Multilingual Thinking: Why Language Shapes How We Build (and Use) Technology

[E-Illingworth]

I sat down recently with Elitza Dublewa-Servatius, a Digital Employee Experience & Change Manager at Coca-Cola Europacific Partners whose work spans language technology, localization, and enterprise AI. She has spoken at SlatorCon and LocWorld, and brings an unusual mix of language, philosophy, and AI experience to the conversation. One thing that surfaced as we chatted: philosophy and multilingual thinking can sharpen how we assess and define a problem. And, as AI becomes more embedded in the enterprise, that ability to think holistically from a variety of perspectives becomes an increasingly valuable part of our human intelligence.

Watch the video below or on Youtube

Before her career in localization, Elitza studied philosophy in Germany. It turns out, the two disciplines aren’t as far apart as you might think. Philosophy trains us to question assumptions, test different framings, and keep asking what sits underneath a problem. One practical method Elitza returns to is the Five Whys: ask why, ask again, and keep going until you get closer to what you are actually trying to solve for.

This philosophical problem-solving instinct carries into her love of learning languages. A native Bulgarian speaker, Elitza has spent years studying languages across several language families, including Germanic, Romance, Slavic, and Uralic languages. She studied Dutch and Finnish at university and is currently learning Portuguese. Her philosophical training and multilingual background hone the same underlying skill: the ability to notice a default perspective instead of mistaking it for the only perspective, and that skill becomes especially valuable as AI systems are trained on highly uneven amounts of data across languages.

Every workflow we design carries some assumption about who the user is and what they're owed. Many of us never notice, because we built it in our own language and never had to hold two versions up next to each other. As Elitza put it, though, “You cannot separate language from culture.” Linguistic equivalence alone does not guarantee cultural or pragmatic equivalence. I shared an anecdote from my college days studying Spanish: in a paper comparing a US standardized test in English and that same test in Spanish, I found the English version framed the questions in an active voice, while the Spanish version used more passive constructions. In that particular example, the difference may simply have reflected translation choices rather than a broad cultural trait, but it was a useful reminder that grammatical framing can subtly shift perceived agency. Take it a step further, and a single piece of software localized into a dozen languages can subtly reposition users in different relationships to the product. Can your users understand not just the words, but the intended meaning in a culturally relevant way?

Enter AI, and now that English-first lens can scale globally, creating similar questions around equity and diversity of thought. As an AI champion for her enterprise, I was curious to know more about the intersection of AI and localization. When at work, Elitza tends to prompt in English, but at home, she defaults to her native Bulgarian. One thing she notices immediately is code-switching inside otherwise Bulgarian output: a response written in Cyrillic may suddenly insert an English word or even a whole phrase in Latin script. At other times, the Bulgarian is understandable but still carries English-shaped phrasing or awkward grammatical agreement around an English term. To a bilingual reader, fluent-looking output can hide these traces of uneven multilingual behavior.

Research offers a useful parallel. In one study of the English-dominant Llama-2 family, researchers found that abstract concept representations in intermediate layers sat closer to English than to the input language, suggesting an English bias inside the model even when the final output is non-English.[1] That does not mean every multilingual model literally translates into English, reasons in English, and translates back; different multilingual training regimes can produce different internal patterns. But it does help explain why output can look fluent in another language while still carrying English fingerprints. That is also where a broader concern appears: when many users rely on the same systems, writing and even patterns of framing can start to converge.

“Because LLMs are trained to capture and reproduce statistical regularities in their training data, which often overrepresent dominant languages and ideologies, their outputs often mirror a narrow and skewed slice of human experience.”

Zhivar Sourati[2]

For example, USC researcher Zhivar Sourati and colleagues argue that as the same handful of LLMs mediate more of our writing and reasoning, distinct linguistic styles, perspectives, and reasoning strategies can become more alike. Related empirical work has found a homogenizing effect on creative diversity, while other studies point to cross-linguistic semantic disagreement and to different internal language patterns depending on multilingual training. This echoes something Elitza raised in our conversation: AI can make language and thinking more uniform, and that is not automatically something to celebrate. [3],[4],[5]

Language carries different histories. Language carries different internal logic. And language can influence what we habitually notice, foreground, and express about time, reality, space, and each other. When diverse perspectives are flattened into outputs shaped by dominant training patterns, we risk narrowing the range of linguistic and cultural perspectives that stay visible. Accessibility has long shown how designing for difference can improve systems for everyone. Diversity breeds innovation, and that holds true for globalization, as well.

In the end, you may be shocked to find out that we didn’t solve all of the world’s translation, localization, or AI problems. But, we did return to philosophy and the fundamental questions of ‘why’. Humans were once thought to have a monopoly on tool use, complex emotion, reasoning, self-awareness, and language. Research across animal cognition has complicated many of those assumptions, showing related capacities in different forms and degrees across species. And now AI systems can process volumes of information no individual could read in a lifetime. As AI takes on more tasks we once considered distinctly human, what is our unique human value, and why does it matter?

Let me know what you think in the comments, and follow Elitza Dublewa-Servatius on LinkedIn.

To learn more about the dolphin language referenced in the video, start with this Medium article by Santiago Castro Zaballa.

[1] https://aclanthology.org/2024.acl-long.820/

[2] https://dornsife.usc.edu/news/stories/ai-may-be-making-us-think-and-write-more-alike/

[3] https://www.sciencedirect.com/science/article/pii/S294988212500091X

[4] https://arxiv.org/pdf/2503.04792

[5] https://aclanthology.org/2025.globalnlp-1.18/

And, check out the suggested reading below (or download the attached PDF to print) for a deeper dive on language, philosophy, and culture.

Suggested Reading

The Alignment Problem: Machine Learning and Human Values — Brian Christian (2020)

A trade-nonfiction look at the core technical and ethical puzzle of AI: how do you get a system to actually want what we want, and whose values get encoded by default.

Beyond Culture — Edward T. Hall (1976)

Argues that culture itself operates mostly below conscious awareness, shaping behavior and perception the way grammar shapes a sentence without the speaker noticing.

The Culture Map: Breaking Through the Invisible Boundaries of Global Business — Erin Meyer (2014)

A practical framework for how communication norms (directness, hierarchy, disagreement) vary by culture and routinely cause cross-border teams to misread each other.

Is That a Fish in Your Ear? Translation and the Meaning of Everything — David Bellos (2011)

A wide-ranging, accessible case that translation isn't a flawed copy of an original but its own creative act, constantly remaking meaning across languages.

Mouse or Rat? Translation as Negotiation — Umberto Eco (2003)

Eco, working from his own experience being translated, argues translation is always a negotiation between what a text says and what a target culture can actually hear.

The Silent Language — Edward T. Hall (1959)

The foundational text on nonverbal and contextual communication, showing how much of meaning travels outside of words entirely.

The Singularity Is Near: When Humans Transcend Biology — Ray Kurzweil (2005)

The book that popularized "the singularity" as shorthand for the moment machine intelligence outpaces and merges with human intelligence.

The Stuff of Thought: Language as a Window into Human Nature — Steven Pinker (2007)

For a skeptical point of view - Pinker argues the relationship mostly runs the opposite direction from the Whorfian claim: thought shapes language more than language shapes thought.

Thinking in Systems: A Primer — Donella Meadows (2008)

A short, foundational guide to seeing problems as interconnected systems rather than isolated parts, with paradigm shift identified as the highest-leverage point of intervention.

Through the Language Glass: Why the World Looks Different in Other Languages — Guy Deutscher (2010)

A careful, moderate case for linguistic relativity, built from real cross-linguistic research rather than the stronger, now-discredited version of the Whorf hypothesis.