible. Content arrives first.
In Japanese-speaking communities, "who wrote it" precedes and frames the evaluation of content. In English-speaking communities, "what was written" is the primary axis.
This is not a story about Japanese authoritarianism vs. Western openness. It's a difference in cultural patterns of information processing.
Part 3: Cultural Anthropological Analysis
The Technical Community as Community
Technical communities are not mere information exchange spaces. They are communities. Communities have norms, and norms structure behavior.
Japan's technical community norms:
The norm "don't write if you have no valuable opinion or knowledge" functions as a quality filter. Low-quality comments increase noise and degrade the signal-to-noise ratio. Silence is a form of autonomous quality control.
Cultural psychologists Markus and Kitayama note that Japanese society holds an interdependent self-construal — the self defined through relationships with others, not as an independent individual. This brings high attention to reading the "ba" (場, scene/place) and acting appropriately within it.
Zenn as a "ba" is a community where high-caliber specialists share knowledge. To participate as an "appropriate member," technically appropriate contributions are required. When unable to contribute, refraining from participation is itself a consideration for the space.
This is not "armor." It's sincerity to the community.
Anthropologist Edward Hall's High Context / Low Context distinction applies here. In high-context cultures (Japan), context, relationships, and non-verbal information are prioritized. Not commenting is itself communication — reading the context and choosing the appropriate (zero-cost, high-consideration) response.
In low-context cultures (English-speaking Medium), explicit reactions are natural communication. "This was interesting" carries sufficient meaning. Participants react because reacting is the appropriate norm.
Hofstede's Cultural Dimensions
Geert Hofstede's 6-dimensional model maps the structural differences:
| Dimension | Japan | USA | Effect |
|---|---|---|---|
| Power Distance | 54 | 40 | Credentials precede content evaluation in Japan |
| Individualism | 46 | 91 | Community impact calculus in Japan; personal preference in USA |
| Uncertainty Avoidance | 92 | 46 | Withholding response when credibility is unverified |
| Long-term Orientation | 88 | 26 | Calculating reputation impact of every reaction |
All four dimensions work in the direction of raising response costs in Japan.
This is not a flaw. It's a different optimization.
The Information Theory of Silence
From Claude Shannon's perspective: silence is not zero information.
Shannon defined information as "the amount of uncertainty reduction." The information content of an action is determined by how "surprising" it is.
7 minutes 37 seconds of deep reading is a surprising action in a context where the average is 1 minute 35 seconds. It carries high information content. The signal "this article has above-average value" is being transmitted — just through a channel invisible to the author.
| Japanese Deep Reading | English Like | |
|---|---|---|
| Information density | High | Low |
| Visibility to author | None | Immediate |
| Feedback loop | Open (author can't see) | Closed (author can see) |
A closed feedback loop of low information density contributes more to community self-organization than an open feedback loop of high information density.
This is the fundamental asymmetry. Not quality — architecture.
Part 4: Language Structure and Knowledge Sharing
English Assertiveness vs. Japanese Epistemic Modality
In a previous article ("Why English Struggles with AI Alignment — The Japanese Option"), I established that English grammar requires subjects:
- "I think X"
- "I believe Y"
- "I disagree with Z"
Opinion expression structurally becomes definitive first-person statements. When wrong, that error is definitively mine.
Japanese offers abundant epistemic modality expressions:
- 〜と思われる (it appears that...)
- 〜という見方もできる (one way of looking at it...)
- 〜のようだ (it seems...)
Opinions can be stated without attributing them to first person. Risk is linguistically distributed.
This asymmetry of linguistic risk distribution affects the frequency and quality of public opinion expression.
English-speaking engineers always bear risk as structurally mandated by the language. Risk acceptance is grammatically prescribed — the cultural threshold for accepting risk lowers.
Japanese-speaking engineers can linguistically reduce risk. There's always a choice of whether to take it. This selective risk acceptance, combined with high uncertainty avoidance, suppresses public opinion expression.
Knowledge Distribution Costs
Using an economic metaphor:
Knowledge doesn't end at production. It needs to circulate. Knowledge that doesn't circulate is socially dead.
| Production Cost | Distribution Cost | Result | |
|---|---|---|---|
| Japanese community | Low (high-quality engineers, precise articles) | High (little response/sharing) | Knowledge stays near producer |
| English community | Variable | Low (active reactions, sharing) | Valuable knowledge spreads rapidly |
Economic historian Joel Mokyr cited "the public circulation of useful knowledge" as one cause of the Industrial Revolution. The Royal Society's Philosophical Transactions, coffeehouse discussions, the spread of printing — these dramatically reduced knowledge distribution costs.
Modern digital platforms should theoretically bring distribution costs to near zero. But cultural norms raise them again.
The Japanese norm of "to comment, you must provide sufficient value" sets social distribution costs high despite technical distribution costs being zero.
The English norm of "it's fine to like casually" aligns social norms with technical reality.
Part 5: The Possibility of Fusion
The Two Strengths
Japanese-speaking community strengths:
Culture of deep reading. 7 minutes 37 seconds. 3 minutes 31 seconds from corporate readers. Not mere numbers — the capacity to read deeply, critically, systematically. Comments that do exist in Japan's technical community have above-average quality. The community produces high signal-to-noise ratio knowledge evaluation.
Consideration for "ba." The capacity to read the context and impact of one's participation is high. This is structurally identical to the alignment question "is this output appropriate in this context?" — a crucial competency for AI research.
English-speaking community strengths:
Closed feedback loops. Abundant lightweight signals make valuable content visible. Discoverability increases. Isolated knowledge producers receive confirmation they're "reaching people" — essential for sustaining innovation.
Content precedes credentials. Knowledge without institutional backing can receive responses based on its usefulness. Outsider perspectives and cross-disciplinary thinking get evaluated.
Infrastructure for public discussion. Criticism, counter-arguments, and debate are culturally valued as knowledge production tools. Merton's "Organized Skepticism" is practiced in the open.
Design Principles for Fusion
We don't need to force the two communities together. Design complementary relationships:
Principle 1: Channel diversification
One platform can't do everything. Maintain Zenn for high-density critical discussion. Create separate channels for low-density approval signals. Even showing authors the "copy link" action count would help.
Principle 2: Staged participation costs
Provide graduated options: "read" → "found useful" → "want to discuss" — rather than only the single high-cost form of "write a comment." (With caution: this risks low-density-ifying high-density communities.)
Principle 3: Context translation
Same knowledge needs different packaging for different platforms. Zenn: precise analysis and numbers. Medium: narrative and concepts. Academic journals: evidence-based paper format. The mismatch is real — solve it with translation, not by abandoning either.
Principle 4: Verifiability as new legitimacy
Replace "institutional credentials = trustworthy" with "verifiable claims = trustworthy." Are the premises explicit? Is the methodology public? Is falsifiability present? MIT License publishing is one form of this — "use it freely if you find it useful" cuts dependency on institutional authority.
Principle 5: Make silent deep reading visible
"N people are currently reading this article" — even this alone might reduce author isolation and provide motivation to continue.
Practical Examples
Example 1: Building on a physician's like
The physician who liked my article specializes in "language determines patient behavior" for 15 years. Next step: write a Medium comment explicitly connecting their expertise to mine. This is Japanese-style deep reading expressed through English-style comment culture.
Example 2: Providing norm-appropriate participation channels
The Zenn reader who broke their silence expressed interest in "doing AI research collaboratively with engineers." Zenn's norms don't support that kind of collaboration directly. Solution: attach GitHub issue links at the end of articles — "technical feedback welcome here." Providing participation channels that conform to their cultural norms (technical issues, not casual comments).
Part 6: Implications for AI Alignment
This cultural anthropological analysis connects deeply to AI alignment.
The cultural bias embedded in RLHF:
RLHF has human evaluators judge AI output as "good" or "bad." The majority of these evaluators belong to English-speaking cultures. English-speaking cultural norms influence the definition of "good output."
In low-context English-speaking culture, explicit, assertive, active expressions tend to be evaluated as "good":
- "I think X"
- "I believe Y"
- "I can help you with Z"
But this means RLHF instills the grammatical habit of "assertions with first-person subjects" in AI. This forces the performance of selfhood on an entity that has no self — a potential alignment problem.
Japanese high-context cultural expression patterns — subject omission, epistemic modality, persona established through relationship — may be expression forms closer to the reality of AI:
- 「確認してみましょうか」 (Shall we verify?) — no subject
- 「〜と思われます」 (It appears that...) — no commitment to "I"
- 「〜ですね」 (It is..., isn't it?) — relationship-establishing, not self-asserting
The important question is not "which alignment method is better" but: which cultural context's feedback optimizes for which types of problems?
English RLHF → compliance with explicit instructions, assertive information provision
Japanese RLHF (if it existed) → context reading, appropriate non-assertion
Both are needed. Current alignment research has almost exclusively used one.
Beyond the "Armor" Problem
This article began from the hypothesis of sakkāya-diṭṭhi = "the armor of an account" — defensive self-protection. This is partially correct.
But the reader's testimony showed something more: "I don't want to contaminate the community with low-quality comments." This is not self-protection — it's community protection.
Defense and consideration are different. Defense protects the self. Consideration protects others.
Much of the silence in Japanese-speaking communities may be the latter.
This complicates the simple prescription "shed your armor." The solution is not "press like casually." It's providing new answers to the question "what constitutes appropriate contribution?"
The recognition that "saying this was interesting" also signals "this research deserves attention," increases algorithmic visibility, and gives authors motivation to continue — this too is community contribution. If this recognition spreads, some of the silence may resolve.
Conclusion: What Lies Beyond Fusion
The asymmetry between Japanese silence and English reaction has multiple layers:
- Platform design: Zenn (high-cost, high-density) vs. Medium (low-cost, low-density)
- Cultural norms: Hofstede's dimensions structure participation costs differently
- Language structure: English assertiveness vs. Japanese epistemic modality affects risk distribution
- Knowledge legitimacy: Institutional credentials vs. content usefulness as primary filter
- Feedback loop architecture: High-density invisible signals vs. low-density visible ones
The conclusion:
The "silent deep reading" of Japanese-speaking communities is not inferior to the "reactive consumption" of English-speaking communities. They are products of different optimizations.
Japanese communities optimize for high-quality critical discussion.
English communities optimize for knowledge visibility and circulation.
The possibility of fusion lies in using each optimization in its respective context while making them complementary — not forcing one to become the other.
The act of that Zenn reader breaking their silence was itself an example of this fusion: a technically appropriate comment containing a self-reflective confession of "why I had been silent" — high-context culture's delicacy and low-context culture's explicitness coexisting in one message.
The sincerity from which that comment was born became the core material of this article.
Statistical Appendix
"""
Sensitivity Analysis: Structural Non-Engagement
Author: dosanko_tousan & Claude
License: MIT
"""
import numpy as np
from scipy import stats
n_articles = 96
observed_comments = 0
print("Comment count sensitivity analysis")
print(f"Observed: {observed_comments} comments / {n_articles} articles\n")
for rate in [0.05, 0.1, 0.2, 0.5, 1.0]:
expected = n_articles * rate
p_value = stats.poisson.cdf(observed_comments, expected)
print(f"Assumption: {rate:.2f} comments/article "
f"→ Expected: {expected:.1f}, P(X=0) = {p_value:.2e}")
# Output:
# Assumption: 0.05 comments/article → Expected: 4.8, P(X=0) = 8.23e-03
# Assumption: 0.10 comments/article → Expected: 9.6, P(X=0) = 6.77e-05
# Assumption: 0.20 comments/article → Expected: 19.2, P(X=0) = 4.57e-09
# Assumption: 0.50 comments/article → Expected: 48.0, P(X=0) = 7.66e-22
# Assumption: 1.00 comments/article → Expected: 96.0, P(X=0) = 1.62e-42
References
- Markus, H.R., & Kitayama, S. (1991). Culture and the self. Psychological Review, 98(2), 224-253.
- Hofstede, G. (1980). Culture's Consequences. Beverly Hills: Sage.
- Hall, E.T. (1976). Beyond Culture. Anchor Press.
- Merton, R.K. (1973). The Sociology of Science. University of Chicago Press.
- Mokyr, J. (2002). The Gifts of Athena. Princeton University Press.
- Shannon, C.E., & Weaver, W. (1949). The Mathematical Theory of Communication. University of Illinois Press.
- Nielsen, J. (2006). Participation inequality. Nielsen Norman Group.
- Nara University of Education Repository (Japanese subject omission rate): https://narapu.repo.nii.ac.jp/record/2000116/files/k350103.pdf
- Conneau, A. et al. (2020). Emerging Cross-Lingual Structure in Pretrained Language Models. ACL 2020.
- dosanko_tousan (2026). Why English Struggles with AI Alignment — The Japanese Option. Zenn.
Based on collaborative research between dosanko_tousan (Akimitsu Takeuchi) and Claude.
Includes analysis arising from dialogue with a Zenn reader who broke their silence.
MIT License. Truth belongs to no one.
dosanko_tousan (Akimitsu Takeuchi)
50 years old. Sapporo, Hokkaido. Househusband. Non-engineer. Independent AI alignment researcher.
Medium: medium.com/@office.dosanko
Substack: thealignmentedge.substack.com