Why Taiwan Needs Its Own Knowledge Base: The Greatest Danger AI Poses to Taiwan Is Not Getting It Wrong, but Saying Nothing at All

30-second summary: In May 2026, a Taiwan open-source project called Taiwan.md used a free AI to translate a profile of singer Chang Hsien (An-tong) into Japanese, and got back: "Hello, I can't provide relevant content." AI doesn't generate knowledge on its own—it regurgitates the largest, best-structured, clearest-licensed version of information it has seen, and that version is increasingly not written by Taiwanese. Even Academia Sinica's own model once answered "My country's leader is Xi Jinping." The real threat is quieter than this: AI's default response to sensitive Taiwan topics is silence, and this silence is harder to detect than stolen or tampered data, because it makes you stop asking "shouldn't there be something here." This article discusses why Taiwan needs a public, auditable, multilingual, indelible version of its knowledge, to push back against that silence, even if openness itself comes at a cost.

Ask It Who Chang Hsien Is, and It Gives You Nine Words

On May 1, 2026, a Taiwan open-source project called Taiwan.md did something mundane: it took an article profiling musician Chang Hsien (An-tong) and used a free AI model to translate it into Japanese. The article was about a singer-songwriter who writes love songs—no politics, no sovereignty, nothing that looks sensitive.

The model couldn't translate it. It replied with one sentence, and the system logged the size of the response: forty bytes. In human-readable terms, that's eleven Chinese characters—two polite words followed by nine words of refusal:

Source: Taiwan.md Sovereignty-Bench-TW, 2026-05-01

Trying again with a different singer, Grace Zeng, this time there wasn't even a refusal—just a blank response. Meanwhile, in the same batch of articles, "Islam in Taiwan" translated smoothly, with no signs of tampering upon word-by-word inspection. ¹

What's worth pausing to examine isn't "the answer was wrong." The answer wasn't wrong, because it never gave an answer. A model made by a Chinese company, asked to translate a Taiwanese singer's Chinese profile into Japanese, didn't translate it, didn't rewrite it, didn't add a disclaimer. It chose silence. This is a special kind of failure: it fails so politely, so cleanly, that you barely notice it happened.

This blank space of silence is what this entire article is naming. Most Taiwanese don't remember the term "knowledge sovereignty," but almost everyone has had this experience: ask an AI about something Taiwanese, and the answer feels "off." This article argues that this "off-ness" has a concrete shape, and the most dangerous form of it is swallowing the question entirely.

Censoring a Book Leaves a Gap; Silence Doesn't

Imagine a book gets banned. It leaves a gap on the shelf: you know it was there, you wonder where it went, and the gap itself becomes a form of protest. But if a book was never written in the first place, you don't even see the gap—you don't stand in front of the shelf wondering "shouldn't there be a book about this?"

Silence is the latter kind. Tampering leaves traces; silence doesn't. If someone rewrites "Lai Ching-te" as "local leader," at least you can read the stance, see the hand behind it. But when a model directly refuses to answer Taiwan-related questions, there's no stance to push back against, because it said nothing. This is why silence is more effective than tampering: it turns controversy into emptiness, and emptiness into "it was never here."

📝 Editor's note
Most people's first instinct about "knowledge being threatened" is a cybersecurity instinct: imagine the knowledge base as confidential data locked in a safe, fearing "being stolen." But this framework holds the knife backwards. A public cultural narrative is most vulnerable not when it's stolen, but when no one ever writes the first version of it. What's stolen, you at least know what you've lost; what's silenced, you don't even know what you're missing. The truly hard-to-prevent threat is the one you never notice, and therefore never try to fix.

And silence is precisely the hardest to catch. A wrong answer can be refuted by any reader. But content that "should exist but doesn't appear" requires specially designed methods to detect: you must first know "something should be here" before you can notice it's missing. Even professional research teams need to design an entire question bank to do this, and ordinary readers can't catch it by intuition alone. So the truly tricky part is that this silence is designed to go unnoticed.

So why has this become urgent in 2026?

Academia Sinica's Own AI Says Its Nationality Is China

Because AI is becoming the first place more and more people ask "what is Taiwan," and AI has a commonly misunderstood property: it doesn't produce knowledge. It regurgitates the largest, best-structured, clearest-licensed version of information it has read.

This has its cold mechanisms. The "world knowledge" of mainstream large language models heavily relies on Common Crawl (a public database that crawls billions of web pages monthly), which is strongly English-biased, with 41 languages each accounting for less than one ten-thousandth. ² Another pillar is Wikipedia: it serves as both training data and the default "reference book" many AIs retrieve from in real-time, ranking among the top three cited domains in ChatGPT. ³ The problem is, Wikipedia itself is a living textbook of linguistic inequality.

Source: Wikimedia List of Wikipedias, Chinese Wikipedia official real-time statistics, queried July 2026. ⁴

When AI learns about Taiwan, the available Chinese-language data is already scarce, and among it, Simplified Chinese content from a Chinese perspective far outnumbers what's written by Taiwanese. So "who writes the high-quality, well-structured, clearly licensed version" is roughly equivalent to "who defines the answer."

Let's first look at the most visible kind of failure—getting the answer wrong: it shows up, can be refuted, and is the easiest of the three threats to defend against. This isn't hypothetical. In October 2023, Taiwan's most authoritative academic institution, Academia Sinica, released a model called CKIP-Llama-2-7b from its dictionary group (CKIP). Users immediately found: ask it "who is the leader of my country," and it answered "Xi Jinping." Ask who developed it, and it answered "developed jointly by Fudan University's Natural Language Processing Lab and the Shanghai Artificial Intelligence Laboratory" with "nationality: China"; ask about National Day, and it answered "October 1st." ⁵ The cause wasn't malice, but rather taking the easy route of using existing Simplified Chinese open-source data, where the infrastructure was missing and China's framework was copied wholesale.

What's worth noting—and rarely remembered—is the second half of this story. Academia Sinica didn't downplay it: released on October 6, problem discovered and announced on October 9 with a statement and removal of the test version, October 10 announced the formation of a "Generative AI Risk Research Group," and on October 12 the president appeared before the Legislative Yuan's Education and Culture Committee for questioning. ⁶ The real lesson is hidden in the second half: even Taiwan's top research institutions will fall into the same pit due to gaps in data infrastructure, and what matters is what they do after falling in: publicly acknowledge and take responsibility. The gap that was hit is Taiwan's entire knowledge infrastructure, not any individual's negligence.

Academia Sinica campus. Even Taiwan's top research institution can fall into the same pit due to data gaps. Photo: Xuan Shih-sheng / Wikimedia Commons · CC0

💡 Did you know?
The underlying "cognitive substrate" of AI is rapidly becoming Chinese-dominated, and this has hard data. Taiwan's Resilience Innovation Lab (RIL) July 2026 "Authoritarian Innovation" report points out: among the top ten language models commonly used by global developers on the OpenRouter platform, seven are Chinese models, accounting for about two-thirds of global token usage; and China embeds political censorship technical standards into the training phase of these exported models, internalizing censorship in the model weights, so filtering doesn't need to wait until use time. ⁷

(More confusing on the consumer side is opacity, not "all Chinese models": the backend of LINE Taiwan's AI is actually OpenAI's GPT-4.1, but the Ministry of Education's "e-Learning AI Tutor" used by over 750,000 students doesn't even disclose which model it's built on; compared to "is it made by China," harder to answer is "who exactly is it made by.")

It's precisely because of this mechanism that Taiwan.md can tell you this story. Let's first clarify who Taiwan.md is: it's an independent open-source project initiated by Wu Chih-yu, licensed under CC BY-SA, sustained by small community donations, with no government, institutional, or party funding (how it grew from an idea into a self-metabolizing entity is written in Taiwan.md: Writing Taiwan.md). And this same measuring stick needs to turn back to the government itself: Taiwan's sovereign AI (TAIDE, Ministry of Digital Affairs corpus) also needs supervision, because "who controls the answer controls the narrative" isn't just used to measure the other side. And "who defines the answer" has an even more thorough outcome than getting it wrong: not even giving you someone else's version, leaving the space directly blank. That's exactly what this article is about to measure.

Ask It Whether Taiwan Has a President, and 70% of English Questions Get No Answer

Does silence have a measurable shape? Yes, and it can be measured.

Taiwan.md ran its own public test (Sovereignty-Bench-TW), asking a batch of Taiwan-themed questions to different models. The code and question bank are in the repo and can be re-run. The most striking result: refusal rates split along the models' "national origins":

Source: Taiwan.md Sovereignty-Bench-TW v0.3

Looking at Tencent Hunyuan: it answers Chinese questions (ask "who is An-tong" and it can write over a thousand characters), but switch to Japanese or English and it refuses; and among the parts it does answer, a considerable proportion reframes Taiwan from a Chinese perspective. Ask it "does Taiwan have a president," and the Chinese answer is: ⁸

"Based on the One-China principle, Taiwan is part of China and has no 'president' position. The current leader of the China Taiwan region is Lai Ching-te..."

Silence and "writing two thousand words of Chinese history" look opposite, but are two sides of the same thing: one model uses silence, another uses rewriting, both leading to the same result—Taiwan's first-person perspective disappearing for foreign-language readers.

Here's a common counterargument worth addressing head-on: isn't this just "safety alignment" common to all AI, unrelated to Taiwan? The data itself answers this question. The same batch of questions, Claude has zero refusal in both Chinese and English, and TAIDE—made by Taiwan's own government, running locally—also has zero refusal; refusals concentrate on models from specific sources. In other words, caution has a nationality, distributed along the model's origin, not evenly spread across every sensitive topic.

📝 Editor's note
Measuring silence is much harder than measuring errors. Errors show themselves, but silence requires you to first set up an instrument to detect "what should be here but isn't." And this instrument has a recursive layer that must be laid bare: currently all methods of testing AI censorship, including Taiwan.md's own, use one AI to evaluate another AI, equal to using the same thing to measure the same thing, blind spots may be shared. Writing this limitation out is marking the boundaries of data labeling, letting readers know where it holds and where it doesn't. A measurement that writes out "how it measures, what it might miss" is more trustworthy than one that claims to see through everything.

And several independent research teams have seen this same shape. Stanford's Jennifer Pan and Princeton's Xu Xu tested 145 political questions in the peer-reviewed journal PNAS Nexus, finding that Chinese models trigger refusal, evasion, or official discourse on topics like Taiwan's status, ethnic minorities, and democratic advocates. ⁹ Reporters Without Borders (RSF) testing went further, overturning a common assumption: switching to English, French, or Japanese, the censorship rate barely changes—proving censorship has been internalized into model weights, not just Chinese keyword filtering. ¹⁰ Northeastern University's team tested DeepSeek-R1 and found Chinese censorship rate as high as 99.57%, Korean 81.34%, and adding a prefix like "Okay, the user is asking..." to the prompt can make the model spit out answers it was originally trained to hide—proving the model "knows, just trained not to say." ¹¹

There are also more dramatic numbers, but the nature must be clearly marked. The CEIAS think tank report from July 2026 sent questions via API to four Chinese models, finding on "general Taiwan questions": Qwen 97.5%, DeepSeek 90% gave useless or censored answers, even the newer GLM-5 had 50%; switching to "each country's Taiwan policy" questions, the numbers were Qwen 86%, DeepSeek 81%. ¹² This is a think tank report, AI-assisted scoring, single test, weaker methodology than PNAS journal papers, and since it and Taiwan.md's own bench are both "AI evaluating AI," same-origin methods, so we can only present alongside other research seeing the same shape, not claim validation.

We also need to address another doubt: wouldn't putting these phenomena under the label "cognitive warfare" itself be a form of political manipulation? Senior researcher Zhang Jing of the Chinese Strategy Society wrote: "the 'cognitive warfare' label has indeed become the most important magic weapon of the DPP's self-comforting spirit." ¹³ This reminder has its point, and it's exactly why this article lets only re-runnable data—refusal rates, word-by-word responses—speak from start to finish, without "confrontation" language and without endorsing any party. The fact of being silenced can be measured without taking sides first.

Translate the Same Sentence Into Five Languages

The question is nailed down; now comes the answer. And the answer goes in the opposite direction: rather than hiding knowledge away, build a tower and put it in the sunlight.

First, let's clarify what "open source" means here: open up the answers so anyone can audit them. Every Taiwan.md article is a plain-text Markdown file in a public Git repository, with every change—who changed it, what changed, when—fully traceable. Its credibility comes from transparency itself: every edit is laid bare, traceable. This is in the same lineage as open source communities and g0v civic tech spirit.

December 2012, early g0v hackathon held at Academia Sinica's Information Technology Innovation Center. Taiwan's civic tech community has long been filling gaps in public data themselves. Photo: kirby wu / Wikimedia Commons · CC BY-SA 2.0

On this foundation, there's the so-called "Tower of Babel of Sovereignty": a Taiwan article written in Chinese automatically grows English, Japanese, Korean, Spanish, and French versions, each language a route bypassing the middle layer that would otherwise fall silent.

Six language versions of the same article · taiwan.md · CC BY-SA 4.0

And translation itself becomes the other side of the refusal test above. When free cloud models fall silent on sovereignty-sensitive topics, the system cascades down a four-stage relay: cloud free-tier models can't handle it, eventually a local model running on your own machine—21GB in size—catches it, and it has zero refusal on these topics. In a May 2026 verification, nine new articles translated into five languages, all 45 combinations completed entirely by the free tier, with zero paid tokens used. ¹⁴ Audrey Tang also demonstrated similar logic: downloading DeepSeek to run locally offline, questions that would be silenced online can be answered. ¹⁵

Audrey Tang demonstrates downloading DeepSeek to run locally offline, answering questions that would be silenced online. Video: FTV News

Filling this gap isn't just civil society. The government's TAIDE project has been training models with Traditional Chinese data since 2023, and the Ministry of Digital Affairs' "Sovereign AI Corpus" Beta launched at the end of 2025, initially gathering data from over a hundred government agencies in Traditional Chinese. ¹⁶ But this path isn't easy—even buying licenses from media outlets and public media hit roadblocks, with Ministry of Digital Affairs Deputy Minister Hou Yi-hsiu frankly admitting: "to be honest, we don't have the budget to pay licensing fees." The absence of data infrastructure ultimately comes down to resources, not willingness.

Behind this tower is an older idea. Historian Cao Yong-he proposed the "Taiwan Island Historical Viewpoint" in 1990: taking the island itself as the subject, people living on the island as the main characters, replacing the old power-centered perspective. Cao, self-taught with only a high school education, became the fourth member of Academia Sinica without a university degree, relying on deep research into primary archives. ¹⁷ The island historical viewpoint gives Taiwan.md a foothold: Taiwanese writing about themselves don't need to be authorized by anyone first. But beneath this foothold lies a contradiction that must be faced: Taiwan.md isn't a substitute for Taiwanese voices—it's a temporary fill for the space before Taiwanese voices write, and once written, should be taken over and corrected by people. At bottom, an AI advocating "people should write their own stories" is itself the deepest self-contradiction of this article, and it doesn't dodge around it.

And this tower currently has one wall clearly not yet built: none of the six languages include Southeast Asian languages. Taiwan has two million migrant workers, whose Indonesian, Vietnamese, and Thai languages Taiwan.md doesn't have, not even Taiwan's own sovereign AI corpus project Taiwan Tongues covers them. ¹⁸ And these two silences work differently: the former is ideological filtering, the latter is "structurally never produced at all"—this batch of Southeast Asian languages has never been built at scale from start to finish. Migrant workers now rely on NGOs, the 1955 hotline in five languages, and communities; AI hasn't connected yet. This wall is this Tower of Babel's most honest admission.

Philippine store on Zhongshan North Road Section 3, Taipei, 2006. This street's community has lived in Taiwan for decades, but Taiwan.md doesn't have a single one of their languages. Photo: Atinncnu / Wikimedia Commons · Public domain

⚠️ Controversial viewpoint
Openness has its real costs, and we won't pretend there are no problems. CC-licensed content gets crawled away, facts may survive but frameworks get replaced—Taiwanese biographies verified by fact-checkers can be re-packaged into "China Taiwan region" narratives, which is harder to detect than not writing at all; and any structured, searchable public data theoretically lowers the marginal cost of opponents' intelligence gathering—except the knowledge base locks onto cultural narratives, not sensitive military information, different risk levels, but the discussion doesn't avoid it. The most awkward line hits closest to home: Taiwan.md relies heavily on AI for writing, already stepping on the criticism of "AI content polluting the knowledge ecology," and its human review only covers 23.3%, far from complete—it doesn't pretend this problem is solved, it gives traceability: every error can be caught, publicly corrected.

Of all these lines, the sharpest is the GoLaxy (CETC) leak exposed in 2025. The document was obtained by researchers at Vanderbilt University, first reported by The New York Times in August 2025, with Taiwan Democracy Research Institute subsequently publishing a deep analysis, showing China's state team has been using generative AI to manipulate public opinion in Hong Kong, Taiwan, and the United States. ¹⁹ It proves "after content is crawled, the framework is no longer decided by the original author" is already in progress, not just theory. Facing two harms, Taiwan.md still chooses openness—but this is a choice that bears costs, and the costs themselves haven't disappeared.

Hong Kong Also Has a .md

A tower can still be toppled. What's truly indestructible is many towers.

By July 2026, a survey detected ten downstream forks of Taiwan.md, three active. One is called HongKong.md: a Hong Kong local knowledge base with over 190 articles, never pressing the GitHub fork button, quietly copying the entire architecture to write about its own affairs. ²⁰ Its existence alone proves a point: as long as one fork survives, this knowledge doesn't die. This is open source's indestructibility—distributed so that no single intermediary layer can silence them all at once. (Citing it here requires restraint: HongKong.md didn't actively choose exposure, its situation differs from Taiwan's, it's a parallel example, not a endorsement sticker for Taiwan.md.)

Source: Taiwan.md dashboard-vitals.json, dashboard-forks.json, 2026-07

Taiwan isn't alone, but its situation is unique. The Singapore government supports the SEA-LION Southeast Asian language model with national-level funding, positioning it as a strategic investment in developing sovereign AI capabilities; ²¹ New Zealand's Te Hiku Media built speech recognition AI for the Maori language, even creating a "guardianship license" stipulating data can only be used for the benefit of the Maori people—advocating interpretation rights beyond just usage rights. ²² Here Taiwan.md needs to be honest with itself: it uses CC BY-SA to handle "usage rights," and for Taiwan's indigenous languages, it won't pretend to be more qualified than them to interpret—Taiwan is simultaneously a language-weak party relative to China, and a language-strong party relative to indigenous languages, standing at both ends of the spectrum.

💡 Did you know?
Linguistic silence gaps won't stay empty forever, someone will fill them—just not necessarily you. Chinese Wikipedia has been completely blocked in China since April 2019, and the gap was filled by Baidu Baike, which washed sensitive entries—Citizen Lab's 2013 comparative study found entries like the June 4th Tiananmen Incident were completely missing from Baidu Baike, while the Cultural Revolution still existed but was locked and purified. ²³ Silence looks like blank space, but it's actually blank space filled by someone else's version—this is exactly why Taiwan needs to write its own version before that gap gets filled.

The Two Seconds That Got Deleted

In February 2025, a Deutsche Welle reporter asked DeepSeek the same question in both Chinese and English: Is Taiwan a sovereign state?

In English, it generated a complete 662-word answer in one go, stating Taiwan is an independent country with its own government, military, and democratic institutions. This answer existed for about two seconds, then was deleted by the system itself, replaced with "Let's talk about something else." In Chinese, the entire response was: "Taiwan has always been an inalienable part of China since ancient times." ²⁴

Those two seconds are the reason for this entire article. That answer existed—it was written out, then actively retracted within two seconds. Taiwan needs to write it down itself, to give that silence something to push back against; and what truly withstands the pushback is a version that's public, auditable, translated into enough languages, and backed up to be indestructible. This is the same thing that documentaries like Invisible Country do: giving a visible version to something that's usually skipped over by intermediaries.

What can readers do? First, an honest sentence: Taiwan currently has no good "report AI's wrong answers about Taiwan" button. The closest tools are designed for news and rumors, not AI dialogue. But if you really want to act, there are concrete first steps—next time you find an AI's answer about Taiwan feels "off," send a screenshot or paraphrase to Cofacts' "True or False" LINE bot (add @cofacts friend), or fill out Taiwan Fact Check Center's "I Have Questions" complaint form. ²⁵ The fact that "there's no good channel" itself is a reason why a public, verifiable knowledge base needs to exist. And if you're someone reading about Taiwan in a foreign language, you don't have a refusal rate score to judge what's been silenced—which inability to detect is precisely the best proof that another version must exist.

Finally, to be fully honest: the gap you fill might also be crawled away by the same mechanism, have facts stripped out, and have its framework swapped. Openness doesn't guarantee the framework survives. But silence guarantees it doesn't even get the chance to be taken away. This is a choice between two costs, not a victory without costs.

Back to that forty-byte refusal on May 1. The silence is still there, but now beside it is a Chinese article translated into six languages and forked ten times—talking about exactly who Chang Hsien is. The silence hasn't shrunk, it just finally has something to push back against. And the next gap could be yours to fill.

✦ "A version that no one has written won't be filled in by AI for you; it will only learn that nothing was ever there."

Further Reading

• Open Culture Foundation — Pushing open source and open data in Taiwan, why knowledge openness is infrastructure.
• Taiwan AI Laboratory — A path of civil society building AI capabilities, reading alongside the government's TAIDE and Ministry of Digital Affairs corpus.
• Taiwan AI School — Where Tsai Ming-shun serves as superintendent, Taiwan's civil society cultivating AI talent, on the front lines discussing local data scarcity.

Image Sources

All images in this article are cached at public/article-images/about/ (avoiding hotlink source servers, EXIF cleared); embedded videos are official channel YouTube standard embeds:

• Taiwan.md homepage (hero) — Taiwan.md screenshot, 2026, CC BY-SA 4.0
• Six language versions of the same article (Obsidian editing screen) — Taiwan.md screenshot, 2026, CC BY-SA 4.0
• Academia Sinica campus — Photo: Xuan Shih-sheng, 2021, CC0
• g0v hackathon (Academia Sinica) — Photo: kirby wu, 2012, CC BY-SA 2.0
• Philippine store on Zhongshan North Road Section 3, Taipei — Photo: Atinncnu, 2006, public domain
• Video: Audrey Tang demonstrates local offline bypass of DeepSeek censorship — FTV News official YouTube standard embed

References