Lost in Translation: What Happens When AI Doesn't Know What It Doesn't Know About Vietnamese Culture

There's a particular kind of frustration that comes from watching a machine confidently say the wrong thing. Not a nervous wrong thing. A certain wrong thing.

That's how Minh Trần, a software engineer based in San Jose, describes the first time he ran a Vietnamese proverb through a popular AI language model. He typed in "Học thầy không tày học bạn" — a saying about how learning from peers can sometimes outpace learning from teachers — and got back a clunky, literal translation that missed the entire cultural point. The AI didn't know it had failed. It just moved on.

"That's the thing," Minh says. "It doesn't know what it doesn't know. And that's actually scarier than it just being wrong."

For Vietnamese-Americans working at the intersection of technology and culture, this is not an abstract concern. It's a daily reality — one that's pushing a growing number of them to stop waiting for Silicon Valley to figure it out and start doing the work themselves.

When the Algorithm Gets Your Face Wrong

Facial recognition bias has been documented extensively in Black and Brown communities, but the conversation around Southeast Asian representation in AI training data is still catching up. Several Vietnamese-American researchers have flagged that facial recognition systems — used in everything from phone unlocking to law enforcement — perform noticeably worse on East and Southeast Asian faces, and within that broad category, Vietnamese faces often fall into an even smaller, underrepresented slice.

Lan Nguyễn, a data scientist in the Bay Area who has spent the last three years studying representation gaps in computer vision datasets, puts it plainly: "Most of these training sets were built fast, built cheap, and built with whoever was easiest to photograph. Vietnamese faces, Vietnamese names, Vietnamese context — it just wasn't a priority."

The consequences aren't always dramatic. Sometimes it's just that your phone unlocks slower than your white coworker's. But sometimes the stakes are higher — in healthcare imaging tools, in hiring software that scans photos, in systems that make quiet decisions about people without anyone in the room to push back.

Lan's current project involves building annotated datasets that include more nuanced Vietnamese visual and cultural context. It's painstaking, unglamorous work. "Nobody's throwing a launch party for this," she laughs. "But somebody has to do it."

The Language Problem Is Bigger Than Bad Grammar

Vietnamese is a tonal language with six distinct tones, a rich system of pronouns that shift based on age and relationship, and idioms that carry centuries of agricultural, Confucian, and Buddhist influence. It is, to put it gently, not a language that forgives lazy modeling.

Large language models trained primarily on English-language internet data — with Vietnamese as a distant afterthought — tend to flatten all of that. They can transliterate. They can approximate. But they regularly miss the weight of words, the social information embedded in how Vietnamese people address each other, and the layered meaning inside expressions that don't have clean English equivalents.

For community organizations trying to use AI tools to reach Vietnamese-speaking elders, this isn't a minor inconvenience. It's a real barrier. Phương Đỗ, who coordinates outreach programs for a Vietnamese senior center in Orange County, tried using an AI chatbot to help draft informational materials in Vietnamese. "It came out sounding like a textbook from 1975," she says. "Our community members thought it was a joke. Some of them were offended."

The formality was wrong. The regional dialect markers were absent. The whole thing felt, as one elder apparently put it, "like a stranger wearing our clothes."

Artists Are Asking the Hard Questions

Not everyone approaching this from the technical side. Some of the most pointed critiques of AI's cultural blind spots are coming from Vietnamese-American artists and writers who are using creative work to expose what the algorithms miss.

Visual artist Bảo Lê, based in Houston, recently completed a series of AI-generated portraits that he deliberately prompted using Vietnamese cultural descriptors — traditional clothing, specific regional landscapes, historical references. What came back was a revealing mess: áo dài that looked vaguely "Asian" but not Vietnamese, landscapes that blended Vietnam with generic Southeast Asian imagery, historical figures who looked nothing like their actual likenesses.

"I wasn't surprised," Bảo says. "But I wanted people to see it. Because a lot of non-Vietnamese people use these tools and just assume what comes out is accurate. It's not. It's a hallucination built on someone else's assumptions."

His work has been shown at two Vietnamese-American cultural festivals this year, and the response has been visceral. People recognize the wrongness immediately. That recognition, he argues, is itself a form of cultural knowledge — one that the machines haven't earned yet.

Building the Fix From the Inside

The encouraging part of this story — and there is one — is that Vietnamese-Americans aren't just complaining from the outside. They're increasingly in rooms where decisions get made, and they're using that access.

A loose network of Vietnamese-American tech professionals has been quietly collaborating on open-source language resources, cultural annotation guidelines, and advocacy within their own companies. Some are pushing their employers to audit AI tools for cultural bias before deploying them in Vietnamese-American communities. Others are contributing to academic research that will, slowly, shift what the next generation of models knows.

There's also a generational dimension here. Younger Vietnamese-Americans who grew up code-switching between two worlds are, in some ways, uniquely equipped to think about what gets lost in translation — because they've been doing that translation their whole lives.

"My parents taught me that you have to carry two versions of yourself," says Minh, the San Jose engineer. "One for home, one for outside. AI only learned the outside version. We're trying to teach it the rest."

Why This Matters Beyond the Vietnamese-American Community

It would be easy to frame this as a niche concern — a specific community pushing back on specific errors. But the implications are broader. As AI becomes the infrastructure of American life — mediating healthcare, education, hiring, news, and social connection — the gaps in its cultural knowledge become structural inequalities baked into the system.

When AI doesn't know Vietnamese culture, it doesn't just inconvenience Vietnamese-Americans. It shapes how other Americans understand (or misunderstand) this community. It affects what stories get told, what faces get seen, what voices get heard.

The Vietnamese-Americans doing this work know they're not just fixing a tech problem. They're fighting for something older and more fundamental: the right to be represented accurately, on their own terms, in a country that has always made them explain themselves.

The machine is learning. Slowly. And the people teaching it — the engineers, the artists, the outreach coordinators, the data scientists — are making sure it learns the real thing.