The Human in the Room: Why Multilingual Research Needs More Than a Good Translation

Twenty-five years ago while conducting my doctoral fieldwork among speakers of Ladino—a dying dialect of Spanish carried across centuries by descendants of Jews expelled from Spain in 1492—I heard a comedian named Vitali describe his language as a travel agency. The audience laughed, and I understood why: speaking Ladino didn't just communicate information. It activated identity, transported people across time and space, and connected them to stories and experiences that no other language could reach in quite the same way.

I've thought about that joke a lot, not only when I started Multilingual Connections back in 2005, but often in the context of market research. The same thing that made Ladino irreplaceable to that community—its capacity to carry meaning that lives below the surface of words—is exactly what gets lost when multilingual research treats language as a final step rather than a design decision.

Language Isn’t a Delivery Mechanism

There's a persistent assumption in research design that language is essentially neutral, and that a well-written English survey question, faithfully translated, will elicit comparable responses across markets.

Words like privacy, trust, satisfaction, and household don't travel cleanly across languages. They carry cultural weight that shifts meaning in ways a grammatically correct translation won't fix. Response scales behave differently across cultures too, as some respondents avoid extremes while others gravitate toward them—introducing measurement error that can make cross-market data look comparable when it's actually measuring different things.

On the qualitative side, I often hear from global researchers that since participants' English levels are high, they're going to conduct interviews in English. But when participants are interviewed in a second language, they simplify. In truncating their answers they trim the nuance, skip the cultural references, and reach for safer or shorter answers. The stories that reveal the most about themselves—the ones built on metaphor, humor, shared cultural memory—often don't surface at all.

This isn't a translation failure. It's a language design failure that happens upstream, before a single word gets sent to a vendor.

What Bilingual Moderators Actually Do

At IIEX Austin in 2024, I presented alongside Kristian Alomá, PhD, founder of Threadline, in a session we called The Art of Emotional Storytelling in Multilingual Research. The presentation grew out of a project that put these challenges in sharp relief. Threadline is a research consultancy whose work centers on uncovering the emotional and narrative layers beneath consumer behavior—the kind of research that depends entirely on capturing not just what people say, but what they mean. The project examined how luxury and wellness resonate across seven markets: Saudi Arabia, the UAE, Kuwait, Egypt, Qatar, India, and China, with ultra-high net worth individuals in each. Our role was translation of guides and stimuli and recruiting bilingual moderators with extensive backgrounds in social sciences and human behavior design.

Understanding concepts of luxury and wellness required language and cultural insiders, not just fluent speakers. Our Hindi moderator on the India interviews intuitively moved between Hindi and English during the sessions—not because the guide called for it, but because that's how the participants thought and spoke. Code-switching isn't a workaround. For multilingual communities, it's how meaning actually gets made, through words that carry weight in one language that their equivalents simply don't in another. One of the things Kristian and I kept coming back to in our presentation was how much fidelity is at stake at every step—translation, moderation, synthesis—and how easy it is to lose the emotional thread without the right people in the room.

That's why the goal isn't just accurate language. It's maintaining the emotional and narrative integrity of the research across every touchpoint.

The AI Translation Question

AI translation has improved significantly, and it belongs in a serious multilingual workflow. It's fast, scalable, and when trained on good glossaries, style guides, and translation memories, can produce clean first drafts. We use it. But "clean draft" and "research-ready" are not the same thing.

Here's what we've seen AI miss consistently in survey work: inconsistent terminology; phrases that are technically neutral in English become emotionally loaded in translation; regional variants collapse into one another when they shouldn't; scale labels that feel natural in English don't always map to equivalents that respondents interpret the same way. Demographics that don't match people's expectations and lived experiences.

The bigger problem is confidence. AI output often looks polished, which makes errors harder to catch. A translation can read fluently in the target language while shifting the meaning of the original question. Unless a human expert is reviewing both versions side by side with research methodology and linguistic accuracy in mind, you may not know you have a problem until the data is already in.

Gender-inclusive language is an important point here, too. In English, the shift toward neutral language has been largely seamless, as it’s easy to incorporate the inclusive “they” and swap out outdated terms like “waitress” for “server” . In Hindi, French, Hebrew, Arabic, and many other languages, gender is structural and is built into the grammar. Workarounds exist, but they can sound artificial or carry unintended political connotations. And well-intentioned efforts to offer inclusive options—asking about non-binary identity, for example—can confuse or even offend respondents in markets where those frameworks aren't culturally familiar, or where honest answers carry real social or legal risk. AI will generate something. It won't tell you whether that something will land the way you intend.

The workflow that actually works pairs AI efficiency with human judgment at every stage: pre-translation planning to flag cultural pitfalls and establish terminology, AI drafting, expert review and localization, platform QA, and in-context testing in the actual survey environment—which catches display and formatting issues in right-to-left languages and character-based scripts that no document review will surface.

Planning Language like You Plan Everything Else

The most common root cause of multilingual research quality problems is a combination of timing and budget. Language gets brought in at the end because it's conceptualized as execution rather than design. But instrument adaptation, demographic localization, and moderator recruiting/briefing are research decisions, not line items to compress.

Demographic questions are a useful test case. Consider household makeup and a common question about children under 18 living at home. In South Korea and some Canadian provinces, the age of majority is 19; in Thailand, it's 20. In many West African, South Asian, and Latin American contexts, adult children routinely live at home until marriage, and multigenerational homes are quite common as well. A question designed around American household assumptions might not just be imprecise but might be genuinely confusing to respondents. Ethnicity and race categories present similar challenges: terminology that is inclusive and familiar in one market can be meaningless or offensive in another. These aren't edge cases. They are the normal condition of global research, and they require expertise and time to get right.

Researchers who do this well treat their language vendor as a partner, similar to how they treat their statistician or their qual moderator: as someone whose expertise shapes the design, not just someone who executes the final deliverable. That means bringing them into the conversation before instruments are locked, building translation timelines that allow for review and adaptation, and budgeting for human expertise rather than assuming AI output only needs a light pass if any.

The Insight You Almost Missed

Language isn't just logistics. It's the medium through which participants tell you who they are, what they want, and what actually matters to them. When it's treated as a checkbox—translated quickly by the cheapest vendor you can find at the end of a project—you get data that looks complete. What you don't get are the stories that only surface when someone feels genuinely understood: the metaphor a participant reaches for in their first language, the joke that unlocks a room, the moment a bilingual moderator follows a thread that a literal transcript would have lost entirely.

Those moments aren't anecdotes. They're the difference between research that confirms what you already thought and research that shows you something you couldn't have found any other way. Building language expertise into the design—not the deadline—is how you get there. And when you do, you're not just improving data quality. You're honoring the thing language actually is: a travel agency, transporting people across the distance between what they mean and what gets heard.